Artificial Intelligence and Social Work

This book marries social work and artificial intelligence (AI) to provide an
introductory guide for using AI for social good. Following an introductory chapter
laying out approaches and ethical principles of using AI for social work interventions,
the book describes in detail an intervention to increase the spread of HIV information
by using algorithms to determine the key individuals in a social network of homeless
youth. Other chapters present interdisciplinary collaborations between AI and social
work students, including a chatbot for sexual health information and algorithms to
determine who is at higher stress among persons with type 2 diabetes.
For students, academic researchers, industry leaders, and practitioners, these
real-life examples from the USC Center for Artificial Intelligence in Society
demonstrate how social work and artificial intelligence can be used in tandem for the
greater good.

m i l i n d ta m b e is Helen N. and Emmett H. Jones Professor of Engineering and

Founding Co-Director of the Center for Artificial Intelligence in Society at the
University of Southern California. He is a fellow of Association for Advancement of
Artificial Intelligence (AAAI) and Association for Computing Machinery (ACM),
recipient of ACM/SIGAI Autonomous Agents Research Award, Christopher
Columbus Fellowship Foundation Homeland security award, INFORMS Wagner prize
in Operations Research, IJCAI John McCarthy Award, and others. He has contributed
several foundational papers in AI in areas of intelligent agents and computational
game theory, which have received best paper awards at major AI conferences
including AAMAS, IJCAI, IAAI.

e r i c r i c e is Associate Professor and Founding Co-Director of the Center for

Artificial Intelligence in Society at the University of Southern California. An expert in
community-based research and social network science and theory, he has spent the
past several years working to merge social work science and AI to create solutions for
major social problems such as homelessness and HIV. Since 2002, he has worked
closely with homeless youth providers in Los Angeles and many other communities
across the country to develop novel solutions to support young people across the
nation who do not have a home, with the goal of ending youth homelessness.

Published online by Cambridge University Press

 Artificial Intelligence for Social Good
general editors
Eric Rice, University of Southern California
Milind Tambe, University of Southern California

editorial board

Jeff Brantingham University of California, Los Angeles

Mark Fox University of Toronto

Carla Gomes Cornell University

Barbara Grosz Harvard University

Michael Littman Brown University

Desmond Patton Columbia University

Andrew Plumptre University of Cambridge

Artificial intelligence has a diversity of applications, many with social benefits. Books in this series will offer a
multidisciplinary perspective on these applications, ranging from social work to health care to environmental
sciences. Each book will survey the AI approaches to a particular societal problem and promising research
directions, with case study examples.
This is the first book in this series.

Published online by Cambridge University Press

 Artificial Intelligence and Social Work

Edited by

m i l i n d ta m b e
University of Southern California

eric rice
University of Southern California

Published online by Cambridge University Press

 University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre, New Delhi – 110025, India
79 Anson Road, #06–04/06, Singapore 079906

Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781108425995
DOI: 10.1017/9781108669016
© Milind Tambe and Eric Rice 2018
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2018
Printed in the United States of America by Sheridan Books, Inc.
A catalogue record for this publication is available from the British Library.
Library of Congress Cataloging-in-Publication Data
Names: Tambe, Milind, 1965– editor. | Rice, Eric (Eric R. W.), editor.
Title: Artificial intelligence and social work / edited by Milind Tambe,
University of Southern California, Eric Rice, University of Southern California.
Description: Cambridge, United Kingdom ; New York, NY : Cambridge University
Press, 2018. | Series: Artificial intelligence for social good |
Includes bibliographical references and index.
Identifiers: LCCN 2018012813 | ISBN 9781108425995 (hardback : alk. paper) |
ISBN 9781108444347 (pbk. : alk. paper)
Subjects: LCSH: Social service–Technological innovations |
Artificial intelligence. | Social justice.
Classification: LCC HV40 .A79155 2018 | DDC 361.30285/63–dc23
LC record available at https://lccn.loc.gov/2018012813
ISBN 978-1-108-42599-5 Hardback
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.

Published online by Cambridge University Press

 Contents

Contributors page vii

PART I 1
1 Merging Social Work Science and Computer Science for
Social Good 3
Eric Rice and Milind Tambe

2 The Causes and Consequences of Youth Homelessness 16

Eric Rice and Hailey Winetrobe

3 Using Social Networks to Raise HIV Awareness among

Homeless Youth 38
Amulya Yadav, Bryan Wilder, Hau Chan, Albert Jiang, Haifeng Xu,
Eric Rice, and Milind Tambe

4 Influence Maximization in the Field: The Arduous Journey

from Emerging to Deployed Application 57
Amulya Yadav, Bryan Wilder, Eric Rice, Robin Petering, Jaih Craddock,
Amanda Yoshioka-Maxwell, Mary Hemler, Laura Onasch-Vera,
Milind Tambe, and Darlene Woo

5 Influence Maximization with Unknown

Network Structure 77
Bryan Wilder, Nicole Immorlica, Eric Rice, and Milind Tambe

Published online by Cambridge University Press

 vi Contents

PART II 91
6 Maximizing the Spread of Sexual Health Information in a
Multimodal Communication Network of Young
Black Women 93
Elizabeth Bondi, Jaih Craddock, Rebecca Funke, Chloe LeGendre,
and Vivek Tiwari

7 Minimizing Violence in Homeless Youth 119

Ajitesh Srivastava, Robin Petering, and Michail Misyrlis

8 Artificial Intelligence for Improving Access to Sexual Health

Necessities for Youth Experiencing Homelessness 136
Aida Rahmattalabi, Laura Onasch-Vera, Orlando Roybal,
Kien Nguyen, Luan Tran, Robin Petering,
Professor Eric Rice, and Professor Milind Tambe

9 Know-Stress: Predictive Modeling of Stress among Diabetes

Patients under Varying Conditions 153
Subhasree Sengupta, Kexin Yu, and Behnam Zahiri

10 A Multidisciplinary Study on the Relationship between

Foster Care Attributes and Posttraumatic Stress Disorder
Symptoms on Foster Youth 169
Amanda Yoshioka-Maxwell, Shahrzad Gholami, Emily Sheng,
Mary Hemler, Tanachat Nilanon, and Ali Jalal-Kamali

11 Artificial Intelligence to Predict Intimate Partner Violence

Perpetration 195
Robin Petering, Mee-Young Um, Nazanin Alipourfard, Nazgol Tavabi,
Rajni Kumari, and Setareh Nasihati Gilani

12 SHIHbot: Sexual Health Information on HIV/AIDS,

chatbot 211
Joshua Rusow, Jacqueline Brixey, Rens Hoegen, Lan Wei, Karan Singla,
and Xusen Yin

13 Ethics and Artificial Intelligence in Public Health

Social Work 231
David Gray Grant

Glossary 250
Elizabeth Bondi and Mee-Young Um

Index 255

Published online by Cambridge University Press

 Contributors

Nazanin Alipourfard
University of Southern California
Elizabeth Bondi
University of Southern California
Jacqueline Brixey
University of Southern California
Hau Chan
University of Nebraska–Lincoln
Jaih Craddock
University of Southern California
Rebecca Funke
University of Southern California
Shahrzad Gholami
University of Southern California
Setareh Nasihati Gilani
University of Southern California
David Gray Grant
Massachusetts Institute of Technology
Mary Hemler
University of Southern California
Rens Hoegen
University of Southern California
Nicole Immorlica
Microsoft Research

vii

Published online by Cambridge University Press

 viii Contributors

Ali Jalal-Kamali
University of Southern California
Albert Jiang
Trinity University
Rajni Kumari
University of Southern California
Chloe Legendre
University of Southern California
Michail Misyrlis
University of Southern California
Kien Nguyen
University of Southern California
Tanachat Nilanon
University of Southern California
Laura Onasch-Vera
University of Southern California
Robin Petering
University of Southern California
Aida Rahmattalabi
University of Southern California
Eric Rice
University of Southern California
Orland Roybal
California State University, Northridge
Joshua Rusow
University of Southern California
Subhasree Sengupta
University of Southern California
Emily Sheng
University of Southern California
Karan Singla
University of Southern California

Published online by Cambridge University Press

 Contributors ix

Ajitesh Srivastava
University of Southern California
Milind Tambe
University of Southern California
Nazgol Tavabi
University of Southern California
Vivek Tiwari
University of Southern California
Luan Tran
University of Southern California
Mee-Young Um
University of Southern California
Lan Wei
University of Southern California
Bryan Wilder
University of Southern California
Hailey Winetrobe
University of Southern California
Darlene Woo
University of Southern California
Haifeng Xu
University of Southern California
Amulya Yadav
University of Southern California
Amanda Yoshioka-Maxwell
University of Southern California
Xusen Yin
University of Southern California
Kexin Yu
University of Southern California
Behnam Zahiri
University of Southern California

Published online by Cambridge University Press

Published online by Cambridge University Press
 P AR T I

Published online by Cambridge University Press

Published online by Cambridge University Press
 1
Merging Social Work Science and Computer
Science for Social Good
Eric Rice and Milind Tambe

Artificial Intelligence (AI) is a ubiquitous part of our daily lives. Particularly

and inextricably, AI is linked to our online lives. Your every click and post is
being mined by increasingly sophisticated algorithms. Whenever you search
for something online, AI is there. AI helps us more successfully navigate real-
world traffic, not just online traffic. And because it helps us find information
quickly or navigate city streets more effectively, it is easy to appreciate how
AI can make our lives better sometimes. There are some in the tech world who
view AI as a panacea, capable of solving any cognitive task given enough time,
computing power, and rich enough data sets. As two authors who live in Los
Angeles, where traffic is awful day and night, we certainly will not argue with
some of the gains AI can bring to the daily life.
But AI is not always viewed with such a kind eye. Indeed, many in the
scientific community as well as the general public are increasingly concerned
about the negative impact AI may have on society. Some of these concerns,
such as killer robots, may seem like dystopic science fiction fantasies, but
weaponized drones guided by AI may not be so far-fetched. There are also
genuine concerns over job loss, as some traditional white-collar jobs may be
replaced by algorithms that are “happy” to perform dreary organizational tasks
for a fraction of the cost of having a human being perform them. Furthermore,
many ethicists, policy makers, pundits, and entrepreneurs with high influence
in the tech world have begun to worry about the future of society, particularly
when AI is applied in arenas where humans have traditionally made decisions,
such as medicine and the law. We are also sympathetic to these fears and think
that they need to be taken seriously.
This book, however, offers a third perspective. AI can be used to improve
society and fight social injustice. From our perspective, the vast majority of
the persons who currently benefit from AI are people who live with a certain

https://doi.org/10.1017/9781108669016.002 Published online by Cambridge University Press

 4 Merging Social Work Science and Computer Science for Social Good

amount of privilege. Largely this is due to the fact that much of AI is wrapped
up in the high-tech economy. Yet, many of the most impoverished parts of the
world still struggle to obtain internet access. The digital footprints of homeless
persons, for example, are far smaller than those of us who live in more affluent
circumstances.
What we propose in this book is a radical shift in perspective. We dare
you to imagine how AI can be used to benefit the disadvantaged. By bringing
together social work scientists and computer scientists, we can tackle thorny,
seemingly intractable social problems in brand-new ways. Consider this: If
problems such as homelessness, HIV, gender discrimination, substance abuse,
domestic violence, and suicide were easy to solve, then we would have solved
those problems already. Right? But the reality is that many social problems are
incredibly complex and require equally complex solutions.
The complexity of the world can rarely be explained by one single
discipline. Solving complex problems with one discipline stands even less
of a chance. We believe that much of the strength of our work rests in
having created an intellectual space in which computer scientists and social
work scientists are equal partners who bring unique knowledge and expertise
to the table. From AI come new technologies and approaches, and from
social work comes a deep understanding of human behavior and how to
intervene effectively in society. As such, we see AI for social good as a new
transdisciplinary intellectual space.

Motivations for Our Research

In some respects, this book is an introduction to the work at the University
of Southern California (USC) Center for AI in Society (CAIS). This center
is one of the first university-based centers focused on AI for Social Good.
More importantly it is the first center to deliberately bring together computer
science and social work science. CAIS is a joint venture between the Suzanne
Dworak-Peck School of Social Work and the Viterbi School of Engineering.
Our mission is to conduct research in Artificial Intelligence to help solve the
most difficult social problems facing our world.
At CAIS, we seek to bring researchers from around the world to focus on
how computer science can be used to solve social problems. This requires an
engagement by social work science and computer science. As a field, AI has
just begun to consider the role it can have in promoting social good. Similarly,
social work has just begun to engage with computer science to enhance the
impact of social work science. It is our intention to have the research emerging

https://doi.org/10.1017/9781108669016.002 Published online by Cambridge University Press

 Eric Rice and Milind Tambe 5

from this new intellectual space impact changes in national and international
policy, enhance specific social programs, and move toward the creation of a
new field of study that merges the technological expertise of engineering with
the domain expertise of social work in solving seemingly intractable social
problems.
There are many ways one can define social problems, but at CAIS we
largely turn to three major sources of inspiration: the Grand Challenges
of Social Work, the Grand Challenges of Engineering, and the Sustainable
Development Goals of the United Nations. To learn more about these, we
encourage you to explore the following web resources:

• Grand Challenges of Social Work (http://aaswsw.org/grand-challenges-

initiative/)
• Grand Challenges of Engineering (www.engineeringchallenges.org/)
• United Nations Sustainable Development Goals (www.un.org/
sustainabledevelopment/sustainable-development-goals/)
• United Nations Millennium Development Goals (www.un.org/
millenniumgoals/environ.shtml)

The Grand Challenges of Social Work is a list of twelve challenges identified

by the National Academy of Social Work to inspire social work practitioners
and scientists in the twenty-first century. Progress made on any of these
challenges will lead to important improvements in human well-being. Some
of the challenges that we are particularly motivated by include issues such
as ensuring healthy development for all youth, eradicating social isolation,
ending homelessness, and harnessing technology for social good. Some of
our projects, such as the HIV prevention program for homeless youth that
we describe in detail in this book, draw inspiration from multiple challenges.
That project in particular is inspired by the desire to harness technology for
social good, end homelessness, and empower homeless youth to ensure healthy
development for a group of youth who are often marginalized and discarded
by mainstream society.
We also draw inspiration from some of the Grand Challenges of Engineer-
ing. Like the Grand Challenges of Social Work, the challenges identified by
the National Academy of Engineering are for engineers in the field as well
as engineering scientists. Two of the challenges we are particularly motivated
by are advancing health informatics and securing cyberspace. We have a set of
emerging projects that work with administrative data from health providers and
social service providers. We are interested in helping build human–machine
collaborative decision-making systems that can improve service delivery in
both the health and social service fields. Likewise, we have a set of projects

https://doi.org/10.1017/9781108669016.002 Published online by Cambridge University Press

 6 Merging Social Work Science and Computer Science for Social Good

that have focused on public safety and security, both cyber and physical. Some
concrete example projects include our collaborations with the US Coast Guard,
the Transportation Security Administration, and different police departments.
Here we have used AI for enhancing security resource optimization; these
algorithms are in daily use for improving randomized patrol routines and
allocation at major airports and ports around the United States, including New
York, Los Angeles, and Boston. In some cases, such as the US Coast Guard
patrols around the Staten Island Ferry in New York, they have radically altered
the patrol patterns and resulted in improved effectiveness.
UN Sustainable Development Goals and UN Millennium Development
Goals also help anchor the work of the center. We include these goals
alongside the Grand Challenges for Social Work because we feel that those
Grand Challenges have a domestic flavor, drawing primarily from issues that
impact the United States and other developed nations. We doubt this was
an intentional oversight, but at CAIS we are interested in issues that may
impact the developing world, and emerging economies. We draw obvious
inspiration from sustainable development goals such as good health and well-
being, equality, no poverty, and decent work and economic growth, as well as
millennium development goals such as eradicating extreme poverty, reducing
child mortality, promoting gender equality and empowering women, and
combating HIV/AIDS, malaria, and other diseases. But we also look to life on
land and to ensuring environmental sustainability. We have actively explored
the ways of using AI for environmental sustainability and conservation.
Concrete ongoing projects include our continued work with nongovernmental
organizations (NGOs) for wildlife conservation, including World Wildlife
Fund (WWF), Wildlife Conservation Society (WCS), and PANTHERA to
implement AI-based techniques for assisting rangers in the field. Here AI
algorithms are used for predicting poacher activity, so that patrols can more
effectively stop the poachers before they deploy their traps to kill animals.
We hope to write a volume dedicated to work on AI for sustainability in the
future.
There is a great deal of overlap across the Grand Challenges and Goals
from which we draw our inspiration. Both Social Work and the United
Nations prioritize fighting poverty, promoting equality, and ensuring healthy
development for children. Both Engineering and the United Nations prioritize
clean drinking water and improved health. What these challenges and goals
share is a focus on improving lives for all people, especially underserved
and marginalized populations throughout the United States and internationally.
These set of goals provide important new directions for engineers and social
workers. It is important to remember that these goals and challenges are not

https://doi.org/10.1017/9781108669016.002 Published online by Cambridge University Press

 Eric Rice and Milind Tambe 7

easy problems to tackle. In fact, a cynic could look at these lists and quickly
dismiss most of the topics. If these were easy things to achieve, they would
have been achieved already. What we have come to appreciate is that a merging
of social work and engineering empowers each to be more effective in tackling
these seemingly impossible challenges. We are not offering a panacea, but we
are offering a new path. It is perhaps an unlikely path, but we hope that the
contributions in this book will do much to demonstrate the potential we have
come to see.

What Is Social Work

Since this book aims to speak to two very different audiences, social work
scientists and computer scientists, a brief overview of just what happens in
these two disciplines seems warranted. Unfortunately, a few paragraphs on the
topics can hardly hope to do each discipline much justice. Moreover, such
a short treatment will inevitably reflect our own personal stances within our
respective disciplines. That said, we hope these next few paragraphs help
you, our readers, to be grounded to some extent in the work of these two
fields.
Social work is fundamentally an intervention-driven discipline. The histor-
ical roots of social work and its pioneers, particularly Jane Adams, Elizabeth
Fry, and Mary Ellen Richmond, were first and foremost in improving the
human condition, particularly for the disadvantaged. Over the past 100-plus
years, social work has emerged as a practice and a science. Individual social
workers work in a variety of settings. The most common are as members of
child welfare agencies, community-based or nongovernmental organizations,
as community organizers, and in policy-making organizations. What ties this
work together is an overarching drive to fight inequality and to help all
people live to their full potential. There has been and continues to be a
fundamental focus on marginalized and disadvantaged groups within society
and on assisting those communities and the individuals who are a part of those
groups to thrive.
Social work science focuses on the complex causal forces that impact
human life. Central to the discipline is an understanding of human well-being
as existing in the nexus of micro-, messo-, and macro-level forces. When social
work thinks about micro-level issues, the focus is on the psychology, biology,
and even spirituality of individuals. Messo-level action is concerned primarily
with the functioning of social networks and local communities. Macro-level
issues are conceptualized as larger social forces, such as institutions and social

https://doi.org/10.1017/9781108669016.002 Published online by Cambridge University Press

 8 Merging Social Work Science and Computer Science for Social Good

norms. The science conducted by social work scientists happens at any of these
three levels and often cuts across these levels.
Research conducted by social work scientists can take many shapes, but
fundamentally there is always a focus on intervention and improving the lives
of humans in need. Some of the goals of social work is specifically about
the design and testing of new behavioral and/or mental health interventions.
Social work scientists tend to want to create change in the world, and many are
not happy to let their interventions be sequestered to university-based work.
Thus, in recent years, there has begun to be a focus within the social work
field on what is called implementation science, that, is the science of bringing
evidence-based interventions to scale in communities.
Social work scientists also do a large amount of observational social science
research. Much of this work is based on using statistical techniques to analyze
survey data collected from a variety of populations, ranging from nationally
representative samples to community-based samples of marginalized groups,
such as homeless youth. There has been a recent turn toward using admin-
istrative data from social welfare and health care service settings as well. In
addition, there are also qualitative methods, such as ethnographic work, focus
groups, and individual open-ended interviews. But all of this work is in the
service of intervention. Almost every article written by a social work scientist
includes a discussion of how their finding impact intervention design or social
policy change. Social work science is fundamentally a science of action.

What Is AI
There is no accepted definition of Artificial Intelligence, but generally speak-
ing, we will consider AI to include computer systems (including robots) that
perform tasks that require intelligence. From a discipline that the public knew
almost nothing about a few decades ago, AI has grown into a discipline
that is often in the news (arguably for the wrong reasons), and has led to
national research strategies being developed in the United States and the United
Kingdom.
Artificial Intelligence began in the late 1950s. Four scientists are generally
considered to be the founding fathers of AI: Allen Newell and Herbert Simon
of Carnegie Mellon University; John McCarthy of Stanford University, and
Marvin Minsky of the Massachusetts Institute of Technology. Since those early
years of optimism, AI has gone through its share of ups and downs; it is worth-
while taking a quick tour of these developments and understanding the roots
of the discipline. The early years of optimism in AI were based on computer

https://doi.org/10.1017/9781108669016.002 Published online by Cambridge University Press

 Eric Rice and Milind Tambe 9

programs performing tasks that could be considered difficult for humans, such
as playing chess, proving theorems, and solving puzzles. Early progress was
dominated by use of logics and symbolic systems, heuristic reasoning, and
search algorithms. If we are playing chess or proving a theorem, we do not face
uncertainty in our reasoning steps: e.g., statements are either true or false; chess
moves are deterministic. Accordingly, probabilistic reasoning and uncertainty
did not play a major role in the field. Since interacting with the outside world
was not a priority in these tasks, a program could reason about a situation in
depth and then react.
Early years of optimism also focused on integrated AI systems that
could combine sensory perception or computer vision, machine learning to
autonomously learn from interactions of the AI system with the world, auto-
mated reasoning and planning to act in the world, possibly speech and dialogue
in natural language, social interactions with humans and other entities, and so
on. Such intelligent agents include robots, but can also be avatars in virtual
worlds. This vision of AI involves gathering data from the world and machine
learning to build a predictive model of how the world would evolve. However,
it does not stop only at making predictions, but also includes planning and
reasoning from first principles, i.e., providing prescriptions and taking actions
in the world.
This optimistic period had reached its end in the 1990s, when AI research
was accused of not living up to its promises, which led to a general cut
in funding and interest in AI. Researchers realized that what was easy for
humans (e.g., recognizing objects in images, understanding human speech as
English words) turned out to be difficult for AI systems. As AI systems started
interacting with the real world, there appeared a need for an entirely new set of
tools to handle reasoning in the presence of uncertainty and to provide quick
reactions to unanticipated events. These difficulties led to new quantitative
tools from decision theory, optimization, and game theory being brought into
AI. This period also coincided with an emphasis on fast reactive reasoning
(rather than purely deep reasoning) in AI, particularly for robots interacting
with the world; this was the beginning of the rise of a new AI, with the early
symbolic AI approaches being called “Good Old-Fashioned AI” (or GOFAI).
With increasing computing power and the ability to gather large quantities
of data, we now have reached yet another new era of AI. Machine learning
using deep neural nets has led to significant new successes in tasks such as
recognizing objects in images and natural language understanding. While these
new approaches to machine learning have not penetrated all of the traditional
areas of AI (e.g., planning and reasoning or social reasoning), this new era
has brought significant new interest and investment from industry. It has also

https://doi.org/10.1017/9781108669016.002 Published online by Cambridge University Press

 10 Merging Social Work Science and Computer Science for Social Good

led to a mistaken perception among the general public that AI is only about
machine learning and prediction, and that unless there are vast quantities of
data, AI cannot function. Often not well understood is AI’s ability to provide
prescriptions, by reasoning through millions of possibilities based on millions
of simulations from first principles of how entities may interact in the world.
In our work described in Part II of this book, we often focus on this ability to
plan interventions.

The Unlikely Partnership

AI and social work may seem like an unlikely combination or even a clashing
set of disciplines. It turns out, however, that there are three major intellectual
points of convergence between our individual research agendas, which make
this a wonderful, albeit surprising, partnership.The research agendas embrace
complexity in social settings. Likewise, both embrace uncertainty. And perhaps
most importantly, both value praxis.
Eric Rice’s research agenda has been rooted in the idea that the social world
is a deeply complex place. The causes of social injustice are complex, and are
the results of the intersections of historical oppression, contemporary societal
structures, psychological issues, biological issues, and even spiritual issues.
Much of his work prior to connecting with that of Milind Tambe focuses on
the complexity of social networks and how the web of interconnections that
surround people impact human well-being. Likewise, Tambe’s research agenda
is driven by complex real-world multi-agent system problems, an area of AI
focused on social interactions. This area traditionally focuses on employing
machine learning, planning, and reasoning in complex social environments
involving interactions among humans, as well as between humans and AI
(including software or robots).
In their research, both Tambe and Rice must wrestle with uncertainty.
From the perspective of computer science modeling, the probabilistic nature
of outcomes is par for the course, and techniques such as sequential decision-
making and planning under uncertainty that we will discuss later are predicated
on not knowing all of what happens in the world. Social work, likewise, has a
deep understanding of the uncertainty of the social world. Human beings are
not only complicated; they are idiosyncratic. Social work loathes “one size fits
all” solutions and embraces understanding of not just what can be known but
also what is not known.
But perhaps the greatest point of interconnection between Tambe and Rice’s
research agenda is a desire to implement solutions. Rice views social work

https://doi.org/10.1017/9781108669016.002 Published online by Cambridge University Press

 Eric Rice and Milind Tambe 11

science as fundamentally a science of intervention. Even when his schol-

arly articles focus on theory, methodological developments, or observational
research, the intent is always to further the well-being of populations in need.
Thus, among Rice’s published work, there are articles that do not explicitly
test new intervention strategies for homeless youth, but nearly all of these
articles end with explicit discussions of how that particular piece of research
may impact program development or policy change. Likewise, if you look at
Tambe’s publications, while some are more theoretical than others, focusing
on the development of new algorithms, particularly over the past decade, all of
this algorithmic work is explicitly focused on trying to assist in the solution of
real-world problems.

New Science for Both Sides

Social work problems and AI may seem far apart, but we argue that, when
brought together they engender powerful new science. We argue that AI for
social good, or AI and social work, is a new intellectual space, because what
has come from our collaborations is unique to both AI and social work. This
is not just an issue of applying “out of the box” AI tools (as is the case
sometimes when social scientists discover that new analytic techniques are
available), nor is it an issue of doing typical social work interventions but
with a technology enhancement (as is sometimes the case with mobile health
technology projects). What we have created is a new intellectual space that
is generative of innovative science from both the computer science and social
work science perspective (Figure 1.1).

Figure 1.1 The USC Center for Artificial Intelligence in Society’s approach to
research

https://doi.org/10.1017/9781108669016.002 Published online by Cambridge University Press

 12 Merging Social Work Science and Computer Science for Social Good

We can illustrate this point best in our work on HIV prevention for homeless
youth. From the perspective of AI, influence maximization in the context
of homeless youth networks introduced several new challenges for which
algorithms did not exist. For example, traditional AI influence maximization
algorithms, focused on domains such as product advertising, have relied on a
“clean” data set of a social network (possibly downloaded from social media)
without uncertainty. However, the social network of homeless youth is not
downloaded from social media; rather, it is obtained from field observations
of social workers, and may be full of uncertainty. Thus, we had to address
new research challenges in influence maximization algorithms focused on
reasoning with uncertainty in social networks and adapting to availability of
new information reducing or adding to such uncertainty. Indeed, this research
work has been lauded for its innovation with a number of awards at major
AI conferences. Likewise, from the perspective of social work, working on
diffusion-based interventions is common, but thinking about diffusion as
a formal problem of influence maximization radically shifted the view of
networks and actions of actors within those networks. And these innovations
were not only significant from a scientific perspective but, as you will see in
Chapter 4, made a real impact on the lives of homeless youth.
More generally, there are many new challenges in this interdisciplinary
space that do not fit the traditional techniques of either field. As we discussed
earlier, the new wave of AI has been driven by large data sets, massive
computational power, and hence deep learning; this AI work has often focused
on learning predictive models. The type of AI research required for this
interdisciplinary space is often devoid of such convenient data sets. Indeed,
we are often focused in low-resource communities that have incomplete and
uncertain data. Often the massive computational power is in short supply;
resources are very limited. Finally, we are often interested in going a step
beyond prediction to prescription, to intervention. This style of AI research
often calls for partnerships between humans and AI. For example, in our work
on HIV prevention for homeless youth, humans are the ones responsible for
interacting with homeless youth, whereas the AI programs focus on selecting
which youth the social workers should interact with.
Similarly for social work, engaging in formal modeling of social processes
in the context of AI allowed us to identify and solve intervention implemen-
tation challenges. As we discuss later in the book, peer-led HIV prevention
programs have not always worked in all contexts. Part of this failure may have
to do with who is picked to be a peer leader. The person who delivers the
messages is as important as the messages being delivered. Two approaches
to peer leader selection are typically used. Peer leaders are selected either
based on attributes that seem to indicate they would possess the capacity to

https://doi.org/10.1017/9781108669016.002 Published online by Cambridge University Press

 Eric Rice and Milind Tambe 13

be a peer leader or by observing a population and selecting the most popular

persons. Thinking about peer leader selection within the context of influence
maximization, however, shifts the thought process toward the importance of
interconnections within the community and the importance of groups of per-
sons working in concert to enact social change. Even static network solutions
such as the persons with the most connections to others were abandoned in
favor of more dynamic solutions. This approach to intervention development
and implementation of social programs is unique to this collaboration between
social work and AI.
In addition to the new techniques discussed earlier, this interdisciplinary
space also calls for support of new scientific traditions. AI research focuses
on improving tools and techniques, and best papers in AI conferences pro-
vide technical breakthroughs in models, representations, algorithms, theorems
and/or new analysis of algorithms or their theoretical or empirical behaviors.
This is fast-paced research, with new breakthroughs coming fast and furious,
and often easily evaluated with massive simulations or other evaluations on
clean datasets. The focus is “inward,” toward the technique. This scientific
progress is exciting and obviously important. However, in our interdisciplinary
space, it is important to go beyond the technical breakthroughs per se and
measure their impact on society. What is required is a careful, patient, and
detailed measurement of the pros and cons, benefits and deficits, conducted on
the scale of human social interactions, potentially carried out over months.
Such detailed understanding of impact of new proposed interventions is
absolutely essential, but it is new to traditional AI. The result of such research
is not new algorithms or new theorems – the staple of publications in AI – but
detailed analysis of impact on society. This is clearly a new kind of science, one
that requires nurturing. Similarly for social work, the process of intervention
development from acceptability trials, to feasibility trials, to efficacy trials,
and finally to effectiveness trials is an incredibly slow-moving process, one
that typically takes more than a decade of work with a single new intervention
program. We need new, more nimble ways of testing interventions in order
to take advantage of the more fast-paced work of computer science. Perhaps
something akin to beta-testing is needed for our work to keep pace with the
emerging opportunities presented by collaboration with AI.

The Structure of the Book

This book has two sections: Part I focuses on our ongoing transdisciplinary
research specifically around social issues such as homelessness and HIV risk.
We focus on how AI can enhance the implementation of HIV prevention

https://doi.org/10.1017/9781108669016.002 Published online by Cambridge University Press

 14 Merging Social Work Science and Computer Science for Social Good

interventions for the homeless youth. Some of this work has been published
in a series of papers at various international AI conferences including IAAI
2015, AAMAS 2016, AAMAS 2017 (two separate papers), as well as several
workshop papers and AI Magazine 2016. We present a version of this
work that is more accessible to an interdisciplinary audience rather than
a purely AI audience. This work with homeless youth drop-in centers in
Los Angeles illustrates that AI algorithms can actually significantly improve
spreading of HIV prevention information among homeless youth. The work
has received significant number of awards, including AAMAS16 best student
paper, AAAI17 best video and best student video award, and a nomination for
AAMAS17 best paper. This section also includes an overview of the social
context of youth homelessness in the United States, provides an exploration of
our first algorithmic solutions to peer-led HIV prevention models, “real-world”
field tests of these algorithms with actual homeless youth in Los Angeles,
and future thoughts about enhanced algorithmic techniques that rely on less
extensive network data collection methods.
Part II of the book features seven aspirational chapters written by graduate
students at USC. All of these chapters represent the work of transdisciplinary
teams of students that include both social work and computer science graduate
students working on a range of difficult social problems using cutting-edge
techniques from AI to attempt to create new solutions to wicked social
problems. These chapters were the product of a course that we co-taught in
the spring of 2017 at USC called “AI for Social Good.” The class included
students from both social work and computer science, mostly PhD students
with a handful of master’s students. These chapters showcase many problems
that can be tackled by this novel collaboration of AI and social work. The
problems covered include interpersonal violence, intimate partner violence,
sexual health information dissemination, posttraumatic stress, depression,
and homelessness. The AI work includes machine learning techniques such
as neural networks, decision trees, and Bayesian networks. There are also
chapters that utilize sequential planning under uncertainty, the Uncertain Voter
Model, and advances to the k-center problem. There is also an initial foray
into natural language processing, with the creation of a Facebook bot to
autonomously deliver sexual health information to sexual-minority youth. This
section of the book concludes with a chapter on ethics of autonomous agents in
the context of social work and public health interventions, written by a CAIS
summer fellow who is a doctoral student in philosophy.
We end the book with a glossary of terms. Perhaps the most challenging
aspect of AI for social work is that social work science and computer science
both have very specific, technical, jargon-filled, academic languages. And

https://doi.org/10.1017/9781108669016.002 Published online by Cambridge University Press

 Eric Rice and Milind Tambe 15

these languages are not at all the same. Over the past few years we have
learned that often we use different words to mean the same thing and the same
word to mean different things. The word “objective” for an AI researcher is
often framed in the context of an optimization problem, and implies a precise
mathematical statement, as opposed to the objective that a social scientist may
wish to achieve for, say, homeless youth in the real world. The “objective” of a
real-world intervention is often something far more nebulous, such as to help
prevent the spread of HIV among homeless youth. What a computer scientist
calls a feature of a data set a social scientist would call a variable. And perhaps
most frustrating, the difference in the word “model” in the context of computer
science and social science is so different as to be comical, if it were not so
disruptive at times. Because we have experienced the hard growth of trying to
become bilingual, we share with you this glossary.

https://doi.org/10.1017/9781108669016.002 Published online by Cambridge University Press

 2
The Causes and Consequences of Youth
Homelessness
Eric Rice and Hailey Winetrobe

In this chapter, we provide the social context and foundation for our work.
“Have You Heard?” is a behavioral intervention that focuses on peer-led HIV
prevention for homeless youth (Rice, Tulbert, et al., 2012). Our recent work
has augmented this intervention by using the artificial intelligence algorithms
HEALER and DOSIM to improve the implementation of this intervention.
This chapter provides a broad overview of why this work is important, by
discussing (1) the extent of youth homelessness, (2) the causes of youth
homelessness, (3) the consequences of youth homelessness (particularly HIV
infection), (4) the importance of social networks in the lives of homeless
youth, (5) the complexity of working with networks of homeless youth,
and (6) how and why AI- and network-based interventions for homeless
youth are needed to increase the success of HIV prevention intervention
implementation.
The “Have You Heard?” intervention, which utilizes HEALER and DOSIM,
is designed to be delivered at homeless youth drop-in centers. These are often
daytime programs providing food, clothing, hygiene supplies, and access to
social workers who work with homeless youth to try to help them achieve
stable, successful, independent lives (e.g., employment services, health care
referrals, housing assistance) (De Rosa et al., 1999; Shillington, Bousman, &
Clapp, 2011; Slesnick, Dashora, Letcher, Erdem, & Serovich, 2009). Drop-in
centers tend to have very few barriers to youth (e.g., not mandating sobriety,
regardless of mental health status), and as such, they have been shown to be
the more successful route for engaging homeless youth compared to shelter
programs, which often have many more restrictions (Slesnick et al., 2016).
Moreover, drop-in centers are available to youth experiencing a range of home-
lessness situations (e.g., couch surfing, temporary living programs, shelters,
street-based), thus creating greater reach.

16

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 Eric Rice and Hailey Winetrobe 17

A systematic study of homeless youth in Los Angeles found that 78% of

sampled youth accessed drop-in centers (even higher among youth recruited
on the streets: 86%) (De Rosa et al., 1999). Moreover, homeless youth rank
drop-in centers as easily accessible and are very satisfied with drop-in center
services (De Rosa et al., 1999). Because the “Have You Heard?” intervention
has been designed to be delivered in the context of these drop-in centers, much
of the research we draw on in this chapter comes from work conducted by
our team and others who sampled youth from such drop-in centers. This is
not to say that work done by other research teams in other contexts (e.g.,
in shelters, outside of Los Angeles, or even outside of the United States) is
unimportant. We focus primarily on the research on homeless youth drawn
from this specific context to provide the most relevant background, describing
the intervention participants’ peers’ demographic characteristics, histories, and
behaviors.

The Extent of Youth Homelessness

We, most other researches, and homeless service providers define homeless
youth as persons between the ages of 13 and 24 who are living on the streets,
in emergency shelters or transitional living programs, or “couch surfing” (i.e.,
staying with friends, family, or strangers on a temporary basis) without the
accompaniment of their parents or guardians. There is not a complete census
of the number of homeless youth in the United States, though it has been
estimated previously that there are approximately 1.6 million youth (7% of
the youth population) in the United States who experience at least one night of
homelessness each year (Ringwalt, Greene, Robertson, & McPheeters, 1998).
(Currently, researchers at Chapin Hall at the University of Chicago are working
to update this number [voicesofyouthcount.org].) Los Angeles, the setting
for our research, is arguably the nation’s capital for homeless youth. The
2017 point-in-time census of homeless individuals in Los Angeles County
found 3,233 unaccompanied youth between the age of 13 and 24 sleeping
on the streets or living in emergency shelters on any given day (Los Angeles
Homeless Services Authority, 2018).
Homeless youth experience a variety of living situations. Based on our prior
work with youth who access drop-in centers, we found that 31% were sleeping
on the streets, 17% were staying in an emergency shelter, 7% were staying
in a motel, and 44% were couch surfing (Rice, 2010). While the point-in-
time census shows an alarming number of youth on the streets each night,
it is invariably an undercount. The point-in-time count numbers always miss

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 18 The Causes and Consequences of Youth Homelessness

the 7% in motels and the 44% who are couch surfing, not to mention that
the point-in-time count is a three-day snapshot of visibly homeless youth,
which is not representative of the transience of youth who cycle in and out
of homelessness. Moreover, couch surfing is the preferred strategy for coping
with homelessness among this population, but it is not a stable form of housing
and typically lasts anywhere from a few days to a few months, but is never a
permanent housing solution.
It is important to understand that homelessness for youth is very cyclical.
Following their first homeless experiences, 73% of youth exited homelessness
within two years, but 49% cycled in and out of homelessness for those two
years (Milburn et al., 2009). On average, homeless youth we surveyed in drop-
in centers were 16.7 years old when they first experienced homelessness, but
there was much variation in this number (standard deviation was 3.9 years)
(Winetrobe, Rhoades, Rice, Milburn, & Petering, 2017). For most of these
youth, homelessness dominated their lives. When asked to sum up their total
time spent homeless, 66% of these youth reported more than two years of their
lives, and only 13% reported less than six months (Winetrobe et al., 2017).
While many youth at any point in time are couch surfing, 68% of youth report
having slept on the streets, park, beach, or some other outdoor environment in
the past year (Rice, Winetrobe, & Rhoades, 2013).
Homeless youth, despite their enormous numbers, are largely invisible to
the world around them. Most of these youth do everything in their power
to blend in with the world at large. The stereotype of homeless individuals
pushing shopping carts or sleeping in “nests” of personal belongings does not
apply to these youth. In our experience with homeless youth over the past 15
years, we would not recognize the vast majority as being homeless if we were
to walk by them on the street. Frankly, if we did not know them personally from
the drop-in centers where we conducted our research, they would be hidden
from us, too. Yes, some homeless youth are identifiable, sitting on a corner with
a cardboard sign, begging for spare change. However, in our experience, the
vast majority of homeless youth we have known and worked with looked like
any other teen or young adult walking down the street, albeit their shoes might
have been a little more worn and their backpacks perhaps a little overstuffed.
This invisibility is by design. A recent study of ours found that only 52% of the
youth we surveyed answered “yes” to the question, “Do you consider yourself
homeless?” (Winetrobe et al., 2017). These are teenagers and young adults
who think of themselves first as teenagers and young adults, not homeless
youth. They want to blend in; they do not want to be targeted because they
are homeless. “Homeless” to them looks more like the older homeless adults
who are entrenched in street life.

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 Eric Rice and Hailey Winetrobe 19

Causes of Youth Homelessness

Reasons for homelessness are often multidimensional. When we asked home-
less youth at drop-in centers to check all of the reasons they felt were why they
became homeless, they cited: being kicked/thrown out of home (42%), leaving
home (33%), conflict with parents (26%), being evicted (16%), escaping abuse
(physical: 10%, sexual: 6%) or violence at home (8%), financial problems
(11%), released from jail (11%), their mental health (9%), their substance
use (12%), aged out of foster care (7%), and sexual identity (4%) (Winetrobe
et al., 2017). It is important to remember that one reason could be the
direct/indirect consequence of another reason(s).

Thrown Out of Home/Run Away from Home

Research on homeless youth thoroughly documented the fact that most
youth come from dysfunctional family backgrounds where households are
consumed by parental substance abuse, neglect, and violence (e.g., Whitbeck
& Hoyt, 1999). Our recent work documented that 44% of youth reported
being physically abused as children, and 39% witnessed family violence
in their homes (Petering, Rice, Rhoades, & Winetrobe, 2014). One multi-
city study found that the majority (80%) of homeless youth were victims
of physical abuse as children (Bender, Brown, Thompson, Ferguson, &
Langenderfer, 2015). Sexual violence is an all too common experience in the
lives of youth who become homeless. Our recent work has shown that 34%
of female youth and 16% of male youth experienced sexual violence prior
to becoming homeless (Harris, Rice, Rhoades, Winetrobe, & Wenzel, 2017).
Many homeless youth are either thrown out of their home or they run away,
making the difficult decision that life on the streets may be safer than life at
home. Indeed, our recent work found that 42% of homeless youth reported
being thrown out or kicked out of home, and 33% reported running away from
home (Winetrobe et al., 2017).

Aging Out of Foster Care

Many youth who experienced violent or neglectful family situations were
at some point removed from those homes by the child welfare system. Our
drop-in center–based samples show that approximately 35% of the youth
we encountered had a history of foster care system involvement (Yoshioka-
Maxwell & Rice, 2017). These youth all experienced being removed from their
family homes by the state because of reports of abuse or neglect perpetrated

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 20 The Causes and Consequences of Youth Homelessness

against them. When children are removed from their families by the child
welfare system, they are placed with relatives, non-kin who may or may not
adopt the child, or into group homes. These children are the responsibility of
the state until they reach adulthood, which is typically defined as age 18, but
has been recently extended to age 21 in some local jurisdictions, and can vary
state to state. Unfortunately, for many children, these removals from home do
not always lead to stable lives, as research has shown that between 11% and
36% of these children will experience homelessness, as youth who age out of
foster care may not have a stable place to live following their last placement
(Dworsky, Dillman, Dion, Coffee-Borden, & Rosenau, 2012).

Sexual- and Gender-Minority Youth

Forty-three percent of homeless youth accessing drop-in centers identify as
lesbian, gay, bisexual, or transgender (LGBT) (Durso & Gates, 2012). LGBT
youth face the usual turmoil of adolescence with the added burdens of social
stigma and discrimination because of their sexual and/or gender identity, often
without the support of peers, family, school, or social services (D’Augelli,
Hershberger, & Pilkington, 1998; Savin-Williams, 1994). The consequences of
this stigma, rejection, and isolation are many, including increased risk of sub-
stance use, depression, suicide, sexual risk taking, and homelessness (Savin-
Williams, 1994). Recent data collected by our team (2012) in conjunction
with the CDC’s Youth Risk Behavior Survey, indicated that 37% of LGBT-
identifying youth experienced at least one night of homelessness each year.
Because of their sexual or gender identity, youth are thrown/kicked out of their
home, or run away because of their family’s rejection and/or abuse (Durso &
Gates, 2012). Many LGBT youth run away from home because they perceive
their community and/or family to be rejecting of their sense of self and/or to
escape abuse. They desire to be accepted and safe in a community, thereby
risking their life on the streets.

Travelers
A subpopulation of homeless youth known as travelers are an important part
of a homeless youth social network. Travelers are youth who migrate from city
to city by themselves or with friends after a short period of time (Um, Rice,
Rhoades, Winetrobe, & Bracken, 2016; Winetrobe et al., 2017). One-third of
homeless youth at Los Angeles drop-in centers identified as travelers and were
less likely to consider themselves to be homeless (Winetrobe et al., 2017).

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 Eric Rice and Hailey Winetrobe 21

Moreover, travelers were more likely to binge drink, use marijuana, and inject
illicit drugs (Martino et al., 2011; Um et al., 2016).
Typically, travelers stick together and are likely to be socially connected
to one another. However, travelers will often learn about a particular city’s
resources and services by meeting “local” homeless youth. We found among
Los Angeles area–drop-in center youth, that those who were connected to
a traveling youth were more likely to use substances (Um et al., 2016).
Because of their short residencies in a given city or at a particular drop-in
center, it becomes more difficult to strategize appropriate social network–
based interventions. Travelers contribute to the overall fluidity and evolution of
homeless youth social networks, and are a prominent segment of the population
to tap into for intervention delivery.

Consequences of Youth Homelessness

Experiences of Violence
Violence is a major consequence of life on the streets. In our recent work,
more than half of the homeless youth had been in at least one fight during the
previous year, and one-fifth sustained an injury during a fight that necessitated
medical attention (Petering, Rice, & Rhoades, 2016). Violence in the context
of romantic relationships is also prevalent. Approximately 38% of homeless
youth in romantic relationships experienced some form of intimate partner
violence in the previous year, the majority of which was bidirectional (26%);
7% reported being solely a perpetrator, and 5% reported being solely a victim
(Petering, Rhoades, Rice, & Yoshioka-Maxwell, 2017). Sexual violence is
also a major threat; 33% of girls and 26% of boys reported sexual assault
victimization while homeless (Harris et al., 2017).

Mental Health
It should not come as a surprise that exposure to this much violence both before
and after becoming homeless can take a toll on the mental health of homeless
youth. Our recent work shows about 41% of homeless youth were likely
suffering from posttraumatic stress disorder (PTSD). Reflecting the higher
rates of sexual assault, more girls than boys reported symptoms of PTSD
(Harris et al., 2017). Depression for homeless youth is often a consequence of
being a victim of assault (Lim, Rice, & Rhoades, 2016). Moreover, homeless
youth with histories of childhood abuse and assault while homeless were more

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 22 The Causes and Consequences of Youth Homelessness

likely to experience PTSD and depressive symptoms (Bender et al., 2015).

In a cross-sectional, multicity study of homeless youth in the United States,
Bender et al. (2015) found that 23% of the homeless youth they surveyed
met diagnostic criteria for PTSD, and 31% for major depression. These are
akin to rates found in a multisite, diagnostic study of homeless and runaway
adolescents (Whitbeck, Johnson, Hoyt, & Cauce, 2004), which compared
depression and PTSD prevalence rates among homeless youth to housed peers.
Whitbeck et al. (2004) reported a lifetime prevalence of depression among
homeless youth that was twice the rate in the same-age general population;
PTSD prevalence among homeless youth was seven times the rate of their
housed peers. Perhaps even more disturbing are the high rates of suicidal
thinking and suicide attempts. Our most recent work found homeless youth’s
suicidal ideation and suicide attempts during the past year were 17% and 7%,
respectively (Barr, Fulginiti, Rhoades, & Rice, 2017).

Substance Use/Abuse
Substance use is quite prevalent among homeless youth. Past-month substance
use among homeless youth was quite high; 50% of homeless youth at
drop-in centers in Los Angeles reported binge drinking (i.e., five or more
alcoholic drinks in a row), 77% reported using marijuana, 25% reported using
methamphetamine, 18% cocaine, 10% crack cocaine, and 12% reported using
heroin (Rhoades, Winetrobe, & Rice, 2014). Past research demonstrated that
homeless youth use substances to help them stay awake to protect themselves
and their property (Christiani, Hudson, Nyamathi, Mutere, & Sweat, 2008)
and to cope with mental distress and suicidality (Christiani et al., 2008; Kidd
& Carroll, 2007). Additionally, homeless youth with foster care experiences
(Hudson & Nandy, 2012; Nyamathi, Hudson, Greengold, & Leake, 2012),
who were victims of abuse (Tyler & Melander, 2010), who are travelers,
who engaged in survival sex (Walls & Bell, 2011), and who started using
substances at a young age (Salomonsen-Sautel et al., 2008) were all more
likely to be using while homeless. And the longer they were homeless, the
more likely they were to use (Thompson, Bender, Ferguson, & Kim, 2015).
Only 12% of homeless youth cited their own substance use as a contributing
factor to their homelessness (Winetrobe et al., 2017). Some homeless youth
come from families where substance use and abuse is common (McMorris,
Tyler, Whitbeck, & Hoyt, 2002; Robertson, Koegel, & Ferguson, 1989), and
thus large numbers began using drugs and alcohol at relatively early ages
(Robertson et al., 1989). Our recent work has shown that 37% reported

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 Eric Rice and Hailey Winetrobe 23

using alcohol and 35% reported using marijuana before age 13 (Rhoades
et al., 2014). Mirroring the recent national trends of prescription opioid drug
abuse (Compton, Jones, & Baldwin, 2016), 22% reported prescription drug
misuse in the prior month, with opioids being the most commonly misused
prescription drug category (Rhoades et al., 2014).

Contraception and Pregnancy

Many homeless youth are also at higher risk of pregnancy than most youth
their age, and pregnancy among homeless youth may exacerbate youths’
homelessness duration (Haley, Roy, Leclerc, Boudreau, & Boivin, 2004;
Thompson, Bender, Lewis, & Watkins, 2008). Sixty-two percent of homeless
girls reported ever having been pregnant, and 44% of males reported having
ever impregnated someone (Winetrobe et al., 2013). One-fifth of Los Angeles
area homeless youth reported having had a child, while 7% of homeless youth
reported having their child(ren) stay with them (Rice et al., 2013).
Access to condoms and other contraception at drop-in centers, shelters, and
other service agencies may increase effective contraception use, though only
36% of homeless youth reported using a condom at last vaginal sex (Winetrobe
et al., 2013). In our recent study, one-third of homeless youth at drop-in centers
reported accessing condoms within the previous month, and female homeless
youth who accessed condoms within the previous month were less likely to
have unprotected sex (Barman-Adhikari, Hsu, Begun, Portillo, & Rice, 2017).
Additionally, male and female homeless youth who thought their street-based
peers were using condoms were less likely themselves to engage in condomless
sex (Barman-Adhikari et al., 2017).
Pregnancy among homeless youth may be a result of (1) sexual assault, (2)
engaging in survival sex (i.e., trading sex for food, money, a place to stay),
(3) a desire to become pregnant, or (4) an unintentional pregnancy because of
lack of contraceptive use altogether or use of ineffective methods (Winetrobe
et al., 2013). Female homeless youth use condoms because they are available
more often and they also protect against sexually transmitted infections (STIs)
(Ensign, 2001). Lack of access to other contraceptive methods may cause
homeless youth to rely on other, less effective forms of contraception (e.g.,
withdrawal, rhythm method), and thus increase their chances of pregnancy.
Perhaps the greatest threat to the sexual health of homeless youth, however,
is not pregnancy, but rather the risk of HIV and other STIs. Rates of HIV have
been documented to be as high as 11.5% among homeless youth (Pfeifer &
Oliver, 1997). Homeless youth are at extremely high risk for contracting HIV

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 24 The Causes and Consequences of Youth Homelessness

because of their risk and survival behavior, including sharing needles in the
context of injection drug use, sex without condoms, and sex without condoms
in the context of survival sex. Based on the data we collected as part of our
National Institute of Mental Health-funded R01 study (Rice R01 MH093336),
from 1,046 homeless youth between 2011 and 2013 in the Los Angeles area,
we calculated the proportion of homeless youth who fall into different HIV
risk categories: 6% were transgender, 24% were young men who have sex with
men (YMSM), 16% injected drugs, 5% were YMSM who inject drugs, 16%
were engaged in sex work, 60% had incarceration experiences, 6% were female
youth who had partners who were YMSM or injection drug users, and 3%
were HIV-positive persons (who thereby are at increased risk for transmitting
the virus).
There is a fundamental need to identify new cases of HIV among homeless
youth and link them to HIV treatment, not only for their own health but to
also limit the spread of HIV. It is encouraging that 85% of homeless youth in
Los Angeles County reported testing for HIV ever in their lifetimes, and 47%
had a HIV test in the past 3 months (Ober, Martino, Ewing, & Tucker, 2012).
The rates of current HIV testing, however, reveal key gaps in testing and
treatment, especially for racial/ethnic minority and sexual minority youth.
From our data we found that 33% of African American youth and 36% of
Latino youth had not been tested for HIV in the past six months. Among
female youth, 28% of African American female homeless youth and 23% of
homeless Latinas had not been tested in the past 6 months. In addition, 19%
of African American YMSM and 27% of Latino YMSM had not tested in the
past 6 months (Rice, 2013). The Centers for Disease Control and Prevention
recommend at least annual testing for HIV, and more frequent (i.e., every 3-
6 months) among higher-risk groups, such as YMSM (Centers for Disease
Control and Prevention, 2017).
Homeless youth are not engaged in regular testing, in part because they are
not well engaged in health care service utilization of any kind (Winetrobe,
Rice, Rhoades, & Milburn, 2016). This arises due to feeling discriminated
against, disrespected, and stigmatized because of their homeless state (Chris-
tiani et al., 2008; Hudson et al., 2010; Martins, 2008; Wen, Hudak, &
Hwang, 2007), and distrust with health care and other social service providers
(Klein et al., 2000; Kurtz, Lindsey, Jarvis, & Nackerud, 2000). Moreover,
homeless youth often lack adequate health insurance, transportation, and
knowledge of how to navigate complicated health systems in order to receive
care (Hudson et al., 2010). Homeless youth are less likely than other youth to
seek help from traditional health care providers, choosing instead to rely on
informal sources such as friends or relatives (Ensign & Gittelsohn, 1998) and

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 Eric Rice and Hailey Winetrobe 25

the internet (Barman-Adhikari & Rice, 2011). In fact, we found that 61% of
homeless youth reported ever looking for general health information online,
and 23% searched specifically online for where to get an HIV test (Barman-
Adhikari & Rice, 2011). It goes without saying that there is no guarantee that
the health information found online is accurate; it requires specific assessment
of the source.

The Importance of Social Networks

In Western developed nations, the normal developmental trajectory for ado-
lescents has been well documented; from early to emerging adulthood, young
people increasingly move toward independence and autonomy with the support
and the relative influence of families, friends, and social institutions as social-
izing agents shifting over time (Arnett, 2000, 2001). By early adolescence,
the role of family changes as the importance of peers and friends, as well
as that of teachers and others in institutional settings, increases (Bauman,
Carver, & Gleiter, 2001; Berndt, 1979). For homeless youth, being thrown out
of home or running away from home exaggerates the process of engagement
with peers and disengagement with family and further disenfranchises youth
from their connections to pro-social institutions, such as school. The best
developmentally focused models of this process are the Risk Amplification
Model (RAM) (Whitbeck, Hoyt, & Yoder, 1999) and its augmentation, the
Risk Amplification and Abatement Model (RAAM) (Milburn et al., 2009).
RAM asserted that the peer networks of homeless youth on the streets
are largely comprised of other homeless adolescents, many of whom come
from problematic backgrounds and engage in risky and/or deviant behavior
(Whitbeck et al., 1999). When social networks are comprised largely of other
deviant youth, the risks associated with living on the streets are magnified
for individuals operating within those networks (e.g., McMorris et al., 2002;
Rice, Milburn, & Monro, 2011). Negative peer influences in the street-based
networks of homeless youth can be seen cutting across a wide spectrum
of risk-taking behaviors, including violence (Petering et al., 2016), mental
health (Fulginiti, Rice, Hsu, Rhoades, & Winetrobe, 2016), substance use
(Barman-Adhikari, Rice, Winetrobe, & Petering, 2015; Yoshioka-Maxwell &
Rice, 2017), and sexual risk-taking (Barman-Adhikari et al., 2017).
RAAM extended this thinking by recognizing that homeless youth networks
are also a major source of resilience (Milburn et al., 2009; Rice, Milburn, &
Rotheram-Borus, 2007). Research showed that the social networks of homeless
youth were more heterogeneous (Johnson, Whitbeck, & Hoyt, 2005) than

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 26 The Causes and Consequences of Youth Homelessness

previously thought (Ennett, Bailey, & Federman, 1999; Whitbeck & Hoyt,
1999). Networks are not just limited to homeless peers, but also include
relationships from youths’ home communities, including relatives (Johnson
et al., 2005; Rice et al., 2011), friends (Rice, 2010; Rice, Kurzban, &
Ray, 2012), and non-familial adults (Rice et al., 2007; Rice et al., 2011).
Furthermore, various dimensions and characteristics of heterogeneous social
networks have differential effects on homeless youth. Homeless youths’ social
networks can and do include pro-social relationships that contribute to positive
outcomes for the individual.
These pro-social relationships reduce homeless youths’ risk behaviors
and antisocial behaviors (Rice, 2010; Rice et al., 2007; Rice et al., 2011).
This potential for positive influence can be seen across a wide spectrum
of issues including improved mental health (Rice, Kurzban, et al., 2012),
reduced substance use (Rice et al., 2011), and reduced sexual risk-taking
(Rice, 2010; Rice, Monro, Barman-Adhikari, & Young, 2010). Most of these
studies, however, point to peers from home and family as the major source of
this pro-social influence. Fortunately for our interests in peer-led prevention
interventions, this work also uncovered subtle, previously ignored positive
impacts of homeless peers, particularly with respect to sexual health. While
youth who were more deeply embedded in social networks of other street
youth were less likely to be using condoms (Rice, Barman-Adhikari, Milburn,
& Monro, 2012), being connected to condom-using peers who were also
homeless was found to be associated with increased condom use (Rice, 2010).

HIV Prevention for Homeless Youth

At present, there is a dearth in peer-led interventions to improve health
behaviors among homeless youth. A few limited interventions and evaluations
that have been conducted produced mixed results. One study found that
peer-based models were effective in increasing homeless youths’ substance
use knowledge and preventive behaviors (Fors & Jarvis, 1995), while
another study found that although a peer-led intervention was effective in
increasing HIV-related knowledge, it was ineffective in reducing sex risk
and substance use behaviors among homeless youth (Booth, Zhang, &
Kwiatkowski, 1999). Another program, YouthCare, included homeless youth
staff in their outreach and HIV prevention education efforts; the result was an
increase in HIV testing among homeless youth (Tenner, Trevithick, Wagner, &
Burch, 1998). In addition to producing inconclusive results, these interventions
are quite old.

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 Eric Rice and Hailey Winetrobe 27

Our recent research investigating homeless youths’ social networks has

shown that peers do play an important role in the HIV risk-taking and
protective behaviors of homeless youth (Rice, 2010; Rice, Tulbert, Cederbaum,
Barman Adhikari, & Milburn, 2012), and it has been suggested that a peer-
based model for HIV prevention be developed for homeless youth (Rice,
Tulbert, et al., 2012; Tenner et al., 1998). Peer-based models, such as Kelly’s
(1997) popular opinion leader model and Latkin et al.’s (2003) Self-Help in
Eliminating Life-Threatening Diseases (SHIELD), may be particularly impact-
ful strategies for homeless youth, due to the importance of social network
engagement both to HIV risk-taking and protective behaviors. Creating an
HIV prevention program that engages homeless youth peers as leaders has
been shown to be acceptable (Rice, Tulbert, et al., 2012), but no work has
demonstrated its feasibility to date.
There are many challenges associated with creating such intervention
programs. Indeed, while peer-based models have been found to be effective
in behavior change for the prevention of HIV in many contexts (Medley,
Kennedy, O’Reilly, & Sweat, 2009), there have been some notable failures
(Booth et al., 1999; The NIMH Collaborative HIV/STD Prevention Trial
Group, 2010). Peer-based models are complex, involving not only careful, cul-
turally tailored, norm-changing messaging but also the effective engagement
of a diverse set of network contexts.
Identifying the right peer leaders is equally important to the intervention
curriculum/program in order for the impact of a peer-based intervention to be
successful. One method for choosing the right peer leaders is to select those
who are in structurally advantageous positions in a social network (Schneider,
Zhou, & Laumann, 2015; Valente & Pumpuang, 2007). Another method, as
demonstrated in our prior work, is to select youth identified by agency staff
and ethnographic methods who are pro-social in their behaviors to be peer
leaders (Rice, Tulbert, et al., 2012). From a structural perspective, however,
selecting peer leaders based on individual pro-social behaviors is an arbitrary
selection process, and such an approach may result in selecting a set of peer
leaders who are not well connected to the rest of the population (Schneider
et al., 2015; Yadav et al., 2015).
Peer-led models for HIV prevention have rarely taken advantage of social
network analytic methods to augment the implementation of these programs,
despite the fact that networks are a crucial component of the Diffusion of
Innovations Theory (Rogers, 1995) on which these interventions were built
(e.g., Kelly et al., 1997; Latkin et al., 2003). As Schneider and colleagues
(2015) suggested, perhaps one reason for the unsuccessful implementation
of peer-based models is that the right persons were not engaged as the

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 28 The Causes and Consequences of Youth Homelessness

peer leaders. Some researchers suggest that those with high degree centrality
(i.e., persons with the highest number of connections) or high betweenness
(i.e., persons who bridge the most other pairs of actors in a network)
could be viable solutions (Schneider et al., 2015; Valente & Pumpuang,
2007).

Why We Need Artificial Intelligence

We need artificial intelligence (AI) for two reasons. First, the static network
solutions proposed by others (Schneider et al., 2015; Valente & Pumpuang,
2007) have limitations, especially among transient and frequently changing
social networks, that dynamic selection algorithms can overcome. Algorithms
that attempt to maximize overall potential influence are theoretically more
likely to influence more people than static solutions that focus on particular
structural characteristics of people in a network (see the simulation results in
the chapters that follow). This becomes increasingly important as the size and
complexity of the social networks increase. In the case of the homeless youth
networks we worked with previously, we were often attempting to impact
networks that had 150 to 300 youth at a particular time, who had several
hundred connections among themselves (Barman-Adhikari, Begun, Rice,
Yoshioka-Maxwell, & Perez-Portillo, 2016; Um, Rice, Rhoades, Winetrobe,
& Bracken, 2016; Yoshioka-Maxwell & Rice, 2017).
Second, homeless youth networks are not only complex, but the instability
of homeless youth social ties necessitates that we consider the uncertainty
present in these networks. In our prior work, we attempted to interview the
entire population of youth who accessed services over a one-month period
at each of our two partner drop-in centers: My Friend’s Place in Hollywood,
CA and Safe Place for Youth in Venice, CA. We repeated these population,
panel surveys every 6 months for 18 months. We found in Hollywood that,
on average, 43% of the population remained stable over a six-month period of
time (44% were included in panel 2 who were seen in panel 1, 42% who were
included in panel 3 were seen in panel 2), and in Venice Beach, only 16% of
the population remained stable over six months (18% stability to panel 2, 14%
stability to panel 3) (Rice, 2013).
Homeless youth social networks are continually changing. New youth are
constantly running away from home or being thrown out of their home and
entering into street life. Youth are also exiting street life, finding either positive
outcomes such as stable housing or negative outcomes such as becoming
incarcerated. Some youth move to another neighborhood or city, and thus

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 Eric Rice and Hailey Winetrobe 29

exit the local social network. Additionally, a subpopulation of travelers may

integrate into a social network of homeless youth for a short period of time
before moving on to another city. Such circumstances, along with changing
relationships between homeless youth in networks, affect the fracture or
formation of connections between homeless youth who stay in the network
over time (Rice, 2013). To make matters even more complicated, the collection
of network data from homeless youth in community settings always involves
a certain amount of error (Petering et al., 2016; Rice et al., 2014; Yoshioka-
Maxwell & Rice, 2017). Youth can forget whom they communicated with
recently, and we as researchers can have difficulty identifying all the persons
who youth may describe in their social networks. (See Brewer, 2000, for a
review of these methodological issues.)
There are a large number of methods for collecting social network data.
One of the most common field techniques is to utilize name generators: a
question or a series of questions designed to elicit the naming of relevant
social connections along some specified criterion (Campbell & Lee, 1991;
Marsden, 2011). Typically, name generators are free-recall name solicitors.
Respondents are given a prompt that defines some criterion, for instance, a
category of persons such as “family,” “friends,” or types of social exchange
relationships (e.g., “Who do you turn to for advice or support?”). Then respon-
dents are asked to list as many people as they can. In some cases, an upper
limit is given to the number of names that can be elicited. Most solutions to
network recall begin with the understanding that a single-item free-recall name
generator will be most subject to recall bias. Brewer and colleagues (Brewer
et al., 2000; Brewer, Garrett, & Rinaldi, 2002) have extensively reviewed
the topic and suggested and tested several viable solutions to the problem,
including nonspecific prompting, reading back lists of names, semantic cues,
multiple elicitation questions, and re-interviewing.
For our work, we chose to follow the solution of multiple elicitation
questions. The following was read: “Think about the last month. Who have
you interacted with? These can be people you interacted with in person, on
the phone, or through the internet. These might be friends; family; people you
hang out with/chill with/kick it with/ have conversations with; people you party
with – use drugs or alcohol; boyfriend/girlfriend; people you are having sex
with; baby mama/baby daddy; case worker; people from school; people from
work; old friends from home; people you talk to (on the phone, by email);
people from where you are staying (squatting with); people you see at this
agency; other people you know from the street.” Interviewers paused between
each prompt to allow youth to nominate additional social connections before
proceeding to the next prompt.

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 30 The Causes and Consequences of Youth Homelessness

After youth finished nominating persons, attributes of each social con-

nection were then collected, including first and last names, aliases, age,
gender, race/ethnicity, visible tattoos, and whether the nominee was a client
of the agency. In the paper-and-pencil version, attributes were entered into a
spreadsheet on a laptop by the interviewer; in the iPad app version, responses
were entered directly into the app by the interviewer.
A sociomatrix (or graph) was created linking participants in the sample.
A directed tie from participant i to participant j was recorded if participant
i nominated participant j in his or her personal network. Although directed
network information was collected, the network information is intrinsically
undirected (a person who interacts with another) and was treated in this
fashion for the current analysis. The initial matrix creation depended on the
information provided by each youth about their personal network members,
and subsequent “matches” were made with other youth in the sample. Matches
were based on name, alias, race/ethnicity, gender, approximate age, tattoos,
and agency attendance. Two independent reviewers, who were part of the team
collecting the field data, made match decisions for all alters who were between
the ages of 13 and 39 years and not identified as agency staff.
A series of decision rules was derived from information that was available
to make the match determination. Decisions were based on a series of
algorithms that included (a) interviewer and recruiter field knowledge (through
the compilation of field notes following each data collection period), (b) how
well the ego knew the alter (i.e., relative, romantic partner, needle sharer,
known for at least one year) and whether the alter was identified as a client,
or (c) via an Access database and form that formulaically paired possible
matches based on names, visible tattoos, and demographic characteristics. If
two distinct youths were similar on all information, they were considered for
the purposes of this research to be the same individuals. If two distinct youths
matched on all information, presence of a third common tie in each personal
network was used to assign the tie/edge. The independent reviewers’ decisions
were compared for agreement. Discrepant matches were discussed with the
independent reviewers and a third reviewer who also served as an interviewer
and recruiter during the data collection. Group consensus led to final match
decisions.
It is not difficult to see that this protocol is not only extremely labor
intensive but also subject to human error at several stages. Error is introduced
by the memory and thus accurate collection of names and descriptors from
youth (Brewer, 2000). Then, more error is introduced by the matching of
youth to one another by research staff, particularly as not all identifying
information collected from a given youth is sufficiently detailed or accurate

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 Eric Rice and Hailey Winetrobe 31

in order to always disambiguate two different youths with similar names

and appearances (e.g., two 20-year-old white youths named Cody who do
not have distinctive visible tattoos). Our procedures involved multiple coders
and interrater reliability checks, but we know this data is not a perfect
representation of the social network of youth at a given time because of these
unavoidable sources of error.
Thus, we need a peer selection procedure within large and complex
networks of homeless youth (or other similarly composed networks), which
can accurately account for the uncertainties we know to be present in this data.
Enter AI.

References
Arnett, J. J. (2000). Emerging adulthood: A theory of development from the late teens
through the twenties. American Psychologist, 55(5), 469
Arnett, J. J. (2001). Conceptions of the transition to adulthood: Perspectives from
adolescence through midlife. Journal of Adult Development, 8(2), 133–143.
doi:10.1023/a:1026450103225
Barman-Adhikari, A., Begun, S., Rice, E., Yoshioka-Maxwell, A., & Perez-Portillo, A.
(2016). Sociometric network structure and its association with methamphetamine
use norms among homeless youth. Social Science Research, 58, 292–308.
doi:10.1016/j.ssresearch.2016.01.004
Barman-Adhikari, A., Hsu, H.-T., Begun, S., Portillo, A. P., & Rice, E. (2017).
Condomless sex among homeless youth: The role of multidimensional social
norms and gender. AIDS and Behavior, 21(3), 688–702. doi:10.1007/s10461-016-
1624-2
Barman-Adhikari, A., & Rice, E. (2011). Sexual health information seeking online
among runaway and homeless youth. Journal of the Society for Social Work and
Research, 2(2), 89–103. doi:10.5243/jsswr.2011.5
Barman-Adhikari, A., Rice, E., Winetrobe, H., & Petering, R. (2015). Social network
correlates of methamphetamine, heroin, and cocaine use in a sociometric network
of homeless youth. Journal of the Society for Social Work and Research, 6(3),
433–457. doi:10.1086/682709
Barr, N., Fulginiti, A., Rhoades, H., & Rice, E. (2017). Can better emotion regulation
protect against suicidality in traumatized homeless youth? Archives of Suicide
Research, 21(3), 490–501. doi:10.1080/13811118.2016.1224989
Bauman, K. E., Carver, K., & Gleiter, K. (2001). Trends in parent and friend
influence during adolescence: The case of adolescent cigarette smoking. Addictive
Behaviors, 26(3), 349–361. doi:10.1016/S0306-4603(00)00110-6
Bender, K., Brown, S. M., Thompson, S. J., Ferguson, K. M., & Langenderfer, L.
(2015). Multiple victimizations before and after leaving home associated with
PTSD, depression, and substance use disorder among homeless youth. Child
Maltreatment, 20(2), 115–124. doi:10.1177/1077559514562859
Berndt, T. J. (1979). Developmental changes in conformity to peers and parents.
Developmental Psychology, 15(6), 608

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 32 The Causes and Consequences of Youth Homelessness

Booth, R. E., Zhang, Y., & Kwiatkowski, C. F. (1999). The challenge of changing
drug and sex risk behaviors of runaway and homeless adolescents. Child Abuse
& Neglect, 23(12), 1295–1306. doi:10.1016/S0145-2134(99)00090-3
Brewer, D. D. (2000). Forgetting in the recall-based elicitation of personal and social
networks. Social Networks, 22(1), 29–43. doi:10.1016/S0378-8733(99)00017-9
Brewer, D. D., Garrett, S. B., & Rinaldi, G. (2002). Free-listed items are effective cues
for eliciting additional items in semantic domains. Applied Cognitive Psychology,
16(3), 343–358. doi:10.1002/acp.797
Campbell, K. E., & Lee, B. A. (1991). Name generators in surveys of personal
networks. Social Networks, 13(3), 203–221. doi:10.1016/0378-8733(91)90006-F
Centers for Disease Control and Prevention. (2017, May 18). HIV testing. Retrieved
from www.cdc.gov/hiv/testing/index.html
Christiani, A., Hudson, A. L., Nyamathi, A., Mutere, M., & Sweat, J. (2008). Attitudes
of homeless and drug-using youth regarding barriers and facilitators in delivery
of quality and culturally sensitive health care. Journal of Child and Adolescent
Psychiatric Nursing, 21(3), 154–163. doi:10.1111/j.1744-6171.2008.00139.x
Compton, W. M., Jones, C. M., & Baldwin, G. T. (2016). Relationship between
nonmedical prescription-opioid use and heroin use. New England Journal of
Medicine, 374(2), 154–163. doi:10.1056/NEJMra1508490
D’Augelli, A. R., Hershberger, S. L., & Pilkington, N. W. (1998). Lesbian, gay,
and bisexual youth and their families: Disclosure of sexual orientation and its
consequences. American Journal of Orthopsychiatry, 68(3), 361
De Rosa, C. J., Montgomery, S. B., Kipke, M. D., Iverson, E., Ma, J. L., & Unger, J. B.
(1999). Service utilization among homeless and runaway youth in Los Angeles,
California: Rates and reasons. Journal of Adolescent Health, 24(3), 190–200.
doi:10.1016/S1054-139X(99)00040-3
Durso, L. E., & Gates, G. J. (2012). Serving our youth: Findings from a
national survey of service providers working with lesbian, gay, bisexual,
and transgender youth who are homeless or at risk of becoming homeless.
Retrieved from: http://williamsinstitute.law.ucla.edu/wp-content/uploads/Durso-
Gates-LGBT-Homeless-Youth-Survey-July-2012.pdf
Dworsky, A., Dillman, K.-N., Dion, R. M., Coffee-Borden, B., & Rosenau, M. (2012).
Housing for youth aging out of foster care: A review of the literature and
program typology. Available at: https://papers.ssrn.com/sol3/papers.cfm?abstract_
id=2112278
Ennett, S. T., Bailey, S. L., & Federman, E. B. (1999). Social network characteristics
associated with risky behaviors among runaway and homeless youth. Journal of
Health and Social Behavior, 40(1), 63–78. doi:10.2307/2676379
Ensign, J. (2001). Reproductive health of homeless adolescent women in Seattle, Wash-
ington, USA. Women & Health, 31(2–3), 133–151. doi:10.1300/J013v31n02_07
Ensign, J., & Gittelsohn, J. (1998). Health and access to care: Perspectives of homeless
youth in Baltimore City, U.S.A. Social Science & Medicine, 47(12), 2087–2099.
doi:10.1016/S0277-9536(98)00273-1
Fors, S. W., & Jarvis, S. (1995). Evaluation of a peer-led drug abuse risk reduction
project for runaway/homeless youths. Journal of Drug Education, 25(4), 321–333.
doi:10.2190/TU92-LX8W-G7FD-9LEM
Fulginiti, A., Rice, E., Hsu, H.-T., Rhoades, H., & Winetrobe, H. (2016). Risky
integration. Crisis: The Journal of Crisis Intervention and Suicide Prevention,
37(3), 184–193. doi:10.1027/0227-5910/a000374

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 Eric Rice and Hailey Winetrobe 33

Haley, N., Roy, E., Leclerc, P., Boudreau, J. F., & Boivin, J. F. (2004). Characteristics
of adolescent street youth with a history of pregnancy. Journal of Pediatric and
Adolescent Gynecology, 17(5), 313–320. doi:10.1016/j.jpag.2004.06.006
Harris, T., Rice, E., Rhoades, H., Winetrobe, H., & Wenzel, S. (2017). Gen-
der differences in the path from sexual victimization to HIV risk behav-
ior among homeless youth. Journal of Child Sexual Abuse, 26(3), 334–351.
doi:10.1080/10538712.2017.1287146
Hudson, A., & Nandy, K. (2012). Comparisons of substance abuse, high-risk sex-
ual behavior and depressive symptoms among homeless youth with and with-
out a history of foster care placement. Contemporary Nurse, 42(2), 178–186.
doi:10.5172/conu.2012.42.2.178
Hudson, A., Nyamathi, A., Greengold, B., Slagle, A., Koniak-Griffin, D., Khalilifard,
F., & Getzoff, D. (2010). Health-seeking challenges among homeless youth.
Nursing Research, 59(3), 212
Johnson, K. D., Whitbeck, L. B., & Hoyt, D. R. (2005). Predictors of social network
composition among homeless and runaway adolescents. Journal of Adolescence,
28(2), 231–248. doi:10.1016/j.adolescence.2005.02.005
Kelly, J. A., Murphy, D. A., Sikkema, K. J., McAuliffe, T. L., Roffman, R. A., Solomon,
L. J., . . . Kalichman, S. C. (1997). Randomised, controlled, community-level HIV-
prevention intervention for sexual-risk behaviour among homosexual men in US
cities. The Lancet, 350(9090), 1500–1505. doi:10.1016/S0140-6736(97)07439-4
Kidd, S. A., & Carroll, M. R. (2007). Coping and suicidality among homeless youth.
Journal of Adolescence, 30(2), 283–296. doi:10.1016/j.adolescence.2006.03.002
Klein, J. D., Woods, A. H., Wilson, K. M., Prospero, M., Greene, J., & Ringwalt,
C. (2000). Homeless and runaway youths’ access to health care. Journal of
Adolescent Health, 27(5), 331–339. doi:10.1016/S1054-139X(00)00146-4
Kurtz, P. D., Lindsey, E. W., Jarvis, S., & Nackerud, L. (2000). How runaway
and homeless youth navigate troubled waters: The role of formal and infor-
mal helpers. Child and Adolescent Social Work Journal, 17(5), 381–402.
doi:10.1023/a:1007507131236
Latkin, C. A., Sherman, S., & Knowlton, A. (2003). HIV prevention among drug users:
Outcome of a network-oriented peer outreach intervention. Health Psychology,
22(4), 332
Lim, C., Rice, E., & Rhoades, H. (2016). Depressive symptoms and their asso-
ciation with adverse environmental factors and substance use in runaway
and homeless youths. Journal of Research on Adolescence, 26(3), 403–417.
doi:10.1111/jora.12200
Los Angeles Homeless Services Authority. (2018). 2017 Greater Los Angeles
Homeless Count Results: Los Angeles County and Continuum of Care. Retrieved
from www.lahsa.org/documents?id=1873-2017-greater-los-angeles-homeless-
count-presentation-los-angeles-county-and-continuum-of-care.pdf
Marsden, P. V. (2011). Survey methods for network data. The SAGE Handbook of Social
Network Analysis, 25, 370–388
Martino, S. C., Tucker, J. S., Ryan, G., Wenzel, S. L., Golinelli, D., & Munjas,
B. (2011). Increased substance use and risky sexual behavior among migratory
homeless youth: Exploring the role of social network composition. Journal of
Youth and Adolescence, 40(12), 1634–1648. doi:10.1007/s10964-011-9646-6
Martins, D. C. (2008). Experiences of homeless people in the health care delivery
system: A descriptive phenomenological study. Public Health Nursing, 25(5),
420–430. doi:10.1111/j.1525-1446.2008.00726.x

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 34 The Causes and Consequences of Youth Homelessness

McMorris, B. J., Tyler, K. A., Whitbeck, L. B., & Hoyt, D. R. (2002). Famil-
ial and “on-the-street” risk factors associated with alcohol use among home-
less and runaway adolescents. Journal of Studies on Alcohol, 63(1), 34–43.
doi:10.15288/jsa.2002.63.34
Medley, A., Kennedy, C., O’Reilly, K., & Sweat, M. (2009). Effectiveness of peer
education interventions for HIV prevention in developing countries: A systematic
review and meta-analysis. AIDS Education and Prevention, 21(3), 181–206.
doi:10.1521/aeap.2009.21.3.181
Milburn, N. G., Rice, E., Rotheram-Borus, M. J., Mallett, S., Rosenthal, D., Batterham,
P., . . . Duan, N. (2009). Adolescents exiting homelessness over two years: The risk
amplification and abatement model. Journal of Research on Adolescence, 19(4),
762–785. doi:10.1111/j.1532-7795.2009.00610.x
Nyamathi, A., Hudson, A., Greengold, B., & Leake, B. (2012). Characteristics of
homeless youth who use cocaine and methamphetamine. The American Journal
on Addictions, 21(3), 243–249. doi:10.1111/j.1521-0391.2012.00233.x
Ober, A. J., Martino, S. C., Ewing, B., & Tucker, J. S. (2012). If you provide the
test, they will take it: Factors associated with HIV/STI testing in a representative
sample of homeless youth in Los Angeles. AIDS Education and Prevention, 24(4),
350–362. doi:10.1521/aeap.2012.24.4.350
Petering, R., Rhoades, H., Rice, E., & Yoshioka-Maxwell, A. (2017). Bidirectional
intimate partner violence and drug use among homeless youth. Journal of
Interpersonal Violence, 32(14), 2209–2217. doi:10.1177/0886260515593298
Petering, R., Rice, E., & Rhoades, H. (2016). Violence in the social net-
works of homeless youths. Journal of Adolescent Research, 31(5), 582–605.
doi:10.1177/0743558415600073
Petering, R., Rice, E., Rhoades, H., & Winetrobe, H. (2014). The social networks of
homeless youth experiencing intimate partner violence. Journal of Interpersonal
Violence, 29(12), 2172–2191. doi:10.1177/0886260513516864
Pfeifer, R. W., & Oliver, J. (1997). A study of HIV seroprevalence in a group of
homeless youth in Hollywood, California. Journal of Adolescent Health, 20(5),
339–342. doi:10.1016/S1054-139X(97)00038-4
Rhoades, H., Winetrobe, H., & Rice, E. (2014). Prescription drug misuse
among homeless youth. Drug and Alcohol Dependence, 138, 229–233.
doi:10.1016/j.drugalcdep.2014.02.011
Rice, E. (2010). The positive role of social networks and social networking technology
in the condom-using behaviors of homeless young people. Public Health Reports,
125(4), 588–595. doi:10.1177/003335491012500414
Rice, E. (2013). Substantial differences in risk behaviors of homeless youth in two
large networks of homeless youth in Los Angeles. Paper presented at the American
Public Health Association Annual Meeting, Boston
Rice, E., Barman-Adhikari, A., Milburn, N. G., & Monro, W. (2012). Position-specific
HIV risk in a large network of homeless youths. American Journal of Public
Health, 102(1), 141–147. doi:10.2105/ajph.2011.300295
Rice, E., Fulginiti, A., Winetrobe, H., Montoya, J., Plant, A., & Kordic, T. (2012).
Sexuality and homelessness in Los Angeles public schools. American Journal of
Public Health, 102(2), 200–201. doi:10.2105/ajph.2011.300411
Rice, E., Holloway, I. W., Barman-Adhikari, A., Fuentes, D., Brown, C. H., & Palinkas,
L. A. (2014). A mixed methods approach to network data collection. Field
Methods, 26(3), 252–268. doi:10.1177/1525822X13518168

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 Eric Rice and Hailey Winetrobe 35

Rice, E., Kurzban, S., & Ray, D. (2012). Homeless but connected: The role of
heterogeneous social network ties and social networking technology in the mental
health outcomes of street-living adolescents. Community Mental Health Journal,
48(6), 692–698. doi:10.1007/s10597-011-9462-1
Rice, E., Milburn, N. G., & Monro, W. (2011). Social networking technology, social
network composition, and reductions in substance use among homeless adoles-
cents. Prevention Science, 12(1), 80–88. doi:10.1007/s11121-010-0191-4
Rice, E., Milburn, N. G., & Rotheram-Borus, M. J. (2007). Pro-social and
problematic social network influences on HIV/AIDS risk behaviours among
newly homeless youth in Los Angeles. AIDS Care, 19(5), 697–704.
doi:10.1080/09540120601087038
Rice, E., Monro, W., Barman-Adhikari, A., & Young, S. D. (2010). Internet use, social
networking, and HIV/AIDS risk for homeless adolescents. Journal of Adolescent
Health, 47(6), 610–613. doi:10.1016/j.jadohealth.2010.04.016
Rice, E., Tulbert, E., Cederbaum, J., Barman Adhikari, A., & Milburn, N. G. (2012).
Mobilizing homeless youth for HIV prevention: A social network analysis of the
acceptability of a face-to-face and online social networking intervention. Health
Education Research, 27(2), 226–236. doi:10.1093/her/cyr113
Rice, E., Winetrobe, H., & Rhoades, H. (2013). Hollywood homeless youth point-
in-time estimate project: An innovative method for enumerating unaccompa-
nied homeless youth. Retrieved from http://hhyp.org/wp-content/uploads/2013/02/
HHYP_Point-in-Time_Brief_5.pdf
Ringwalt, C. L., Greene, J. M., Robertson, M., & McPheeters, M. (1998). The
prevalence of homelessness among adolescents in the United States. American
Journal of Public Health, 88(9), 1325–1329. doi:10.2105/ajph.88.9.1325
Robertson, M. J., Koegel, P., & Ferguson, L. (1989). Alcohol use and abuse among
homeless adolescents in Hollywood. Contemporary Drug Problems, 16, 415
Rogers, E. M. (1995). Diffusion of Innovations (4th ed.). New York: Free Press
Salomonsen-Sautel, S., Van Leeuwen, J. M., Gilroy, C., Boyle, S., Malberg,
D., & Hopfer, C. (2008). Correlates of substance use among homeless
youths in eight cities. The American Journal on Addictions, 17(3), 224–234.
doi:10.1080/10550490802019964
Savin-Williams, R. C. (1994). Verbal and physical abuse as stressors in the lives
of lesbian, gay male, and bisexual youths: Associations with school problems,
running away, substance abuse, prostitution, and suicide. Journal of Consulting
and Clinical Psychology, 62(2), 261
Schneider, J. A., Zhou, A. N., & Laumann, E. O. (2015). A new HIV prevention
network approach: Sociometric peer change agent selection. Social Science &
Medicine, 125, 192–202. doi:10.1016/j.socscimed.2013.12.034
Shillington, A. M., Bousman, C. A., & Clapp, J. D. (2011). Characteristics of homeless
youth attending two different youth drop-in centers. Youth & Society, 43(1), 28–43.
doi:10.1177/0044118X09351277
Slesnick, N., Dashora, P., Letcher, A., Erdem, G., & Serovich, J. (2009).
A review of services and interventions for runaway and homeless youth:
Moving forward Children and Youth Services Review, 31(7), 732–742.
doi:10.1016/j.childyouth.2009.01.006
Slesnick, N., Feng, X., Guo, X., Brakenhoff, B., Carmona, J., Murnan, A., . . . McRee,
A.-L. (2016). A test of outreach and drop-in linkage versus shelter linkage
for connecting homeless youth to services. Prevention Science, 17(4), 450–460.
doi:10.1007/s11121-015-0630-3

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 36 The Causes and Consequences of Youth Homelessness

Tenner, A. D., Trevithick, L. A., Wagner, V., & Burch, R. (1998). Seattle YouthCare’s
prevention, intervention, and education program. Journal of Adolescent Health,
23(2), 96–106. doi:10.1016/S1054-139X(98)00057-3
The NIMH Collaborative HIV/STD Prevention Trial Group. (2010). Results of the
NIMH Collaborative HIV/STD Prevention Trial of a Community Popular Opinion
Leader Intervention. Journal of Acquired Immune Deficiency Syndromes (1999),
54(2), 204–214. doi:10.1097/QAI.0b013e3181d61def
Thompson, S. J., Bender, K., Ferguson, K. M., & Kim, Y. (2015). Factors asso-
ciated with substance use disorders among traumatized homeless youth. Jour-
nal of Social Work Practice in the Addictions, 15(1), 66–89. doi:10.1080/
1533256X.2014.996229
Thompson, S. J., Bender, K. A., Lewis, C. M., & Watkins, R. (2008). Runaway
and pregnant: risk factors associated with pregnancy in a national sample of
runaway/homeless female adolescents. Journal of Adolescent Health, 43(2), 125–
132. doi:10.1016/j.jadohealth.2007.12.015
Tyler, K. A., & Melander, L. A. (2010). Foster care placement, poor parenting, and
negative outcomes among homeless young adults. Journal of Child and Family
Studies, 19(6), 787–794. doi:10.1007/s10826-010-9370-y
Um, M.-Y., Rice, E., Rhoades, H., Winetrobe, H., & Bracken, N. (2016). Influence of
“traveling” youth and substance use behaviors in the social networks of homeless
youth. Network Science, 4(1), 28–47. doi:10.1017/nws.2015.32
Valente, T. W., & Pumpuang, P. (2007). Identifying opinion leaders to pro-
mote behavior change. Health Education & Behavior, 34(6), 881–896.
doi:10.1177/1090198106297855
Walls, N. E., & Bell, S. (2011). Correlates of engaging in survival sex among
homeless youth and young adults. The Journal of Sex Research, 48(5), 423–436.
doi:10.1080/00224499.2010.501916
Wen, C. K., Hudak, P. L., & Hwang, S. W. (2007). Homeless people’s perceptions of
welcomeness and unwelcomeness in healthcare encounters. Journal of General
Internal Medicine, 22(7), 1011–1017. doi:10.1007/s11606-007-0183-7
Whitbeck, L. B., & Hoyt, D. R. (1999). Nowhere to grow: Homeless and runaway
adolescents and their families. New York: Aldine De Gruyter
Whitbeck, L. B., Hoyt, D. R., & Yoder, K. A. (1999). A risk-amplification
model of victimization and depressive symptoms among runaway and home-
less adolescents. American Journal of Community Psychology, 27(2), 273–296.
doi:10.1023/A:1022891802943
Whitbeck, L. B., Johnson, K. D., Hoyt, D. R., & Cauce, A. M. (2004). Mental disorder
and comorbidity among runaway and homeless adolescents. Journal of Adolescent
Health, 35(2), 132–140. doi:10.1016/j.jadohealth.2003.08.011
Winetrobe, H., Rhoades, H., Barman-Adhikari, A., Cederbaum, J., Rice, E., & Milburn,
N. (2013). Pregnancy attitudes, contraceptive service utilization, and other factors
associated with Los Angeles Homeless Youths’ use of effective contraception and
withdrawal. Journal of Pediatric and Adolescent Gynecology, 26(6), 314–322.
doi:10.1016/j.jpag.2013.06.007
Winetrobe, H., Rhoades, H., Rice, E., Milburn, N., & Petering, R. (2017). “I’m not
homeless, I’m houseless”: Identifying as homeless and associations with service
utilization among Los Angeles homeless young people. Journal of Social Distress
and the Homeless, 26(1), 16–24. doi:10.1080/10530789.2017.1280204

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 Eric Rice and Hailey Winetrobe 37

Winetrobe, H., Rice, E., Rhoades, H., & Milburn, N. (2016). Health insurance coverage
and healthcare utilization among homeless young adults in Venice, CA. Journal of
Public Health, 38(1), 147–155. doi:10.1093/pubmed/fdv001
Yadav, A., Marcolino, L. S., Rice, E., Petering, R., Winetrobe, H., Rhoades, H., . . .
Carmichael, H. (2015). Preventing HIV Spread in Homeless Populations Using
PSINET. Paper presented at the AAAI, Austin, Texas, January 25–30
Yoshioka-Maxwell, A., & Rice, E. (2017). Exploring the impact of network characteris-
tics on substance use outcomes among homeless former foster youth. International
Journal of Public Health, 62(3), 371–378. doi:10.1007/s00038-016-0845-5

https://doi.org/10.1017/9781108669016.003 Published online by Cambridge University Press

 3
Using Social Networks to Raise HIV Awareness
among Homeless Youth
Amulya Yadav, Bryan Wilder, Hau Chan, Albert Jiang, Haifeng Xu,
Eric Rice, and Milind Tambe

Introduction
Homelessness has reached a crisis level, with more than 565,000 homeless
people in the United States on any given night. Homeless youth (i.e., people
below the age of 25) account for almost 34% of the total homeless population
[1]. These homeless youth face significant difficulties, having to struggle for
basic amenities such as health care, nutritious food, and primary education.
HIV has an extremely high incidence among homeless youth, as they are
more likely to engage in high HIV-risk behaviors (e.g., unprotected sexual
activity, injection drug use) than other sub-populations. In fact, previous
studies show that homeless youth face a 10-times greater risk of HIV infection
than do stably housed populations [1].
To help prevent HIV infection among homeless youth, many homeless
drop-in centers implement social network–based peer leader intervention
programs, where a select number of youth, called peer leaders, are taught
strategies for reducing risk of contracting HIV. These intervention programs
consist of day-long educational sessions in which these peer leaders are
provided with information about HIV prevention measures [2]. These leaders
are then encouraged to share these messages with their peers in their social
circles, in order to lead all their peers toward safer behaviors and practices.
These peer leader–based intervention programs are motivated by the drop-
in centers’ limited financial resources, which prevents them from directly
assisting the entire homeless youth population. This leads to the question
well known in the field of influence maximization, namely how to select
“influential” nodes (i.e., homeless youth) to maximize the spread of awareness
within a given social network. In the context of homeless drop-in centers,
this problem is further complicated by two factors. First, the social network
structure is imperfectly known, which makes identifying “influential” nodes

38

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 Amulya Yadav et al. 39

challenging [3]. Although some connections (friendships) are known, other

connections may be uncertain. This is because homeless youth are a hard-to-
reach population, and their social networks are harder to characterize than are
networks of stably housed youth [2]. Second, managing homeless youth (some
of whom have emotional and behavioral problems) during an intervention, with
the homeless drop-in center’s limited personnel, is not easy. As a result, the
drop-in center officials can only manage small groups composed of three or
four youths at one time. Therefore, the drop-in center officials prefer a series
of small-sized intervention camps organized sequentially (i.e., one after the
other) to maximize the impact of their intervention [4]. In such camps, youth
may reveal some additional information about the network, which can be used
to inform future interventions.
The drop-in centers need a plan to choose the participants (i.e., peer
leaders) of their sequentially organized interventions. This plan must address
four key points: (1) it must efficiently deal with uncertainties in the network
structure, i.e., uncertainty about existence or absence of some friendships in the
network; (2) it needs to take into account new information uncovered during
the interventions, which reduces the uncertainty in our understanding of the
network; (3) the plan needs to be deviation tolerant, as sometimes homeless
youth may choose not to be a peer leader, thereby forcing the drop-in center to
modify its plan; (4) our approach should address the challenge of gathering
information about social networks of homeless youth, which usually costs
thousands of dollars and many months of time [4].
This chapter presents three key contributions in addressing the sequential
planning needs of homeless drop-in centers. First, we model the drop-in
centers’ sequential planning needs by introducing the dynamic influence
maximization under uncertainty (or DIME) problem. The sequential selection
of intervention participants under network uncertainty in DIME sets it apart
from any other previous work on influence maximization, which mostly
focuses on single-shot decision problems (i.e., a set of nodes are selected
just once, instead of selecting sets of nodes repeatedly) [5–8]. We analyze
several novel theoretical aspects of the DIME problem, which illustrates its
computational hardness.
Second, we propose a new software agent, HEALER (hierarchical ensem-
bling based agent, which plans for effective reduction in HIV spread), to
provide an end-to-end solution to the DIME problem. First, HEALER casts the
DIME problem as a partially observable Markov decision process (POMDP)
and solves it using HEAL (hierarchical ensembling algorithm for planning), a
novel POMDP planner that quickly generates high-quality recommendations
(of intervention participants) for homeless drop-in center officials. In this

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 40 HIV Awareness in Homeless Youth

chapter, we discuss the design of HEALER and explain its method of

gathering information about the homeless youth social network (at low cost)
by interacting with youth via a network construction application. We also give
a high-level overview of the HEAL algorithm and refer the reader to Yadav
et al. [9] for a more complete understanding.
Third, we introduce the DOSIM algorithm, a novel algorithm for solving
the DIME problem when the parameters of the problem are unknown. DOSIM
solves a zero-sum game in which nature adversarially chooses the true
parameters (e.g., the probability that influence will spread along a given edge).
Lastly, we present experimental simulation results on social networks
gathered from homeless youth. We find that both HEALER and DOSIM
substantially outperform the heuristic that is currently used in the field.

Related Work
There are three distinct areas of work related to the homeless drop-in center
problem that we introduced in the previous section. The primary problem in
computational influence maximization is to find optimal “seed sets” of nodes
in social networks, which can maximize the spread of information or influence
in the social network (according to some a priori known influence model).
While there are many algorithms for finding such seed sets [5–8], most of
these algorithms assume no uncertainty in the network structure and select a
single seed set. In contrast, HEALER selects several seed sets sequentially
in our work to select intervention participants for each successive training
program, taking into account updates to the network structure revealed in
past interventions. The DIME problem also incorporates uncertainty about
the network structure and influence status of network nodes (i.e., whether a
node is influenced or not). Finally, unlike [5–8], HEALER uses a different
diffusion model as we explain later here. Golovin et al. [9] introduced adaptive
submodularity and discussed adaptive sequential selection (similar to our
problem), and they proved that a Greedy algorithm has a (1-1/e) approximation
guarantee. However, unlike the DIME problem, they assume no uncertainty in
the network structure. We show that while the DIME problem can be cast into
the adaptive stochastic optimization framework of [10], its influence function
is not adaptive submodular (see the “DIME Problem Statement” section) and
because of this their Greedy algorithm loses its approximation guarantees.
Finally, Lei et al. [11] use multi–armed bandit algorithms to pick influential
nodes in social networks when influence probabilities are not known, but their

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 Amulya Yadav et al. 41

approach requires many iterations to converge, thereby making it unsuitable

for a real-world domain like ours.
Next, we discuss literature from social work. The general approach to
these interventions is to use peer change agents (PCA) (i.e., peers who bring
about change in attitudes) to engage homeless youth in interventions, but
most studies do not use network characteristics to choose these PCAs [12].
A notable exception is Valente et al. [13], who proposed selecting intervention
participants with highest degree centrality (the most ties to other homeless
youth). However, previous studies [14, 15] show that degree centrality per-
forms poorly, as it does not account for potential overlaps in influence of two
high-degree centrality nodes.
The final field of related work is planning for reward and cost optimization.
We only focus on the literature on Monte-Carlo (MC) sampling based online
POMDP solvers, since this approach allows significant scale-up [16]. The
POMCP (Partially Observable Monte-Carlo Planning) solver [17] uses Monte-
Carlo UCT (upper confidence bound) tree search in online POMDP planning.
Also, Somani et al. [18] present the DESPOT (determinized sparse partially
observable tree) algorithm, which improves the worst-case performance of
POMCP. Our initial experiments with POMCP and DESPOT showed that
they run out of memory on even our small-sized networks. A recent paper
introduced PSINET-W [14], which is a MC sampling–based online POMDP
planner. We have discussed PSINET’s shortcomings earlier, and how HEALER
remedies them with the use of its heuristics. In particular, HEALER scales up
whereas PSINET fails to do so. HEALER’s algorithmic approach also offers
significant novelties in comparison with PSINET.

HEALER’s Design
HEALER has a modular design [9], and consists of two major components.
First, it has a network construction application for gathering information
about social networks. Second, it has an algorithm called HEAL, which
solves the DIME problem (introduced later) using heuristics. We first explain
HEALER’s components individually, and then explain how they are used
inside HEALER’s design.

Network Construction Application

HEALER gathers information about social ties among homeless youth by
interacting with these youth via its network construction application. Once

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 42 HIV Awareness in Homeless Youth

a fixed number of homeless youth register in its network application (which

is hosted as a website to ensure ease of access for the youth), HEALER
parses contact lists (on Facebook) of all the registered homeless youth and
generates the social network that connects these youth. We choose Facebook
for gathering information because previous studies [19] show that a large
proportion (∼80%) of homeless youth are regularly active on Facebook.
Specifically, HEALER adds a link between two homeless youth if and only if
both youth are (1) friends on Facebook and (2) are registered in its application.
Unfortunately, there is uncertainty in the generated network, as friendship links
between people who are only friends in real life (and have not added each other
as friends on Facebook) are not captured by HEALER’s network construction
application.
Previously, collecting accurate social network data on homeless youth was
a technical and financial burden beyond the capacity of most agencies working
with these youth [19]. Research staff at homeless drop-in centers conducted
tedious face-to-face interviews with homeless youth to infer ties between
these youth, a process that costs thousands of dollars and many months of
time. HEALER’s network construction application enables homeless drop-
in centers to quickly generate a first approximation of the homeless youth
social network at low cost. The HEAL algorithm (the second component
in HEALER) subsequently corrects and improves the social network struc-
ture iteratively (as explained later), which is one of the major strengths
of this approach. This network construction application has been tested
multiple times by our collaborating homeless drop-in centers with positive
feedback.

DIME Solver
The DIME Solver then takes the approximate social network (generated by
HEALER’s network construction application) as input and solves the DIME
problem (formally defined later in the paper) using HEAL, the core algorithm
running inside HEALER. The HEAL algorithm is an online POMDP solver,
i.e., it interleaves planning and execution for each time step (explained later in
the paper). The solution of the DIME problem generated by HEAL is provided
as a series of recommendations (of intervention participants) to homeless
drop-in center officials. Each recommendation would urge the officials to
invite a particular set of youth for their intervention camp. For example,
in Figure 3.1, HEALER would recommend inviting nodes D and A for the
intervention.

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 Amulya Yadav et al. 43

HEALER Design
HEALER’s design begins with the network construction application construct-
ing an uncertain network (as explained above). HEALER has a sense-reason-
act cycle, where it repeats the following process for T interventions. It reasons
about different long-term plans to solve the DIME problem, and it acts by
providing DIME’s solution as a recommendation (of intervention participants)
to homeless drop-in center officials. The officials may choose to not use
HEALER’s recommendation in selecting their intervention’s participants.
After finalizing the selection of participants, the drop-in center officials contact
the chosen participants (via phone/email) and conduct the intervention with
them. Upon the intervention’s completion, HEALER senses feedback about
the conducted intervention from the officials. This feedback includes new
observations about the network, e.g., uncertainties in some links may be
resolved as intervention participants are interviewed by the drop-in center
officials (explained more later). HEALER uses this feedback to update and
improve its future recommendations.

DIME Problem Statement

HEALER represents social networks as directed graphs (consisting of nodes
and directed edges) where each node represents a person in the social network
and a directed edge between two nodes A and B (say) represents that node A
considers node B as their friend. HEALER assumes directed-ness of edges, as
sometimes homeless drop-in centers assess that the influence in a friendship is
very much unidirectional, and so it has to account for uni-directional follower
links. Otherwise friendships are encoded as two unidirectional links. In the
following, we provide some background information that helps define a precise
problem statement for DIME. After that, we will show some hardness results
about this problem statement.

Uncertain Network
The uncertain network is a directed graph G = (V, E) with |V|= N nodes and
|E|= M edges. The edges E in an uncertain network are of two distinct types:
(1) the set of certain edges (Ec ) that consists of friendships we are certain
about; and (2) the set of uncertain edges (Eu ) that consists of friendships
we are uncertain about. Recall that uncertainties about friendships exist
because HEALER’s network construction application misses out on some links
between people who are friends in real life but not on Facebook.

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 44 HIV Awareness in Homeless Youth

To model the uncertainty about missing edges, every uncertain edge has an
existence probability u(e) associated with it, which represents the likelihood
of “existence” of that uncertain edge in the real world. For example, if there
is an uncertain edge (A, B) (i.e., we are unsure whether node B is node A’s
friend), then u(A, B) = 0.75 implies that B is A’s friend with a 0.75 chance. This
existence probability allows us to measure the potential value of influencing a
given node. For example, if node A is connected to many uncertain edges with
low u(e) values, then it is unlikely that node A is highly influential (as most of
his supposed friendships may not exist in reality).
In addition, every edge in the network (both certain and uncertain) has
a propagation probability p(e) associated with it. A propagation probability
of 0.5 on directed edge (A, B) denotes that if node A is influenced (i.e.,
has information about HIV prevention), it influences node B (i.e., gives
information to node B) with a 0.5 probability in each subsequent time step (our
full influence model is defined later here). This graph G with all relevant p(e)
and u(e) values represents an uncertain network and serves as an input to the
DIME problem. Figure 3.1 shows an example of an uncertain network, where
the dotted edges represent uncertain edges. We now explain how HEALER
generates an uncertain social network.
First, HEALER uses its network construction application to generate a
network with no uncertain edges. Next, we use well-known link prediction
techniques such as KronEM [20] to infer existence probabilities u(e) for
additional friendships that might have been missed by the network construction
application. This process gives us an uncertain network, which is then used by
HEALER to generate recommendations, as we explain next.

Figure 3.1 Illustration of the value of information in the DIME Problem. A, B, C,

and D represent nodes, and the edges between them represent friendships. There
are two kinds of edges: certain edges (denoted by solid edges as shown in the
left figure) and uncertain edges (denoted by dotted edges as shown in the right
figure). The propagation and existence probabilities on all the edges is assumed
to be fixed.

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 Amulya Yadav et al. 45

Given the uncertain network as input, HEALER runs for T rounds (cor-
responding to the number of interventions organized by the homeless drop-
in center). In each round, HEALER chooses K nodes (youth) as intervention
participants. These participants are assumed to be influenced after the inter-
vention (i.e., our intervention deterministically influences the participants).
Upon influencing the chosen nodes, HEALER “observes” the true state of the
uncertain edges (friendships) outgoing from the selected nodes. This translates
into asking intervention participants about their one-hop social circles, which
is within the capabilities of the homeless drop-in center [2].
After each round, influence spreads in the network according to our
influence model (explained later here) for L time steps, before we begin the
next round. This L is the time duration in between two successive intervention
camps. In between rounds, HEALER does not observe the nodes that get
influenced during L time steps. Thus, while HEALER knows the influence
model, it does not observe the random samples from the influence model that
led to some nodes getting influenced. HEALER only knows that explicitly
chosen nodes (our intervention participants in all past rounds) are influenced.
Informally then, given an uncertain network G0 = (V, E) and integers T, K, and
L (as defined earlier), HEALER finds an online policy for choosing exactly K
nodes for T successive rounds (interventions) that maximizes influence spread
in the network at the end of T rounds.

Influence Model
Unlike most previous work in influence maximization [5–8], HEALER uses a
variation of the independent cascade model [21]. In the standard independent
cascade model, all nodes that get influenced at time t get a single chance to
influence their un-influenced neighbors at time t + 1. If they fail to spread
influence in this single chance, they do not spread influence to their neighbors
in future rounds. On the other hand, HEALER’s model assumes that nodes get
multiple chances to influence their un-influenced neighbors. If they succeed
in influencing a neighbor at a given time step t , they stop influencing that
neighbor for all future time steps. Otherwise, if they fail in step t , they try to
influence again with the same propagation probability in the next time step.
This variant of independent cascade has been shown to empirically provide a
better approximation of the real influence spread than the standard independent
cascade model [21, 22]. Further, we assume that nodes that get influenced at a
certain time step remain influenced for all future time steps.
We now provide notation for defining HEALER’s policy formally. Let
Δ = {A ⊂ V s.t. |A|= K} denote the set of K sized subsets of V, which represents

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 46 HIV Awareness in Homeless Youth

the set of possible choices that HEALER can make at every time step t ∈ [1, T ].
Let Ai ∈ Δ ∀i ∈ [1, T ] denote HEALER’s choice in the ith time step. Upon
making choice Ai , HEALER “observes” uncertain edges adjacent to nodes in
Ai , which updates its understanding of the network. Let Gi ∀ i ∈ [1, T ] denote
the uncertain network resulting from Gi−1 with observed (additional edge)
information from Ai . Formally, we define a history Hi ∀ i ∈ [1, T ] of length i as
a tuple of past choices and observations Hi =< G0 , A1 , G1 , A2 , . . . , Ai−1 , Gi >.
Denote by Yi = {Hk s.t. k ≤ i} the set of all possible histories of length less
than or equal to i. Finally, we define an i-step policy Πi : Yi → Δ as a
function that takes in histories of length less than or equal to i and outputs
a K node choice for the current time step. We now provide an explicit problem
statement for DIME.

Problem Statement
Given as input are an uncertain network G0 = (V, E) and integers T, K, and
L (as defined earlier). Denote by R(HT , AT ) the expected total number of
influenced nodes at the end of round T, given the T -length history of previous
observations and actions HT , along with AT , the action chosen at time T. Let
E{HT ,AT ∼ΠT } R(HT , AT ) denote the expectation over the random variables HT =
< G0 , A1 , G1 , A2 , . . . , AT −1 , GT > and influence of AT , where Ai are chosen
according to ΠT (Hi ) ∀ i ∈ [1, T ], and Gi are drawn according to the distribution
over uncertain edges of Gi−1 that are revealed by Ai . The objective of DIME is
to find an optimal T-step policy ΠT∗ = argmaxΠT E{HT ,AT ∼ΠT } R(HT , AT ).
Next, we show hardness results about the DIME problem. First, we analyze
the value of having complete information in DIME. Then, we characterize the
computational hardness of DIME.

The Value of Information

We characterize the impact of insufficient information (about the uncertain
edges) on the achieved solution value. We show that no algorithm for DIME
is able to provide a sufficiently good approximation to the full-information
solution value (i.e., the best solution achieved with regard to the underlying
ground-truth network), even with infinite computational power.
Theorem 1 Given an uncertain network with n nodes, for any  > 0, there is
no algorithm for the DIME problem that can guarantee a n−1+ approximation
to OPT full , the full-information solution value.
Proof: We prove this statement by providing a counterexample in the form
of a specific (ground-truth) network for which there can exist no algorithm

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 Amulya Yadav et al. 47

that can guarantee a n−1+ approximation to OPT full . Consider an input to the
DIME problem, an uncertain network with n nodes with 2 × nC2 uncertain
edges between the n nodes, i.e., it is a completely connected uncertain network
consisting of only uncertain edges (an example with n = 3 is shown in Figure
3.1). Let p(e) = 1 and u(e) = 0.5 on all edges in the uncertain network, i.e., all
edges have the same propagation and existence probability. Let K = 1, L = 1,
and T = 1, i.e., we just select a single node in one shot (in a single round). 
Further, consider a star graph (as the ground truth network) with n nodes
such that propagation probability p(e) =1 on all edges of the star graph (shown
in Figure 3.1). Now, any algorithm for the DIME problem would select a single
node in the uncertain network uniformly at random with equal probability
of 1/n (as information about all nodes is symmetrical).
   In expectation, the
algorithm will achieve an expected reward n × n + 1n × 1 + . . . + 1n × 1 =
1

n × n + n × 1 = 2 − n . However, given the ground truth network, we get

1 n−1 1

OPT full = n because we always select the star node. As n goes to infinity, we
can at best achieve a n−1 approximation to OPT full . Thus, no algorithm can
achieve a n−1+ approximation to OPT full for any  > 0.

Computational Hardness
We now analyze the hardness of computation in the DIME problem in the next
two theorems.
Theorem 2 The DIME problem is NP-Hard.
Proof: Consider the case where Eu = Φ, L = 1, T = 1, and p(e) = 1 ∀e ∈ E.
This degenerates to the classical influence maximization problem, which is
known to be NP-hard. Thus, the DIME problem is also NP-hard. 
Some NP-Hard problems exhibit nice properties that enable approximation
guarantees for them. Golovin et al. [23] introduced adaptive submodularity, an
analog of submodularity for adaptive settings. Intuitively, adaptive submod-
ularity deals with cases in which actions/items are to be picked in multiple
stages, and newer information is revealed every time an action is picked.
Adaptive submodularity requires that the expected marginal gain of picking
an action can only decrease as more actions are picked and more information
is revealed. Formally, adaptive submodularity requires that Marginal(A, Ψ) ≤
Marginal(A, Ψ ) ∀ Ψ ⊂ Ψ , where Marginal(A, Ψ) represents the marginal
gain/benefit of picking action A, conditioned on getting information Ψ . This
makes the adaptive submodularity framework a natural fit for the DIME
problem. Presence of adaptive submodularity ensures that a simply greedy

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 48 HIV Awareness in Homeless Youth

algorithm provides a (1-1/e) approximation guarantee with regard to the

optimal solution defined on the uncertain network. However, as we show next,
while DIME can be cast into the adaptive stochastic optimization framework
of [23], our influence function is not adaptive submodular, because of which
their Greedy algorithm does not have a (1-1/e) approximation guarantee.
Theorem 3 The influence function of DIME is not adaptive submodular.
Proof: The definition of adaptive submodularity requires that the expected
marginal increase of influence by picking an additional node is more when
we have less observation. Here the expectation is taken over the random states
that are consistent with current observation. We show that this is not the case in
DIME problem. Consider a path with three nodes, A, B, and C and two directed
edges, e1 = (A, B) and e2 = (B, C). Let p(e1 ) = p(e2 ) = 1, i.e., propagation
probability is 1; and u(e1 ) = 1, u(e2 ) = 1 −  for some small enough  to be set.
Thus, the only uncertainty comes from incomplete knowledge of the existence
of edges. 
Let us assume that we pick node A. After picking node A, the expected
marginal benefit of picking node C is . However, after picking node B, if we
get information Ψ = {(B, C)does not exist}, then the expected marginal benefit
of picking node C goes to 1 (up from ). Since the expected marginal benefit
of picking node C increased from  to 1 upon receiving more information and
picking more actions, this contradicts the definition of adaptive submodularity.
This shows that the influence function of DIME is not adaptive submodular.

HEAL: DIME Problem Solver

The above theorems show that DIME is a hard problem, as it is difficult
to even obtain any reasonable approximations. HEALER models DIME as
a partially observable Markov decision process (POMDP) [24], which is a
logical fit for the problem, for two reasons. First, several interventions are
conducted sequentially, similar to sequential POMDP actions. Second, there
is partial observability (similar to POMDPs) due to uncertainties in network
structure and influence status of nodes. We now provide a high-level overview
of HEALER’s POMDP model.

POMDP Model
A state in this model includes the influence status of all network nodes (i.e.,
which nodes are influenced and which nodes are not) and the true state of
the uncertain edges (i.e., whether each uncertain edge exists or not in the

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 Amulya Yadav et al. 49

real world). Thus, there are 2N+M possible POMDP states in a network with
N nodes and M uncertain edges. Similarly, an action in this model is any
possible subset of K network nodes, which can be called for an intervention.
Thus, if K nodes are being selected in every intervention on a network with N
nodes, there are N C K possible POMDP actions. Finally, an observation in this
model is based on the assumption that when a set of K nodes (i.e., K distinct
homeless youth) are called in for intervention, the drop-in center officials can
talk to these nodes (or youth) and resolve the status of the uncertain edges in
their local neighborhood. Specifically, the drop-in center official observes the
true state of each uncertain edge (i.e., whether it exists in the real world or
not) outgoing from the K nodes chosen in that action. The observation of the
true state of uncertain edge (A, B) leads to resetting of u(A, B) to either 1 or
0 (depending on whether edge (A, B) actually exists or not). Thus, when M
uncertain edges are outgoing from the K nodes chosen in a POMDP action,
there are 2 M possible POMDP observations. Finally, the rewards in this model
keep track of the number of new nodes that get influenced upon taking a
POMDP action. Refer to Yadav et al. [9] for the full POMDP model.

HEAL
HEAL is a heuristic-based online POMDP planner for solving the DIME prob-
lem. HEAL solves the original POMDP using a novel hierarchical ensembling
heuristic: it creates ensembles of imperfect (and smaller) POMDPs at two
different layers, in a hierarchical manner (see Figure 3.2). HEAL’s top layer
creates an ensemble of smaller sized intermediate POMDPs by subdividing the
original uncertain network into several smaller sized partitioned networks by
using graph partitioning techniques [25]. Each of these partitioned networks is
then mapped onto a POMDP, and these intermediate POMDPs form the top
layer ensemble of POMDP solvers.
In the bottom layer, each intermediate POMDP is solved using TASP
(tree aggregation for sequential planning), HEAL’s POMDP planner, which
subdivides the POMDP into another ensemble of smaller-sized sampled
POMDPs. Each member of this bottom layer ensemble is created by randomly
sampling uncertain edges of the partitioned network to get a sampled network
having no uncertain edges, and this sampled network is then mapped onto a
sampled POMDP. Finally, the solutions of POMDPs in both the bottom and
top layer ensembles are aggregated using novel techniques to get the solution
for HEAL’s original POMDP.
These heuristics enable scale up to real-world sizes (at the expense of
sacrificing performance guarantees), as instead of solving one huge problem,

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 50 HIV Awareness in Homeless Youth

Figure 3.2 Hierarchical decomposition in HEAL (from Yadav et. al. [9]).

HEAL solves several smaller problems. The primary difference between

HEAL and PSINET (the previous state-of-the-art) is in the top layer of
HEAL, which uses the graph partitioning heuristic. This heuristic divides
up the network into different partitions, with each partition corresponding
to an intermediate POMDP (Figure 3.2). The partitions are chosen in a way
that minimizes the number of cross-edges going across the partitions while
ensuring that the partitions have similar sizes. Since these partitions are
almost disconnected, we solve each partition separately without accounting for
influence going across the partitions. Simulations show that even on smaller
settings, HEAL achieves a 100-fold speed-up over PSINET, while providing a
70% improvement in solution quality; and on larger problems, where PSINET
is unable to run at all, HEAL continues to provide high solution quality.

DOSIM: A New Algorithm for the DIME Problem

DOSIM [26] is a novel algorithm that solves a generalization of the DIME
problem. The key motivation behind DOSIM is to be able to select actions

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 Amulya Yadav et al. 51

(i.e., set of K nodes) for T stages without knowing the exact model parameters
(i.e., p(e) and u(e) values for each network edge). HEALER dealt with this
issue by assuming a specific p(e) and u(e) value based on suggestions by
service providers. DOSIM instead works with interval uncertainty over these
model parameters. That is, the exact value of each u(e) and p(e) does not have
be exactly supplied; they are just assumed to lie within some interval. This
generalizes the model used by HEALER to include higher-order uncertainty
over the probabilities in addition to the uncertainty induced by the probabilities
themselves. DOSIM chooses an action that is robust to this interval uncertainty.
Specifically, it finds a policy that achieves close to optimal value regardless of
where the unknown probabilities lie within the interval.
The problem is formalized as a zero-sum game between the algorithm,
which picks a policy, and an adversary (nature) who chooses the model
parameters. A game models a situation where two players must each select
an action. Each player receives a payoff depending on the actions that were
taken. “Zero sum” indicates that one player’s loss is the other’s gain. In our
situation, we model nature as selecting the model parameters that result in
the worst performance for the algorithm. This game formulation represents a
key advance over HEALER’s POMDP policy (which was constrained to fixed
propagation probabilities), as it enables DOSIM to output mixed strategies
(distributions) over POMDP policies, which make it robust against worst-
case propagation probability values. Moreover, DOSIM receives periodic
observations, which are used to update its understanding of its belief state
(i.e., probability distribution over different model parameters). However,
the number of possible actions for the game is intractably large, because there
are an exponential number of policies (each of which specifies an action to take
for any possible set of observations).
To resolve this issue, DOSIM uses a double oracle approach, illustrated in
Figure 3.3. Each of the tables shown is a payoff table, which lists the actions
available for each player. Each cell of the table gives the payoff that each
player obtains when the corresponding strategies are chosen. Since the game
is zero sum, these payoffs always total zero. In a double oracle algorithm,
each player starts with only a limited number of arbitrarily chosen strategies
available to them. In our case, this means that the algorithm has perhaps
one or two policies available, and nature has one or two sets of parameters.
Standard techniques from game theory can be used to find the best strategy
for each player in such a small game (the best strategies are referred to as an
equilibrium). Then, DOSIM incrementally adds new strategies to these small
starting sets. Specifically, DOSIM has an oracle for each player. The oracle
computes that player’s best strategy to play in light of the options that the

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 52 HIV Awareness in Homeless Youth

Figure 3.3 Example of the double oracle algorithm.

other player currently has available. This new strategy is then added to the
list of those currently under consideration. DOSIM alternates between adding
strategies for the two players until neither can benefit from adding any more
strategies. At this point, it has reached an equilibrium.

Experimental Results
We compare three algorithms: HEALER, DOSIM, and degree centrality (DC).
DC is the current standard practice for influence maximization in this domain,
where drop-in center officials recruit the set of youth with the highest degree to
serve as peer leaders. In order to compare HEALER and DOSIM, we assume
that the problem parameters are known, fixing u(e) = 0.1 and p(e) = 0.6 for
all edges in the network. This was done based on findings in Kelly et. al. [27].
This experiment was run on a 2.33 GHz 12-core Intel machine having 48 GB of
RAM, and was averaged over 100 runs. All comparison results are statistically
significant under bootstrap-t (α = 0.05).
We run the algorithms on two networks, Network A and Network B. Each
represents the social network of homeless youth gathered from a different
drop-in center in the Los Angeles area. The networks had between 140 and
170 nodes each. Figure 3.4 shows the results for Network A and Figure 3.5
shows the results for Network B. Each figure plots the average influence spread
obtained by each algorithm over 100 runs. On the x-axis, we vary the number
of seeds K. The number of time steps T is fixed to 5 throughout. We see

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 Amulya Yadav et al. 53

Figure 3.4 Simulated influence spread of each algorithm on Network A.

Figure 3.5 Simulated influence spread of each algorithm on Network B.

that both HEALER and DOSIM significantly outperform DC. This holds on
both networks and across all values of K. HEALER and DOSIM benefit from
considering the global structure of the network. For example, they will (either
explicitly or implicitly) avoid placing many seed nodes close together. DC
obtains significantly less influence spread because it looks only at the degree
of each node in isolation.

Conclusion
In this chapter, we explored design of HEALER and DOSIM, two adaptive
software agents that recommend intervention attendees to homeless drop-in

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 54 HIV Awareness in Homeless Youth

center officials. HEALER solves a POMDP on a social network to come

up with recommendations for which homeless youth in a social network
should be chosen as intervention attendees. We first formally characterized
the computational problem (called DIME) solved by HEALER, and showed
that it is an NP-Hard problem. Moreover, well-known algorithms such as
Greedy lose their approximation guarantees in the DIME problem due to the
feedback about network structure received during interventions. We inferred
that DIME’s computational hardness forces HEALER to rely on heuristic
methods for solving DIME. Further, we analyzed these heuristic methods
and showed that the primary reason behind the superior performance of
HEALER is its graph partitioning heuristic, which works well due to the
small-world nature of the real-world networks of homeless youth. We then
examined a new algorithm for the problem, DOSIM. DOSIM is motivated
by uncertainty about the underlying parameters of the problem. Rather than
assuming a fixed value for the parameters, it solves a zero-sum game to
determine a policy that is robust to a worst-case set of parameters. We then
compared the performance in simulation of both algorithms and of a commonly
used heuristic (degree centrality). Both DOSIM and HEALER substantially
outperformed the baseline.

References
[1] National HCH Council, “HIV/AIDS among Persons Experiencing Homelessness:
Risk Factors, Predictors of Testing, and Promising Testing Strategies.” [Online].
Available: www.nhchc.org/wp-content/uploads/2011/09/InFocus_Dec2012.pdf.
[2] E. Rice, E. Tulbert, J. Cederbaum, A. B. Adhikari, N. G. Milburn, “Mobilizing
Homeless Youth for HIV Prevention: A Social Network Analysis of the Accept-
ability of a Face-to-Face and Online Social Networking Intervention,” Health
Education Research, vol. 27, no. 2, pp. 226–236, 2012.
[3] E. Rice, “The Positive Role of Social Networks and Social Networking Technol-
ogy in the Condom-using Behaviors of Homeless Young People,” Public Health
Reports, vol. 125, no. 4, pp. 588–595, 2010.
[4] E. Rice, A. Fulginti, H. Winetrobe, J. Montoya, A. Plant, T. Kordic, “Sexuality
and Homelessness in Los Angeles Public Schools,” American Journal of Public
Health, vol. 102, pp. 200–201, 2012.
[5] C. Borgs, M. Brautbar, J. Chayes, B. Lucier, “Maximizing Social Influence in
Nearly Optimal Time,” Proc. 25th Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), pp. 946–957, 2014.
[6] Y. Tang, X. Xiao, Y. Shi, “Influence Maximization: Near-Optimal Time Com-
plexity Meets Practical Efficiency,” Proc. 2014 ACM SIGMOD International
Conference on Management of Data, pp. 75–86, 2014.

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 Amulya Yadav et al. 55

[7] D. Kempe, J. Kleinberg, E. Tardos, “Maximizing the Spread of Influence through

a Social Network,” Proc. 9th ACM SIGKDD International Conference on Knowl-
edge Discovery and Data Mining, pp. 137–146, 2003.
[8] J. Leskovec, A. Krause, C. Guestrin, C. Faloutsos, J. VanBriesen, N. Glance,
“Cost Effective Outbreak Detection in Networks,” Proc. 13th ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, pp. 420–
429, 2007.
[9] A. Yadav, H. Chan, A.X. Jiang, H. Xu, E. Rice, M. Tambe, “Using Social
Networks to Aid Homeless Shelters: Dynamic Influence Maximization under
Uncertainty,” Proc. International Conference on Autonomous Agents and Mul-
tiagent Systems (AAMAS) 2016, Singapore, pp. 740–748.
[10] D. Golovin, A. Krause, “Adaptive Submodularity: Theory and Applications in
Active Learning and Stochastic Optimization,” Journal of Artificial Intelligence
Research, vol. 42, pp. 427–486, 2011.
[11] S. Lei, S. Maniu, L. Mo, R. Cheng, P. Senellart, “Online Influence Maximization,”
Proceedings of the 21th ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, pp. 645–654, August 2015.
[12] J. Schneider, A. Ning. Zhou, E. O. Laumann, “A New HIV Prevention Net-
work Approach: Sociometric Peer Change Agent Selection,” Social Science &
Medicine, vol. 125, pp. 192–202, 2015.
[13] T. W. Valente, P. Pumpuang, “Identifying Opinion Leaders to Promote Behavior
Change,” Health Education and Behavior, pp. 881–896, 2007.
[14] A. Yadav, L. Marcolino, E. Rice, R. Petering, H. Winetrobe, H. Rhoades,
M. Tambe, H. Carmichael, “Preventing HIV Spread in Homeless Populations
Using PSINET,” Proc. 27th Conference on Innovative Applications of Artificial
Intelligence (IAAI), Austin, 2015.
[15] E. Cohen, D. Delling, T. Pajor, R. F. Werneck, “Sketch-Based Influence Max-
imization and Computation: Scaling up with Guarantees,” Proc. 23rd ACM
International Conference on Information and Knowledge Management, pp. 629–
638, 2014.
[16] S. Ross, J. Pineau, S. Paquet, B. Chaib-Draa, “Online Planning Algorithms for
POMDPs,” Journal of Artificial Intelligence Research, pp. 663–704, 2008.
[17] D. Silver, J. Veness, “Monte-Carlo Planning in Large POMDPs,” Proc. Advances
in Neural Information Processing Systems (NIPS), pp. 2164–2172, 2010.
[18] A. Somani, N. Ye, D. Hsu, W. S. Lee, “DESPOT: Online POMDP Planning
with Regularization,” Proc. Advances in Neural Information Processing Systems
(NIPS), pp. 1772–1780, 2013.
[19] S. D. Young, E. Rice, “Online Social Networking Technologies, HIV Knowledge,
and Sexual Risk and Testing Behaviors among Homeless Youth,” AIDS and
Behavior, vol. 15, no. 2, pp. 253–260, 2011.
[20] M. Kim, J. Leskovec, “The Network Completion Problem: Inferring Missing
Nodes and Edges in Networks,” In Proc. of the 2011 SIAM International
Conference on Data Mining, Society for Industrial and Applied Mathematics, pp.
47–58, April 2011.

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 56 HIV Awareness in Homeless Youth

[21] Q. Yan, S. Guo, D. Yang, “Influence Maximizing and Local Influenced Commu-
nity Detection Based on Multiple Spread Model,” Advanced Data Mining and
Applications, pp. 82–95, 2011.
[22] J. P. Cointet, C. Roth, “How Realistic Should Knowledge Diffusion Models Be?”
Journal of Artificial Societies and Social Simulation, vol. 10, no. 3, pp. 1–5, 2007.
[23] A. Yadav, B. Wilder, E. Rice, R. Petering, J. Craddock, A. Yoshioka-Maxwell,
M. Hemler, L. Onasch-Vera, M. Tambe, D. Woo, “Influence Maximization in the
Field: The Arduous Journey from Emerging to Deployed Application,” Interna-
tional Conference on Autonomous Agents and Multiagent Systems (AAMAS) 2017,
Sau Paulo, Brazil, pp. 150–158.
[24] L. P. Kaelbing, M. L. Littman, A. R. Cassandra, “Planning and Acting in Partially
Observable Stochastic Domains,” Artificial Intelligence, Vol. 101, no, 1, pp. 99–
134, 1998.
[25] D. LaSalle, G. Karypis, “Multi-Threaded Graph Partitioning,” Proc. 27th IEEE
International Symposium on Parallel & Distributed Processing (IPDPS), pp. 225–
236, 2013.
[26] B. Wilder, A. Yadav, N. Immorlica, E. Rice, M. Tambe, “Uncharted but Not
Uninfluenced: Influence Maximization with an Uncertain Network,” Proc. of the
2017 International Conference on Autonomous Agents and Multi-Agent Systems,
International Foundation for Autonomous Agents and Multiagent Systems, pp.
1305–1313, 2017.
[27] J. A. Kelly, D. A. Murphy, K. J. Sikkema, T. L. McAuliffe, R. A. Roffman, L. J.
Solomon, R. A. Winett, S. C. Kalichman, “Randomised, Controlled, Community-
Level HIV-Prevention Intervention for Sexual-Risk Behaviour among Homosex-
ual Men in US Cities,” The Lancet, vol. 350, no. 9090. pp. 1500–1505, 1997.

https://doi.org/10.1017/9781108669016.004 Published online by Cambridge University Press

 4
Influence Maximization in the Field
The Arduous Journey from Emerging to Deployed Application

Amulya Yadav, Bryan Wilder, Eric Rice, Robin Petering,

Jaih Craddock, Amanda Yoshioka-Maxwell, Mary Hemler,
Laura Onasch-Vera, Milind Tambe, and Darlene Woo
Center for Artificial Intelligence in Society, University of Southern
California, LA, CA, 90089

Introduction
The process of building a software agent that can be deployed regularly in
the real world to assist underserved communities is very difficult. While
significant attention has been paid in the literature to build agents for innovative
applications, the topic of transitioning agents from an emerging phase in the
lab to a deployed application in the field has not received significant attention
[1]. This chapter illustrates the research challenges and complexities of this
topic by focusing on agents for a particular health-critical domain, i.e., raising
awareness about HIV among homeless youth.
Homeless youth are twenty times more likely to be HIV positive than stably
housed youth, due to high-risk behaviors (such as unprotected sex, survival
sex, sharing drug needles, etc.) [2, 3]. To reduce rates of HIV infection among
youth, many providers of services for homeless youth (henceforth simply
“service providers”) conduct peer-leader-based social network interventions
[4], in which a select group of homeless youth are trained as peer lead-
ers (PLs). This peer-led approach is particularly desirable because service
providers have limited resources and homeless youth tend to distrust adults.
The training program of these peer leaders includes detailed information
about how HIV spreads and what one can do to prevent infection. The PLs
are also taught effective ways of communicating this information to their
peers [5]. Because of their limited financial and human resources, service
providers can only train a small number of these youth and not the entire
population. As a result, the selected peer leaders in these intervention trainings
are tasked with spreading messages about HIV prevention to their peers in

57

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 58 Influence Maximization in the Field

their social circles, thereby encouraging them to move to safer practices.

Using these interventions, service providers aim to leverage social network
effects to disseminate information about HIV and induce behavior change
(increased HIV testing) among more and more people in the social network
of homeless youth.
In fact, there are further constraints that service providers face: behavioral
struggles of homeless youth means that service providers can only train three
or four peer leaders in every intervention. This makes sequential training
the most optimal approach, with groups of up to four youth trained one
after another. They are trained as peer leaders in the intervention, and are
asked information about friendships that they observe in the real-world social
network. This new information about the social network is then used to
improve the selection of the peer leaders for the next intervention. As a result,
the peer leaders for these limited interventions need to be chosen strategically
so that awareness spread about HIV is maximized in the social network of
homeless youth.
The previous chapter discussed HEALER [6] and DOSIM [7], two agents
that assist service providers in optimizing their intervention strategies. These
agents recommend “good” intervention attendees, i.e., homeless youth who
maximize HIV awareness in the real-world social network of youth. In essence,
both HEALER and DOSIM reason strategically about the multiagent system
of homeless youth to select a sequence of three or four of them at a time to
maximize HIV awareness. While HEALER [6] is an adaptive software agent
that solves POMDPs to select the best set of peer leaders, DOSIM [7] uses
robust optimization techniques to find the correct set of peer leaders, even
when the influence probability parameters are not known. Unfortunately, while
earlier research [6, 7] published promising simulation results from the lab,
neither of these agent-based systems have even been tested so far in the real
world. This chapter illustrates that transitioning these agents from the lab into
the real world is not straightforward.
Several questions need to be answered before final deployment of these
agents. First, do peer leaders actually spread HIV information in a homeless
youth social network, and are they able to provide meaningful information
about the social network structure during intervention training (as assumed
by HEALER and DOSIM)? Second, would agents (which use POMDPs and
robust optimization approaches to reason about underlying social networks)
outperform standard techniques used by service providers to select peer
leaders? If they do not, then a large-scale deployment is unwarranted. Third,
which agent out of HEALER or DOSIM performs better in the field? Finally, is
the system robust and/or flexible enough to resolve any unforeseen challenges
that may arise before deployment?

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 Amulya Yadav et al. 59

To answer these questions, it is necessary to conduct real-world pilot tests

before deployment of these agents on a large scale. Indeed, the health-critical
nature of the domain and complex influence spread models used by social
influence maximization agents makes conducting pilot tests even more impor-
tant to validate their real-world effectiveness. This chapter presents results
from three real-world pilot studies involving 173 homeless youth in an
American city. This is an actual test involving word-of-mouth spread of
information, and actual changes in youth behavior in the real world as a
result. To the best of our knowledge, these are the first such pilot studies that
provide head-to-head comparison of different software agent (with POMDP,
robust optimization driven) approaches for social influence maximization,
including a comparison with a baseline approach. Our pilot study results
show that HEALER and DOSIM achieve 184% more information spread
than Degree Centrality (baseline), and do significantly better at inducing
behavior change among homeless youth. Second, we present analyses of
these real-world results, illustrating the strengths and weaknesses of different
influence maximization approaches we compare. Specifically, we illustrate
how HEALER and DOSIM cleverly exploit the community structure of real-
world social networks to outperform Degree Centrality. Third, we present
research challenges revealed in conducting these pilot tests, and propose
solutions to address them. These challenges dispel any misguided notions
about the ease of taking applications from the emerging to the deployed
application phase. Finally, the promising results obtained in these pilot studies
open the door to future deployment of HEALER and DOSIM by service
providers on a regular basis.

Pilot Study Pipeline

Starting in the spring of 2016, we conducted three different pilot studies at
two service providers in a large American city over a seven-month period.
Each pilot study recruited a unique network of youth. Recall that these pilot
studies serve three purposes. First, they help in justifying our assumptions
about whether peer leaders actually disseminate HIV information in their
social network, and whether they provide meaningful information about the
social network structure (i.e., observations) during the intervention training.
Second, these pilot studies help in exposing unforeseen challenges that need
to be solved convincingly before these agents can be deployed in the field.
Third, they provide a head-to-head comparison of two different software agent
approaches for social influence maximization, including a comparison with a
baseline approach.

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 60 Influence Maximization in the Field

Figure 4.1 Real-world pilot study pipeline

Each of these pilot studies had a different intervention mechanism, i.e., a

different way of selecting actions (or a set of K peer leaders). The first and
second studies used HEALER and DOSIM (respectively) to select actions,
whereas the third study served as the control group, where actions were
selected using Degree Centrality (i.e., picking K nodes in order of decreasing
degrees). We chose Degree Centrality (DC) as the control group mechanism,
because this is the current modus operandi of service providers in conducting
these network based interventions [8].
Pilot Study Process The pilot study process consisted of five sequential
steps. Figure 4.1 illustrates these five steps.

1. Recruitment: First, we recruited homeless youth from a service provider

into our study (n = 173 at baseline across the three tests). We provided
youth with general information about our study and what was expected of
them (i.e., if selected as a peer leader, they would be expected to
disseminate information among their peers). The youth took a 20-minute
baseline survey, which enabled us to determine their current risk-taking
behaviors (e.g., they were asked about the last time they got an HIV test).
Every youth was given a 20 USD gift card as compensation for
participating in the pilot study. All study procedures were approved by our
Institutional Review Board.
2. Network Generation: After recruitment, the friendship-based social
network that connects these homeless youth was generated. We relied on
two information sources to generate this network: (1) online contacts of
homeless youth and (2) field observations made by the authors while
collecting baseline survey data. To expedite the network generation phase,

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 Amulya Yadav et al. 61

online contacts of homeless youth were used (via a software application

that the youth were asked to use) to build a first approximation of the
real-world social network of homeless youth. This network was then
refined using field observations (about additional real-world friendships)
made by the authors. All edges inferred in this manner were assumed to be
certain edges. More information on uncertain edges is provided later.
3. Interventions: Next, the generated network was used by the software
agents to select actions (i.e., K peer leaders) for T stages. In each stage, an
action was selected using the pilot’s intervention strategy. The K peer
leaders of this chosen action were then trained as peer leaders (i.e.,
informed about HIV) by pilot study staff during the intervention. These
peer leaders also revealed more information (i.e., provide observation)
about newer friendships that we did not know about. These friendships
were incorporated into the network, so that the agents could select better
actions in the next stage of interventions. Every peer leader was given a 60
USD gift card.
Only four to six peer leaders could be trained at once, and thus peer
change agents were enrolled in three subsequent rounds. Training was
delivered by three facilitators: Rice, his PhD students with MSWs, or
MSW interns. Training was interactive and broken into six hour-long
modules on: the mission of peer advocacy, sexual health, HIV prevention,
communication skills, leadership skills, and self-care. Peer leaders were
asked to focus their communications on their social ties, particularly other
youth at the drop-in center, and to promote regular HIV testing.
The training was designed to be engaging and included a variety of
learning activities such as group discussion, games, journaling and
reflection, experiential learning, and role-playing. The training minimized
lecture-based learning. The small group setting was critical for
maintaining a safe and manageable space for youth to learn and reflect.
Further, the training was developed from a youth empowerment model.
Consistent language was used throughout to reiterate the participants’ role
as a leader and advocate within their community.
4. Follow-Up: The follow-up phase consisted of meetings during which the
peer leaders were asked about any difficulties they faced in talking to their
friends about HIV. They were given further encouragement to keep
spreading HIV awareness among their peers. These follow-up meetings
occured on a weekly basis, for a period of seven weeks after Step 3 ended.
5. Analysis: For analysis, we conducted in-person surveys. Of the initial 173
youth, 122 (71%) were surveyed a second time, one month after their first
survey and after all initial intervention training sessions were completed

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 62 Influence Maximization in the Field

(but before the completion of Step 4). Every youth in our study was given
a 25 USD gift card to show up for these surveys. During the surveys they
were asked if some youth from within the pilot study talked to them about
HIV prevention methods after the pilot study began. Their answers helped
determine if information about HIV reached them in the social network or
not. Thus, these surveys were used to find out the number of youth who
became informed about HIV as a result of our interventions. Furthermore,
they were asked to take the same survey about HIV risk that they took
during recruitment. These post-intervention surveys enabled us to compare
HEALER, DOSIM, and DC in terms of information spread (i.e., how
successful were the agents in spreading HIV information through the
social network) and behavior change (i.e., how successful were the agents
in causing homeless youth to test for HIV), the two major metrics that we
used in our evaluation section. These two metrics were based on
self-reports. First, youth were asked when (if ever) they had been tested for
HIV; answer choices included never, in the past month, three to six months
ago, and more than six months ago. Second, commmunication was
assessed with a series of eight questions about their attitudes toward
conversations with peer leaders (if they had such a conversation; if not,
they could select “I didn’t have a conversation with a Have You Heard
Peer Leader about HIV or AIDS.” Those who endorsed the
non-communication responses were scored as not receiving infromation
directly from a peer leader.

We provide these behavior change results in order to quantify the true

impact of these social influence maximization agents in the homeless youth
domain. In these results, we measured behavior change by asking youth if
they have taken an HIV test at baseline and repeating this question during
the follow-up surveys. If the youth reported taking an HIV test at one month
(after interventions) but not at baseline and that youth also reported getting
informed about HIV from a peer leader, we attributed this behavior change to
our intervention. This allowed us to measure whether our interventions led to
a reduction in risk behaviors.
Uncertain network parameters. While there exist many link prediction
techniques [9] to infer uncertain edges in social networks, the efficacy of these
techniques is untested on homeless youth social networks. Therefore, we took
a simpler, less “risky” approach – each edge not created during the network
generation phase (i.e., Step 2 above) was added to the network as an uncertain
edge. Thus, after adding these uncertain edges, the social network in each pilot
study became a completely connected network, consisting of certain edges
(inferred from Step 2) and uncertain edges. The existence probability on each

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 Amulya Yadav et al. 63

Figure 4.2 Information spread with pe on HEALER’s pilot network

uncertain edge was set to u = 0.01. Our approach to adding uncertain edges
ensures that no potential friendship is missed in the social network because of
our lack of accurate knowledge.
Getting propagation probabilities (pe ) values was also challenging. In the
HEALER’s pilot, service providers estimated that the true pe value would be
somewhere around 0.5. Since the exact value was unknown, we assumed an
interval of [0.4, 0.8] and simulated HEALER’s performance with pe values in
this range. Figure 4.2 shows how information spread achieved by HEALER on
its pilot study network is relatively stable in simulation for pe values around
0.5. The y-axis shows the information spread in simulation and the x-axis
shows increasing pe values. This figure shows that information spread achieved
by HEALER varied by ∼11.6% with pe in the range [0.4, 0.8]. Since influence
spread is relatively stable in this range, we selected pe = 0.6 (the mid-point
of [0.4, 0.8]) on all network edges. Later in this chapter we provide ex post
justification for why pe = 0.6 was a good choice, at least for this pilot study.
In DOSIM’s pilot, we did not have to deal with the issue of assigning
accurate pe values to edges in the network. This is because DOSIM can work
with intervals in which the exact pe is assumed to lie. For the pilot study, we
used the same interval of [0.4, 0.8] to run DOSIM. Finally, the control group
pilot study did not require finding pe values, as peer leaders were selected using
Degree Centrality, which does not require knowledge of pe .

Results from the Field

We now provide results from all three pilot studies. In each study, three
interventions were conducted (or T = 3), i.e., Step 3 of the pilot study process

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 64 Influence Maximization in the Field

Figure 4.3 Set of surveyed non–peer leaders

(Figure 4.1) was repeated three times. The actions (i.e., set of K peer leaders)
were chosen using intervention strategies (policies) provided by HEALER
[6], DOSIM [7], and Degree Centrality (DC) in the first, second, and third
pilot studies, respectively. Recall that we provide comparison results on two
different metrics. First, we provide results on information spread, i.e., how
well different software agents were able to spread information about HIV
through the social network. Second, even though HEALER and DOSIM do not
explicitly model behavior change in their objective function (both maximize
the information spread in the network), we provide results on behavior change
among homeless youth, i.e., how successful the agents were in inducing
behavior change among homeless youth.
Figure 4.3 shows a matrix diagram that explains the results that we collect
from the pilot studies. To begin with, we exclude PLs from all our results and
focus only on non–peer leaders (non-PLs). This is done because PLs cannot
be used to differentiate the information spread (and behavior change) achieved
by HEALER, DOSIM, and DC. In terms of information spread, all PLs are
informed about HIV directly by study staff in the intervention trainings. In
terms of behavior change, the proportion of PLs who change their behavior
does not depend on the strategies recommended by HEALER, DOSIM, and
DC. Thus, Figure 4.3 shows a matrix diagram of the set of all non-PLs (who
were surveyed at the end of one month). This set of non-PLs can be divided
into four quadrants based on (1) whether or not they were informed about HIV
(by the end of one-month surveys in Step 5 of Figure 4.1) and (2) whether or
not they were already tested for HIV at baseline (i.e., during recruitment, they
reported that they had been tested for HIV in the last six months).
For information spread results, we report on the percentage of youth in this
big rectangle, who were informed about HIV by the end of one month (i.e.,
boxes A + B as a fraction of the big box). For behavior change results, we
exclude youth who were already tested at baseline (as they do not need to

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 Amulya Yadav et al. 65

Table 4.1. Logistic details of different pilot studies

HEALER DOSIM DC

Youth Recruited 62 56 55
PL Trained 17.7% 17.85% 20%
Retention % 73% 73% 65%
Avg. Observation Size 16 8 15

undergo behavior change because they are already exhibiting desired behavior
of testing). Thus, we only report on the percentage of untested informed youth
(i.e., box B), who now tested for HIV (i.e., changed behavior) by the end of
one month (which is a fraction of youth in box B). We do this because we can
only attribute conversions (to testers) among youth in box B (Figure 4.3) to
strategies recommended by HEALER and DOSIM (or the DC baseline). For
example, non-PLs in box D who convert to testers (due to some exogenous
reasons) cannot be attributed to HEALER or DOSIM’s strategies (as they
converted to testers without getting HIV information).
Study details. Table 4.1 shows details of the pilot studies. It shows that
the three pilots had fairly similar conditions, (1) all three pilots recruited ∼60
homeless youth; (2) peer leader training was done on 17–20% of these youth,
which is recommended in social sciences literature [4]; and (3) retention rates
of youth (i.e., percentage of youth showing up for post-intervention surveys)
were fairly similar (∼70%) in all three pilots. This figure also shows that
peer leaders provided information about 13 uncertain friendships on average
in every intervention stage (across all three pilot studies), which validates
HEALER and DOSIM’s assumption that peer leaders provide observations
about friendships [6, 7].
Information spread. Figure 4.4 compares the information spread achieved
by HEALER, DOSIM, and DC in the pilot studies. The x-axis shows the three
different intervention strategies and the y-axis shows the percentage of non-
PLs to whom information spread (box A + B as a percentage of total number
of non–peer leaders in Figure 4.3). This figure shows that PLs chosen by
HEALER and DOSIM are able to spread information among ∼70% of the
non-PLs in the social network by the end of one month. Surprisingly, PLs
chosen by DC were only able to inform ∼27% of the non-PLs. This result
is surprising, as it means that HEALER and DOSIM’s strategies were able
to improve over DC’s information spread by over 184%. We now explain
reasons behind this significant improvement in information spread achieved
by HEALER and DOSIM over DC.

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 66 Influence Maximization in the Field

Informed Uninformed

% of Non–Peer Leaders Informed

100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
HEALER DOSIM DC
Different Algorithms

Figure 4.4 Comparison of information spread among non–peer leaders

25
% of edges between peer leaders

20

15

10

0
HEALER DOSIM DC
Different pilots

Figure 4.5 Percentage of edges between PLs

Figure 4.5 illustrates a big reason behind DC’s poor performance. The
x-axis shows different pilots and the y-axis shows what percentage of network
edges were redundant, i.e., they connected two peer leaders. Such edges are
redundant, as both its nodes (peer leaders) already have the information. This
figure shows that redundant edges accounted for only 8% (and 4%) of the total
edges in HEALER and DOSIM pilot studies. On the other hand, 21% of the
edges in DC’s pilot study were redundant. Thus, DC’s strategies pick PLs in
a way that creates a lot of redundant edges, whereas HEALER picks PLs that
create one-third fewer redundant edges. DOSIM performs best in this regard,
by selecting nodes that create the fewest redundant edges (∼5 times fewer than
DC and half as many as HEALER), and is the key reason behind its good

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 Amulya Yadav et al. 67

25

% of edges between communities

20

15

10

0
HEALER Net DOSIM Net DC Net
Different Networks

Figure 4.6 Community structure comparison

performance in Figure 4.4. Concomitantly to the presence of redundant edges,

HEALER also spreads out its PLs selection across different communities
within the homeless youth network, which also aids in information spreading,
as discussed later here.
Figure 4.6 shows the community structure of the three pilot studies’ social
networks. To generate this figure, the three networks were partitioned into
communities using METIS [10], an off-the-shelf graph-partitioning tool. We
partitioned each network into four different communities (as shown in Figure
4.8) to match the number of PLs (i.e., K = 4) chosen in every stage. The x-axis
shows the three pilot study networks and the y-axis shows the percentage
of edges that go across these four communities. This figure shows that all
three networks can be fairly well represented as a set of reasonably disjointed
communities, as only 15% of edges (averaged across all three networks) went
across the communities. Next, we show how HEALER and DOSIM exploit
this community structure by balancing their efforts across these communities
simultaneously to achieve greater information spread as compared to DC.
Figure 4.7 illustrates patterns of PL selection (for each stage of intervention)
by HEALER, DOSIM, and DC across the four different communities uncov-
ered in Figure 4.6. Recall that each pilot study consisted of three stages of
intervention (each with four selected PL). The x-axis shows the three different
pilots. The y-axis shows what percentage of communities had a PL chosen
from within them. For example, in DC’s pilot, the chosen PL covered 50%
(i.e., two out of four) communities in the first stage, 75% (i.e., three out of four)
communities in the second stage, and so on. This figure shows that HEALER’s
chosen PLs cover all possible communities (i.e., 100% communities touched)

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 68 Influence Maximization in the Field

100
1st Intervention
2nd Intervention

% of communities touched
80 3rd Intervention

60

40

20

0
HEALER DOSIM DC
Different Pilots

Figure 4.7 Exploiting community structure of real-world networks

Figure 4.8 Four partitions of DC’s pilot network

in the social network in all three stages. On the other hand, DC concentrates
its efforts on just a few clusters in the network, leaving approximately 50%
communities untouched (on average). Therefore, while HEALER ensures that
its chosen PLs covered most real-world communities in every intervention,
the PLs chosen by DC focused on a single or a few communities in each
intervention. This further explains why HEALER is able to achieve greater
information spread, as it spreads its efforts across communities, unlike DC.
While DOSIM’s coverage of communities is similar to DC, it outperforms
DC because it generates five times fewer redundant edges than does DC
(Figure 4.5).

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 Amulya Yadav et al. 69

Converted Not Converted

% of informed and untested youth

100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
HEALER DOSIM DC
Different algorithms

Figure 4.9 Behavior change comparison

Behavior change. Figure 4.9 compares behavior change observed in home-

less youth in the three pilot studies. The x-axis shows different intervention
strategies, and the y-axis shows the percentage of non-PLs who were untested
for HIV at baseline and were informed about HIV during the pilots (i.e., youth
in box B in Figure 4.3). This figure shows that PLs chosen by HEALER (and
DOSIM) converted 37% (and 25%) of the youth in box B to HIV testers. In
contrast, PLs chosen by DC did not convert any youth in box B to testers.
DC’s information spread reached a far smaller fraction of youth (Figure 4.4),
and therefore it is unsurprising that DC did not get adequate opportunity to
convert anyone of them to testing. This shows that even though HEALER and
DOSIM do not explicitly model behavior change in their objective function,
the agents’ strategies still end up outperforming DC significantly in terms of
behavior change.

Challenges Uncovered
This section highlights research and methodological challenges that we uncov-
ered while deploying these agent based interventions in the field. While
handling these challenges in a principled manner is a subject for future
research, we explain some heuristic solutions used to tackle these challenges in
the three pilot studies (which may help in addressing the longer-term research
challenges).
Research challenges. While conducting interventions, we often encounter
an inability to execute actions (i.e., conduct intervention with chosen peer

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 70 Influence Maximization in the Field

leaders), because a subset of the chosen peer leaders may fail to show up
for the intervention (because they may get incarcerated, or find temporary
accommodation). Handling this inability to execute actions in a principled
manner is a research challenge. Therefore, it is necessary that algorithms and
techniques developed for this problem are robust to these errors in execution of
intervention strategy. Specifically, we require our algorithms to be able to come
up with alternate recommendations for peer leaders, when some homeless
youth in their original recommendation are not found. We now explain how
HEALER, DOSIM, and DC handle this challenge by using heuristic solutions.
Recall that for the first pilot, HEALER’s intervention strategies were found
by using online planning techniques for POMDPs [6]. Instead of offline
computation of the entire policy (strategy), online planning only finds the
best POMDP action (i.e., selection of K network nodes) for the current belief
state (i.e., probability distribution over state of influence of nodes). Upon
reaching a new belief state, online planning again plans for this new belief. This
interleaving of planning and execution works to our advantage in this domain,
as every time we have a failure that was not anticipated in the POMDP model
(i.e., a PL that was chosen in the current POMDP action did not show up), we
can recompute a policy quickly by marking these unavailable nodes, so that
they are ineligible for future peer leader selection. After recomputing the plan,
the new peer leader recommendation is again given to the service providers to
conduct the intervention.
For the second pilot study, we augmented DOSIM to account for unavail-
able nodes by using its computed policy to produce a list of alternates for each
peer leader. This alternate list ensures that, unlike HEALER, DOSIM does
not require rerunning in the event of a failure. Thus, if a given peer leader
does not show up, then study staff work down the list of alternates to find a
replacement. DOSIM computes these alternates by maintaining a parameter qv
(for each node v), which gives the probability that node v will show up for the
intervention. This qv parameter enables DOSIM to reason about the inability
to execute actions, thereby making DOSIM’s policies robust to such failures.
To compute the alternate for v, we condition on the following event σv : node v
fails to show up (i.e., set qv = 0), while every other peer leader u shows up with
probability qu . Conditioned on this event σv , we find the node that maximizes
the conditional marginal gain in influence spread, and use it as the alternate for
node v. Hence, each alternate is selected in a manner that is robust with respect
to possible failures on other peer leader nodes. Finally, in the DC pilot, in case
of a failure, the node with the next highest degree is chosen as a peer leader.
Methodological challenges. A methodological challenge was to ensure a
fair comparison of the performance of different agents in the field. In the real

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 Amulya Yadav et al. 71

Table 4.2. Similarity of social networks in different

pilot studies

HEALER DOSIM DC

Network Diameter 8 8 7
Network Density 0.079 0.059 0.062
Avg. Clustering Coefficient 0.397 0.195 0.229
Avg. Path Length 3.38 3.15 3.03
Modularity 0.568 0.568 0.602

world, HEALER, DOSIM, and DC could not be tested on the same network,
as once we disseminate HIV messages in one network as part of one pilot
study, fewer youth are unaware about HIV (or uninfluenced) for the remaining
pilots. Therefore, each agent (HEALER, DOSIM, or DC) is tested in a different
pilot study with a different social network (possibly with a different structure).
Since HEALER’s, DOSIM’s, and DC’s performances are not compared on the
same network, it is important to ensure that HEALER and DOSIM’s superior
performance (observed in Figure 4.4) is not due to differences in network
structure or any extraneous factors.
First, we compare several well-known graph metrics for the three distinct
pilot study social networks. Table 4.2 shows that most metrics are similar on
all three networks, which establishes that the social networks generated in
the three pilot studies were structurally similar. This suggests that comparison
results would not have been very different had all three algorithms been tested
on the same network. Next, we attempt to show that HEALER and DOSIM’s
superior performance (Figure 4.4) was not due to extraneous factors.
Figure 4.10 compares information spread achieved by PLs in the actual
pilot studies with that achieved by the same PLs in simulation. The simulation
(averaged over 50 runs) was done with propagation probability set to pe = 0.6
in our influence model. The x-axis shows the different pilots and the y-axis
shows the percentage of non-PLs informed in the pilot study networks. First,
this figure shows that information spread in the simulation closely mirrors pilot
study results in the HEALER’s and DC’s pilots (∼10% difference), whereas it
differs greatly in the DOSIM’s pilot. This shows that using pe = 0.6 as the
propagation probability modeled the real-world process of influence spread
in the HEALER’s and DC’s pilot study network fairly well, whereas it was
not a good model for the DOSIM’s pilot network. This further suggests that
information spread achieved in the real world (at least in HEALER’s and
DC’s pilot studies) was indeed due to the respective strategies used, and

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 72 Influence Maximization in the Field

80
Pilot Study
70

% of non-PLs informed
In Simulation
60
50
40
30
20
10
0
HEALER Pilot DOSIM Pilot DC Pilot
Different pilot studies

Figure 4.10 Simulation of information spread

not some extraneous factors. In other words, DC’s poor performance may
not be attributed to some real-world external factors at play, since its poor
performance is mimicked in simulation results (which are insulated from real-
world external factors) as well. Similarly, HEALER’s superior performance
may not be attributed to external factors working in its favor, for the same
reason.
On the other hand, since DOSIM’s performance in the pilot study does
not mirror simulation results in Figure 4.10, it suggests the role of some
external factors, which were not considered in our models. However, the
comparison of simulation results in this figure is statistically significant (p-
value = 9.43E − 12), which shows that even if DOSIM’s performance in the
pilot study matched its simulation results, i.e., even if DOSIM achieved only
∼ 40% information spread in its pilot study (as opposed to the 70% spread that
it actually achieved), it would still outperform DC by ∼ 33%.
Having established that DC’s poor performance in the field was not due to
any external factors, we now show that it also was not tied to some peculiar
property/structure of the network used in its pilot study. Figure 4.11 compares
information spread achieved by different agents (in simulation over 50 runs)
when each agent was run on DC’s pilot study network. Again, the simulation
was done using pe = 0.6 as propagation probability, which was found to
be a reasonable model for real-world influence spread in DC’s network (see
Figure 4.10). The x-axis in Figure 4.11 shows different algorithms being run
on DC’s pilot study network (in simulation). The y-axis shows the percentage
of non-PLs informed. This figure shows that even on DC’s pilot study network,
HEALER (and DOSIM) outperform DC in simulation by ∼ 53% (and 76%)
(p − value = 9.842E − 31), thereby establishing that HEALER’s and DOSIM’s

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 Amulya Yadav et al. 73

60

50

% of non-PLs informed
40

30

20

10

0
DC (Real) DC (Sim) HEALER DOSIM
(Sim) (Sim)
Different algorithms

Figure 4.11 Comparison on the DC’s network

6
5% Edges Removed
% difference in information

5
spread after perturbation

5% Edges Added

0
HEALER DOSIM DC
Different networks

Figure 4.12 Comparison on perturbed networks

improvement over DC was not due to specific properties of the networks in

their pilot studies, i.e., HEALER’s and DOSIM’s superior performances may
not be attributed to specific properties of networks (in their pilot studies)
working in their favor. In other words, this shows that DC’s poor performance
may not be attributed to peculiarities in its network structure working against
it, as otherwise this peculiarity should have affected HEALER’s and DOSIM’s
performances as well, when they are run on DC’s pilot study network (which
does not happen as shown in Figure 4.11).
Figure 4.12 shows information spread achieved by PLs (chosen in the pilot
studies) in simulation (50 runs), averaged across 30 different networks, which
were generated by perturbation of the three pilot study networks. The x-axis

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 74 Influence Maximization in the Field

shows the networks that were perturbed. The y-axis shows the percentage
difference in information spread achieved on the perturbed networks, in
comparison with the unperturbed network. For example, adding 5% edges
randomly to HEALER’s pilot study network results in only ∼2% difference
(p-value = 1.16E − 08) in information spread (averaged across 30 perturbed
networks). These results support the view that HEALER’s, DOSIM’s, and
DC’s performances are not due to their pilot study networks being on the
knife’s edge in terms of specific peculiarities. Thus, HEALER and DOSIM
outperform DC on a variety of slightly perturbed networks as well.

Conclusion, Limitations, and Lessons Learned

This chapter illustrates challenges faced in transitioning agents from an emerg-
ing phase in the lab to a deployed application in the field. It presents first-of-its-
kind results from three real-world pilot studies, involving 173 homeless youth
in an American city. Conducting these pilot studies underlined their importance
in this transition process: they are crucial milestones in the arduous journey
of an agent from an emerging phase in the lab, to a deployed application in
the field. The pilot studies helped in answering several questions that were
raised in the first section of this chapter. First, we learned that peer-leader-
based interventions are indeed successful in spreading information about HIV
through a homeless youth social network (as seen in Figure 4.4). Moreover,
we learned that peer leaders are very adept at providing lots of information
about newer friendships in the social network (Table 4.1), which helps software
agents refine their future strategies.
These pilot studies also helped establish the superiority (and hence, their
need) of HEALER and DOSIM – we are using complex agents (involving
POMDPs and robust optimization), and they outperform DC (the modus
operandi of conducting peer-led interventions) by 184% (Table 4.1, Figure
4.9). The pilot studies also helped us gain a deeper understanding of how
HEALER and DOSIM beat DC (shown in Figures 4.5–4.7) – by minimizing
redundant edges and exploiting community structure of real-world networks.
Out of HEALER and DOSIM, the pilot tests do not reveal a significant
difference in terms of either information spread or behavior change (Table 4.1,
Figure 4.9). Thus, carrying either of them forward would lead to significant
improvement over the current state-of-the-art techniques for conducting
peer-leader-based interventions. However, DOSIM runs significantly faster
than HEALER (∼40times), and thus is more beneficial in time-constrained
settings [7].

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 Amulya Yadav et al. 75

These pilot studies also helped uncover several key challenges (e.g.,
inability to execute actions, estimating propagation probabilities), which were
tackled in the pilot studies using heuristic solutions. However, handling these
challenges in a principled manner is a subject for future research. Thus, while
these pilot studies opened the door to future deployment of these agents in
the field (by providing positive results about the performance of HEALER
and DOSIM), they also revealed some challenges that need to be resolved
convincingly before these agents can be deployed.
As with any pilot study, some limitations must be acknowledged. The first
is that we have a relatively small sample size. While we can demonstrate
impressive effect size with these results, we cannot say that these results are
statistically significant. Second, we must be cautious about the generalizability
of this work. To the extent that homeless populations differ across urban,
suburban, and rural settings, this intervention may or may not be effective
outside of urban populations of homeless youth. It would also be valuable to
include a second control group in which no intervention is delivered to assess
the impact of repeated observations of homeless youth with respect to HIV
testing and communication. These limitations, however, are typical of many
intervention development pilot studies. Despite these limitations, we are very
excited and encouraged by these results. Our community partners are likewise
excited. At the time of writing this chapter, we have just begun to enroll youth
into a large-scale study to establish statistical significance and the efficacy of
this intervention model.

Acknowledgments
This research was supported by MURI Grant W911NF-11-1-0332 & NIMH
Grant R01-MH093336.

References
[1] M. Jain, B. An, and M. Tambe. An overview of recent application trends at the
AAMAS conference: Security, sustainability and safety. AI Magazine, 33(3):14,
2012.
[2] N. H. Council. HIV/AIDS among persons experiencing homelessness: Risk
factors, predictors of testing, and promising testing strategies. www.nhchc.org/
wp-content/uploads/2011/09/InFocus_Dec2012.pdf, Dec. 2012.
[3] CDC. HIV Surveillance Report. www.cdc.gov/hiv/pdf/g-l/hiv_surveillance_
report_vol_25.pdf, Mar. 2013.

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 76 Influence Maximization in the Field

[4] E. Rice. The positive role of social networks and social networking technology in
the condom-using behaviors of homeless young people. Public Health Reports,
125(4):588, 2010.
[5] E. Rice, E. Tulbert, J. Cederbaum, A. B. Adhikari, and N. G. Milburn. Mobilizing
homeless youth for HIV prevention: A social network analysis of the acceptability
of a face-to-face and online social networking intervention. Health Education
Research, 27(2):226, 2012.
[6] A. Yadav, H. Chan, A. X. Jiang, H. Xu, E. Rice, and M. Tambe. Using
social networks to aid homeless shelters: Dynamic influence maximization under
uncertainty. In International Conference on Autonomous Agents and Multiagent
Systems (AAMAS), 2016.
[7] B. Wilder, A. Yadav, N. Immorlica, E. Rice, and M. Tambe. Uncharted but not
uninfluenced: Influence maximization with an uncertain network. In International
Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2017.
[8] T. W. Valente. Network interventions. Science, 337(6090):49–53, 2012.
[9] M. Kim and J. Leskovec. The network completion problem: Inferring missing
nodes and edges in networks. In Proceedings of the SIAM Conference on Data
Mining. SIAM, 2011.
[10] D. LaSalle and G. Karypis. Multi-threaded graph partitioning. In Parallel &
Distributed Processing (IPDPS), 2013 IEEE 27th International Symposium on
Parallel and Distributed Processing, pages 225–236. IEEE, 2013.

https://doi.org/10.1017/9781108669016.005 Published online by Cambridge University Press

 5
Influence Maximization with Unknown
Network Structure
Bryan Wilder, Nicole Immorlica, Eric Rice, and Milind Tambe

Introduction
In contexts ranging from health to international development, practitioners
have used the social network of their target population to spread information
and change behavior. The challenge is to identify the most influential people.
While previous work has delivered computationally efficient algorithms for
this influence maximization problem [1–3], this work assumes that the social
network is given explicitly as input. However, in many real-world domains,
the network is not initially known and must be gathered via laborious
field observations. For example, collecting network data from vulnerable
populations such as homeless youth, while crucial for health interventions,
requires significant time spent gathering field observations [4]. Social media
data are often unavailable when access to technology is limited, for instance in
developing countries or with vulnerable populations. Even when such data are
available, it often includes many weak links that are not effective at spreading
influence [5]. For instance, a person may have hundreds of Facebook friends
whom they barely know. In principle, the entire network could be reconstructed
via surveys, and then existing influence maximization algorithms applied.
However, exhaustive surveys are very labor-intensive and often considered
impractical [6]. For influence maximization to be relevant to many real-world
problems, it must contend with limited information about the network, not just
limited computation.
The major informational restriction is the number of nodes that may be
surveyed to explore the network. Thus, a key question is: How can we find
influential nodes with a small number of queries? Existing fieldwork uses
heuristics, such as sampling some percentage of the nodes and asking them to
nominate influencers [6]. We formalize this problem as exploratory influence
maximization and seek a principled algorithmic solution, i.e., an algorithm

77

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press

 78 Influence Maximization with Unknown Network Structure

that makes a small number of queries and returns a set of seed nodes that
are approximately as influential as the globally optimal seed set. To the best
of our knowledge, no previous work directly addresses this question from an
algorithmic perspective (we survey the closest work in the third section).
Real-world networks often have strong community structure, where nodes
form tightly connected subgroups that are only weakly connected to the rest
of the network [7, 8]. Consequently, influence mostly propagates locally.
Community structure has been used to develop computationally efficient
influence maximization algorithms [9, 10]. Here, we use it to design a
highly information-efficient algorithm. We make three contributions. First, we
introduce exploratory influence maximization and show that it is intractable
for general graphs. Second, we present the ARISEN algorithm, which exploits
community structure to find influential nodes. Third, experiments on both
synthetic and real networks verify ARISEN’s performance.

Exploratory Influence Maximization

As a motivating example, consider a homeless youth drop-in center that wishes
to spread HIV prevention information [4]. It would try to select the most
influential peer leaders to spread information, but their social network is not
initially known. Constructing the network requires a laborious survey [4]. Our
motivation is to mitigate this effort by querying only a few youth. Such queries
require much less time than the day-long training peer leaders receive. We now
formalize this problem.

Influence maximization: The influence maximization problem [11] starts

with a graph G = (V, E), where |V|= n and |E|= m. We assume that G is
undirected; social links are typically reciprocal [12]. An influencer selects K
seed nodes, aiming to maximize the expected size of the resulting influence
cascade. We assume that influence propagates according to the independent
cascade model (ICM), the most prevalent model in the literature. Initially, all
nodes are inactive except for the seeds. When a node activates, it independently
activates each of its neighbors with probability q. q is typically assumed
to be the same for all edges [1, 11, 13]. Let f (S ) denote the expected
number of activated nodes with seed set S ⊆ V. The objective is to compute
arg max|S |≤K f (S ).

Local information: The edge set E is not initially known. Instead, the
algorithm explores portions of the graph using local operations. We use the

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press

 Bryan Wilder et al. 79

popular “Jump-Crawl” model [14], where the algorithm may either jump to a
uniformly random node or crawl along an edge from an already surveyed node
to one of its neighbors. When visited, a node reveals all of its edges. We say
that the query cost of an algorithm is the total number of nodes visited using
either operation. Our goal is to find influential nodes with a query cost that is
much less than n, the total number of nodes.

Stochastic Block Model (SBM): In our formal analysis, we assume that the
graph is drawn from the SBM. The SBM originated in sociology [15] and
lately has been intensively studied in computer science and statistics (see e.g.,
[16–18]). In the SBM, the network is partitioned into disjoint communities
C1 . . . .C L . Each within-community edge is present independently with prob-
ability pw and each between-community edge is present independently with
probability pb . Recall that the Erdős-Rényi random graph G(n, p) is the graph
on n nodes where every edge is independently present with probability p.
In the SBM, community Ci is internally drawn as G(|Ci |, pw ) with additional
random edges to other communities. While the SBM is a simplified model,
our experimental results show that ARISEN also performs well on real-world
graphs. ARISEN takes as input the parameters n, pw , and pb , but is not given
any prior information about the realized draw of the network. It is reasonable
to assume that the model parameters are known, since they can be estimated
using existing network data from a similar population (in our experiments, we
show that this approach works well).

Objective: We compare to the globally optimal solution, i.e, the best perfor-
mance if the entire network were known. Let fE (S ) give the expected number
of nodes influenced by seed set S when the set of realized edges are E. Let
A(E) be the (possibly random) seed set containing our algorithm’s selections
given edge set E. Let OPT be the expected value of the globally optimal
solution that seeds K nodes. We aim for an algorithm that maximizes the value
E[ fE (A(E))]/OPT , where the expectation is over the randomness in the graph,
the algorithm’s choices, and the ICM.

Related Work
First, Yadav et al. [13] and Wilder et al. [19] studied dynamic influence
maximization over a series of rounds. Some edges are “uncertain” and are
only present with some probability; the algorithm can gain information about
these edges in each round. However, most edges are known in advance. By

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press

 80 Influence Maximization with Unknown Network Structure

contrast, our work does not require any known edges. Mihara et al. [20]
also consider influence maximization over a series of rounds, but in their
work the network is initially unknown. In each round, the algorithm makes
some queries, selects some seed nodes, and observes all of the nodes that are
activated by its chosen seeds. The ability to observe activated nodes makes
our problem incomparable with theirs, because activations can reveal a great
deal about the network and give the algorithm information that even the global
optimizer does not have (their benchmark does not use the activations). Further,
activations are unobservable in many domains (e.g., medical ones) for privacy
and legal reasons. Carpentier and Valko [21] study a bandit setting where the
algorithm does not know the network but observes the number of activations
at each round. However, in applications of interest (e.g., HIV prevention) it is
more feasible to gather some network data in advance than to conduct many
low-reward trial campaigns.
Another line of work concerns local graph algorithms, where a local algo-
rithm only uses the neighborhoods around individual nodes. Borgs, Brautbar,
Chayes and Lucier [22] study local algorithms for finding the root node in
a preferential attachment graph and for constructing a minimum dominating
set. Other work, including Bressen et al. [23] and Borgs et al. [24], aims to
find nodes with high PageRank using local queries. These algorithms are not
suitable for our problem, since a great deal of previous work has observed
that picking high PageRank nodes as seeds can prove highly suboptimal for
influence maximization [1, 2, 25]. Essentially, PageRank identifies a set of
nodes that are individually central, while influence maximization aims to
find a set of nodes that are collectively best at diffusing information. We
also emphasize that our technical approach is entirely distinct from work on
PageRank. Lastly, Alon et al. [26] attempt to infer a ground truth from the
opinions of agents with an unknown social network, a different task than
influence maximization with correspondingly distinct techniques.

The ARISEN Algorithm

We now introduce our main contribution, the ARISEN algorithm (Approxi-
mating with Random walks to Influence a Socially Explored Network); see
Figure 5.1. The idea behind ARISEN (Algorithm 1) is to sample a set of T
random nodes {v1 . . . vT } from G and explore a small subgraph Hi around each
vi by taking R steps of a random walk (Lines 1–3). R, T and B (explained later)
are inputs. Intuitively, T should be greater than K (the number of seeds) so we
can be sure of sampling each of the largest K communities. The subgraphs Hi

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press

 Bryan Wilder et al. 81

1 ARISEN 2 InitializeWeights
Lines 1–3 Lines 1–6

A network with 7 Each block is a sample Weights after

communities from a community InitializeWeights Line 6

4 RefineWeights
3 InitializeWeights
5 ARISEN
Line 7
Line 7

Final sampled seeds Output of RefineWeights Output of InitializeWeights

Figure 5.1 Example run of ARISEN with K = 3 (explained further in text). Each
block is one sample, with current weight proportional to its height (e.g., in Frame
2, C5 has one sample with very high weight).

Algorithm 1 ARISEN
Require: R, T, B, K, n, pw , pb
1: for i = 1...T do
2: Sample vi uniformly
random from G.
3: Hi = R nodes on a
random walk from vi .
4: end for
5: w = InitializeWeights

6: w = RefineWeights(w)
iid 
7: Sample u1 ...uK ∼ w
8: return u1 ...uK

are used to construct a weight vector w where wi gives the weight associated
with vi (Lines 5–6). The algorithm then independently samples each seed from
{v1 . . . vT } with probability proportional to w (Line 7).
The challenge is to construct weights w that balance two opposing goals.
First, we would oftentimes like to disperse the seed nodes throughout the
network. For instance, if each community has equal size, we would like to
seed K different communities. Second, we would other times like to place

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press

 82 Influence Maximization with Unknown Network Structure

Algorithm 2 InitializeWeights
1: for i = 1...T do
2: Form Hi by discarding the first B nodes of Hi and keeping each
remaining node v j w.p. d(v1 j )

3: d̂ = R1 u∈Hi d(u)
4: Sˆ i = d̂−pb n
pw −pb
5: end for
6: w j = n .
Ŝ T
j

7: τ = max{Ŝ j | {i|Ŝ i ≥Ŝ j } wi ≥ K}
8: For any j with Ŝ j < τ, set w j = 0.
9: return w, H

Algorithm 3 RefineWeights
1: for i = 1...T do
2: vi = arg maxv∈Hi f (v, Hi )
Sˆi )2
3: wi = Tnβ(
f (si ,Hi )
4: end for
5: sort w in increasing order
6: for i = 1...T do
7: while EstVal(2wi , w−i ) > EstVal(w ) do
8: wi = 2wi
9: end while
10: wi = BinarySearch(wi , 2wi )
11: end for
12: return w

more seeds in large communities. For instance, if one community has 10,000
nodes and others have only 100 nodes each, we should seed the large com-
munity more. ARISEN navigates this trade-off with the following ingredients
(Algorithm 2). First, InitializeWeights uses the random walk around each
vi to estimate the size of the community that vi lies in (see next page “Initial
Weights”). From these estimates it constructs a w that, in expectation, seeds
the largest K communities once each. Second, RefineWeights (See “Refining
the Weights” later in the chapter) tests if using a w that puts more weight on
large communities would increase the expected influence. The main novelty

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press

 Bryan Wilder et al. 83

is to carry out these steps using purely local information, since we will not
generally be able to tell which of the vi are in the same community.
We first formalize the objective that ARISEN optimizes, which is a lower
bound on its true influence. Let f (X, Ci ) denote the influence of seed set X on
L
the subgraph Ci and g(X) = i=1 f (X, Ci ), i.e., the influence spread within each
community without considering between-community edges. ARISEN aims to
optimize E[g(X)]. Note that f (X, G) ≥ g(X) always holds. When pb is low and
little influence spreads between communities (which is the case that we study),
g is a good proxy for the true influence. We now explain ARISEN in detail, and
how it optimizes the surrogate objective g.

Initial weights
In the SBM, each community Ci has expected average degree di = |Ci |pw +
(n − |Ci |)pb . Solving for |Ci |, we can estimate the size of the community from
its average degree (Algorithm 2, Lines 3–4). InitializeWeights uses the
nodes sampled in the random walk (after a burn-in period B) to estimate di .
Since a random walk is biased toward high degree nodes, we use rejection
sampling (Line 2) to obtain an unbiased estimate. In order to choose seed
nodes using these estimates, a natural idea would be to choose the K samples
with the largest estimated size. However, this fails because large communities
are sampled more often and will be seeded many times, which is redundant.
For example, in Figure 5.1, placing all of the seeds in C1 would be suboptimal
compared to also seeding C2 . Complicating matters, the estimated sizes are
only approximate, which rules out many simple solutions. One solution is
to weight each sample inversely to its size (Line 6), and then sample seeds
with probability proportional to the weights. This evens out the sampling
bias toward large communities. Using a weighted sampling scheme gives us
a principled way to prioritize samples and facilitates later steps where we
tune the weights to improve performance. In the example in Figure 5.1, all
communities have total weight of 1 after this inverse weighting (Frame 2).
Next, the weights are truncated so that only the largest K communities
receive nonzero weight (Line 7). After this step, the largest K communities
have weight 1 and all smaller communities have weight 0 (at least approxi-
mately, due to sampling errors). For example, Frame 3 of Figure 5.1 shows that
only C1 , C2 and C3 have nonzero weight. We draw K seeds; in each draw, each
of the top K communities is seeded with probability approximately K1 . Thus,
the cumulative probability that each is seeded is nearly 1 − (1 − K1 )K ≥ 1 − 1/e.

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press

 84 Influence Maximization with Unknown Network Structure

Refining the weights

The initial weights suffice to obtain the approximation guarantee proved later
here and are the best possible for some networks. However, they are overly
pessimistic in other cases, such as when some communities are much larger
than others. In such cases, it would be better to focus more seed nodes on
large communities. We now outline RefineWeights, which tunes the weights
produced by InitializeWeights to account for such scenarios. In essence,
RefineWeights tries to exploit easier cases where some communities are
much larger than others by producing new weights w .
RefineWeights (Algorithm 3) starts in Line 2 by defining vi to be the
most influential node in the sampled subgraph Hi (instead of the random
starting node). Line 3 then sets wi according to vi ’s influence instead of the
estimated size of its community. Asymptotically, the two weighting schemes
are identical, but using influence spread instead of size is more accurate for
small networks. Lines 5–11 successively modify each element of w. Starting
with the weights corresponding to the largest communities, RefineWeights
asks whether g would be increased by doubling the wi under consideration
(Line 7). If yes, we set wi = 2wi and ask if it can be doubled again. If no,
RefineWeights performs a binary search between wi and 2wi to find the best
setting (Line 10). Then it moves on to the weight corresponding to the next
smallest community. In the example in Figure 5.1, Frame 4 shows that the
weights of samples from C1 and C2 have been increased. Each change is made
only if it improves g, so we have:
Proposition 5.0.1 Let w the output of InitializeWeights and w be the
   
output of RefineWeights. Then, EX∼w g(X) ≥ EX∼w g(X) .
The key difficulty is determining if each modification increases g. In the
EstVal procedure, we provide a way to estimate g using only local knowledge:
We give the main idea here. Take any seed set X. Note that the influence within
each Ci depends only on nodes in X ∩ Ci , which we write as XCi . So, g can
L
be rewritten as g(X) = i=1 E[ f (XCi , Ci )]. If we knew XCi , then we could
 
calculate E f (XCi , Ci ) by simulating draws from the SBM for the unobserved
portions of Ci . Concretely, let Hi be the subgraph observed in community Ci ,
with estimated size Sˆ i . We simulate the rest of Ci by adding Sˆ i − |Hi | new
nodes, with edges between them and Hi randomly generated from the SBM.
This is sufficient to choose the best seed within Hi , as in Line 2. For Line 7, we
need to estimate g. The obstacle is not knowing which of the v1 . . . vT lie in the
same community (since a node will contribute less influence if there is another
seed from the same community). However, we do know (approximately) how

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press

 Bryan Wilder et al. 85

Figure 5.2 Influence spread compared to OPT as K varies with q = 0.15.

Figure 5.3 Influence compared to OPT as q varies.

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press

 86 Influence Maximization with Unknown Network Structure

Figure 5.4 Query complexity as K varies.

many other times each community is sampled, and the (approximate) weight
that those samples will receive, so g can be estimated by averaging some
careful simulations. While runtime is not our focus, we note that via a standard
2
Hoeffding bound (see Kempe et al. [27]), O( n 2 log 1 ) simulations per EstVal
call guarantee multiplicative error  with high probability.

Experiments
We now present experiments on an array of data sets. First, homeless: two
networks (a and b) gathered from the social network of homeless youth in
Los Angeles and used to study HIV prevention with 150–200 nodes each.
Second, india: three networks of the household-level social contacts of villages
in rural India, gathered by Banerjee et al. [28] to study diffusion of information
about microfinance programs, with 250–350 nodes each. Third, netscience1 :
a collaboration network of network science researchers with 1,461 nodes.
Fourth, SBM: synthetic SBM graphs with 1,000 nodes each. SBM-equal has
10 communities of equal size (pw = 4n , pb = 0.2 n ) and SBM-unequal has 10
communities with size ranging from 3 n to 30 n (pw = 6n , pb = 0.2
1 1
n ).
We focus on networks with about 100–1,000 nodes because this is the size
of real-world social groups of interest to us. We present results for ARISEN
and three benchmarks. First, random, which simply selects K nodes uniformly
at random. Second, recommend, which for each of the K nodes first queries
a random node and then seeds their highest-degree friend. Third, snowball,
which starts from a random node and seeds that node’s highest-degree
neighbor. It then seeds the highest-degree neighbor of the first seed, and so
1
www-personal.umich.edu/mejn/netdata/

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press

 Bryan Wilder et al. 87

on. We compare to recommend and snowball because these are the most
common methods used in the field [6]. For each real network, pw and pb
are estimated from a different network in the same category (for netscience,
we use another collaboration network, astro-ph2 ). For the SBM data sets, we
use another network from the same distribution. We present a cross-section of
results across the data sets, but the general trends are the same for all networks.
Our first set of results measures the influence spread of each algorithm
against the optimal value. We approximate the optimal value using TIM [3], a
state-of-the-art influence maximization algorithm, run on the full network. As
in previous work [1], we focus on when K is a small fraction of n. Figure 5.3
shows results as q is varied with K = 0.01 · n. Each point gives the fraction of
OPT achieved for that setting of q. For example, the point at q = 0.2 indicates
that ARISEN obtained a value of 0.8 · OPT . Each point averages over 50 runs.
We see that ARISEN substantially outperforms all baselines, particularly when
q is low. All differences between algorithms are statistically significant (t-test,
p < 10−7 ). Previous work [1] has also observed that when q is very high,
influence maximization is sufficiently easy that nearly any algorithm performs
well. Thus, Figure 5.2 presents results where K is varied with q = 0.15 fixed
(since low q is when the problem is hard). We see that ARISEN uniformly
outperforms the baselines, particularly when K is small. As K becomes larger,
the baselines improve (again because the problem becomes easier). However,
they are still outperformed by ARISEN.
Figure 5.4 examines each algorithm’s query cost. The supplement lists R
and T for every data set and value of K; here we just focus on the total number
of queries. ARISEN uses more queries than any of the baselines. However, its
query cost is uniformly in the range 0.20 · n − 0.35 · n, a relatively small portion
of the network in absolute terms.

References
[1] Wei Chen, Chi Wang, and Yajun Wang. Scalable influence maximization for
prevalent viral marketing in large-scale social networks. In KDD, pages 1029–
1038. ACM, 2010.
[2] Kyomin Jung, Wooram Heo, and Wei Chen. Irie: Scalable and robust influence
maximization in social networks. In ICDM, pages 918–923. IEEE, 2012.
[3] Youze Tang, Xiaokui Xiao, and Yanchen Shi. Influence maximization: Near-
optimal time complexity meets practical efficiency. In KDD. ACM, 2014.

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press

 88 Influence Maximization with Unknown Network Structure

[4] Eric Rice, Eve Tulbert, Julie Cederbaum, Anamika Barman Adhikari, and
Norweeta G Milburn. Mobilizing homeless youth for HIV prevention. Health
Education Research, 27(2):226–236, 2012.
[5] Robert M Bond, Christopher J Fariss, Jason J Jones, Adam DI Kramer, Cameron
Marlow, Jaime E Settle, and James H Fowler. A 61-million-person experiment in
social influence and political mobilization. Nature, 489(7415):295–298, 2012.
[6] Thomas W Valente and Patchareeya Pumpuang. Identifying opinion leaders to
promote behavior change. Health Education & Behavior, 2007.
[7] Michelle Girvan and Mark EJ Newman. Community structure in social and
biological networks. PNAS, 99(12):7821–7826, 2002.
[8] Jure Leskovec, Kevin J Lang, Anirban Dasgupta, and Michael W Mahoney.
Community structure in large networks: Natural cluster sizes and the absence of
large well-defined clusters. Internet Mathematics, 6(1):29–123, 2009.
[9] Yi-Cheng Chen, Wen-Yuan Zhu, Wen-Chih Peng, Wang-Chien Lee, and Suh-Yin
Lee. CIM: community-based influence maximization in social networks. ACM
Transactions on Intelligent Systems and Technology (TIST), 5(2):25, 2014.
[10] Yu Wang, Gao Cong, Guojie Song, and Kunqing Xie. Community-based greedy
algorithm for mining top-k influential nodes in mobile social networks. In KDD,
pages 1039–1048. ACM, 2010.
[11] David Kempe, Jon Kleinberg, and Éva Tardos. Maximizing the spread of influence
through a social network. In KDD, pages 137–146. ACM, 2003.
[12] Tiziano Squartini, Francesco Picciolo, Franco Ruzzenenti, and Diego
Garlaschelli. Reciprocity of weighted networks. Scientific Reports, 2012.
[13] Amulya Yadav, Hau Chan, Albert Xin Jiang, Haifeng Xu, Eric Rice, and Milind
Tambe. Using social networks to aid homeless shelters: Dynamic influence
maximization under uncertainty. In AAMAS, pages 740–748, 2016.
[14] Michael Brautbar and Michael J Kearns. Local algorithms for finding interesting
individuals in large networks. In Innovations in Theoretical Computer Science,
pages 188–199, 2010.
[15] Stephen E Fienberg and Stanley S Wasserman. Categorical data analysis of single
sociometric relations. Sociological Methodology, 12:156–192, 1981.
[16] Emmanuel Abbe and Colin Sandon. Community detection in general stochastic
block models: Fundamental limits and efficient algorithms for recovery. In FOCS,
pages 670–688. IEEE, 2015.
[17] Florent Krzakala, Cristopher Moore, Elchanan Mossel, Joe Neeman, Allan Sly,
Lenka Zdeborová, and Pan Zhang. Spectral redemption in clustering sparse
networks. PNAS, 110(52):20935–20940, 2013.
[18] Elchanan Mossel, Joe Neeman, and Allan Sly. Reconstruction and estimation in
the planted partition model. Probability Theory and Related Fields, 162(3-4):
431–461, 2015.
[19] Bryan Wilder, Amulya Yadav, Nicole Immorlica, Eric Rice, and Milind Tambe.
Uncharted but not uninfluenced: Influence maximization with an uncertain net-
work. In AAMAS, pages 740–748, 2017.
[20] Shodai Mihara, Sho Tsugawa, and Hiroyuki Ohsaki. Influence maximization
problem for unknown social networks. In ASONAM, pages 1539–1546. ACM,
2015.

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press

 Bryan Wilder et al. 89

[21] Alexandra Carpentier and Michal Valko. Revealing graph bandits for maximizing
local influence. In International Conference on Artificial Intelligence and Statis-
tics, pages 10–18, 2016.
[22] Christian Borgs, Michael Brautbar, Jennifer Chayes, Sanjeev Khanna, and
Brendan Lucier. The power of local information in social networks. In WINE,
pages 406–419. Springer, 2012.
[23] Marco Bressan, Enoch Peserico, and Luca Pretto. The power of local information
in pagerank. In WWW, pages 179–180. ACM, 2013.
[24] Christian Borgs, Michael Brautbar, Jennifer Chayes, and Shang-Hua Teng.
Multiscale matrix sampling and sublinear-time pagerank computation. Internet
Mathematics, 10(1–2):20–48, 2014.
[25] Masahiro Kimura, Kazumi Saito, Ryohei Nakano, and Hiroshi Motoda. Finding
influential nodes in a social network from information diffusion data. In Social
Computing and Behavioral Modeling, pages 1–8. Springer, 2009.
[26] Noga Alon, Michal Feldman, Omer Lev, and Moshe Tennenholtz. How robust is
the wisdom of the crowds? In IJCAI, pages 2055–2061, 2015.
[27] David Kempe, Jon M Kleinberg, and Éva Tardos. Maximizing the spread of
influence through a social network. Theory of Computing, 11(4):105–147, 2015.
[28] Abhijit Banerjee, Arun G Chandrasekhar, Esther Duflo, and Matthew O Jackson.
Gossip: Identifying central individuals in a social network. Technical report,
National Bureau of Economic Research, 2014.

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press

https://doi.org/10.1017/9781108669016.006 Published online by Cambridge University Press
 P AR T II

Published online by Cambridge University Press

Published online by Cambridge University Press
 6
Maximizing the Spread of Sexual Health
Information in a Multimodal Communication
Network of Young Black Women
Elizabeth Bondi, Jaih Craddock, Rebecca Funke, Chloe LeGendre,
and Vivek Tiwari

Introduction
Young Black women (YBW) between the ages of 18 and 25 have higher rates
of HIV than any other age or race and ethnicity group of women (Center for
Disease Control and Prevention (CDC), 2016). Social networks of YBW might
be considered both protective and risky from a public health perspective, as
information about sexual behaviors and sexual health may spread within the
communication network to the health benefit or detriment of the network’s
members. However, there is a dearth of literature on the social networks of
YBW, particularly regarding how such social networks can be best leveraged
to aid in HIV prevention within such networks. With the current HIV/AIDS
priority of decreasing incidence of HIV among high-risk groups, examining
ways to leverage such networks is essential.
Research has demonstrated that reduction in HIV rates for various pop-
ulations can be achieved using methodologies from social network theory.
Specifically, peer-led interventions have been used to reduce health risk
behaviors in high-risk populations (Convey, Dickson-Gomez, Weeks, & Li,
2010; Davey-Rothwell, Tobin,Yang, Sun, & Latkin, 2011; Jennings, Howard,
& Perotte, 2014; Mahat, Scoloveno, De Leon, & Frenkel, 2008; McKirnan,
Tolou-Shams, & Courtenay-Quirk, 2010; O’Grady, Wilson, & Harman, 2009).
Interventions that use peer leaders in network-oriented HIV intervention
programs to endorse changes in risk behaviors accelerate behavioral changes
by capitalizing on social influence processes (Kelly et al., 1991; Latkin et al.,
1996). Diffusion of Innovation is a theoretical framework that posits that
people are more likely to adopt new ideas or behaviors if similar people in
their network have either already adopted them or have favorable evaluations
of these ideas or behaviors (Rogers, 2010). The theory of Diffusion of

93

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 94 Spread of Sexual Health Information in Black Women

Innovation has been used in social network–based interventions to guide the

understanding of how information is spread throughout networks.
Ultimately, our goal is to spread sexual health information throughout
networks of YBW. Ideally, the information would reach as many individuals
in the networks as possible, increasing the likelihood of a reduction of
HIV rates within this higher-risk population. To achieve this goal, as with
typical peer-led interventions, key members of a social network would be
trained to disseminate sexual health information throughout their networks.
More specifically, in our approach, the selected individuals would be taught
how to effectively communicate about sexual health using one preferred
communication platform (for instance, via text messaging).
In this work, we propose that if information about a given social network
is known – including the connectivity between individuals, the modes of
communication typically employed across each connection, and whether or not
the conversations may include topics related to sex or sexual health – we can
maximize the spread of information in the network when constrained by lim-
ited resources for intervention and training. For example, if we observe some
level of communication about sexual health via texting, we might thus consider
that this mode of communication might be effective to use for disseminating
information pertinent to HIV prevention throughout the network of YBW.
Therefore, an effective intervention for information diffusion could perhaps
involve training some small subset of the network on how to effectively spread
information about sexual health via text messaging specifically.
The communication networks formed by YBW are particularly complex,
as such communication is known to be multimodal – composed of a variety
of communication methods including in-person conversations, phone calls,
and many social media– and mobile phone–based platforms. Communication
between YBW and their social network members (SNMs) via social media
is becoming normalized due to a rapidly changing social media landscape
(Noar & Willoughby, 2012). Black youth and young adults are using social
media websites and apps at higher rates than any other American group
(Lenhart, 2015; Smith, 2014). Considering that social media is increasingly
becoming the mode of communication for many Black young adults, it is
essential to understand through what mode (i.e., social media, text message,
phone, or in-person) YBW are speaking to their SNMs regarding sexual
health, and how these communication networks can be leveraged by social
workers to spread information that will benefit the health of the population.
Other researchers have examined the possibility of using social network–
based peer-led HIV intervention via social media (Young et al., 2014).
However, to our knowledge, no HIV intervention studies have considered

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 Elizabeth Bondi et al. 95

social networks of YBW, particularly examining the modes of communication

used by YBW.
Based on quantitative survey information, we identify four ways that
YBW typically communicate with each other about sexual health: in-person
(talk), texting (txt), calling on the phone (call), or over social media (soc),
including a variety of social media platforms. We present a hypothetical
relationship example to demonstrate the potential complexity of a multimodal
communication network of YBW, specifically as it relates to sexual health.
First, each individual in a social network might currently communicate about
sexual health–related topics only with a fraction of the people in her network.
Each individual also might only communicate with these people using certain
platforms. For instance, take four YBW: individuals denoted as A, B, C, and
D. Perhaps A is friends with B, C, and D. Individual A might feel comfortable
communicating with B about sexual health via in-person conversation, text
message, and social media, but might only talk to C about sexual health
via social media. Perhaps, based on the nature of the relationship between
A and C and their past interactions, individual A might feel uncomfortable
discussing potentially personal or sensitive topics in face-to-face conversation
with individual C. Additionally, perhaps A and D do not communicate about
sexual health on any platform. The quantitative survey we describe and
utilize in this work was designed to elucidate such complexities, which are
inherent in the multimodal communication networks of YBW, but even further
complicated by the inclusion of potentially sensitive or personal subject matter.
By analyzing the spread of sexual health information over a real network of
73 Los Angeles–based YBW for each different communication modality, we
propose to select the best mode of communication to utilize for peer-led inter-
ventions for YBW (i.e., choosing among in-person, text message, phone call,
or social media–based interventions). We also propose that selecting which
individuals should be trained as intervention peer leaders varies depending on
the communication mode selected for the intervention, which is fundamental to
understanding how these network structures should be best used to effectively
spread sexual health and HIV prevention information.
In summary, in order to achieve the goal of spreading information to the
greatest number of individuals within a given communication network, we
propose that one must answer two questions: (1) what mode of communication
will be best for a given network, and (2) given this best communication
mode, which individuals should be trained as a part of the intervention? To
answer these questions, a social worker could employ the approach of Valente
and Pumpuang (2007), where individuals are selected based on having high
“degree centrality,” or the most ties with other individuals in the network.

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 96 Spread of Sexual Health Information in Black Women

This approach, however, does not consider that these “popular” individuals
may share many connections, and as a result, the network model is not
accurate (Yadav et al., 2015; Cohen, Delling, Pajor, & Werneck, 2014). Instead,
we propose to analyze one real-world multimodal communication network
using tools and techniques developed by computer science researchers. More
specifically, we model this information diffusion problem as one of “maximum
influence.”
One frequent goal of computer scientists is to determine how to most
“optimally” perform some given task. For instance, there are numerous
situations where a community is faced with a limited budget, and its members
or leadership must decide how to best allocate resources. A computer scientist
would try to: (1) find out if such an optimal allocation exists; (2) find this
optimal allocation; or, alternatively, if this task is impossible or difficult (i.e.,
the problem-solving approach is too slow to be useful for impactful decision-
making), (3) find a “good” way of approximating the optimal solution. This
approach can be applied to the HIV prevention domain, where interventionists
will have a limited budget to spend on training peer leaders about sexual
health but ultimately want the majority of a network to adopt a new behavior
to prevent HIV. Thus, interventionists will try to select for training the “k”
best people in the network that can spread information to the most network
members, where k can be any number of individuals (however many the
training budget will allow).
Because maximizing the information spread throughout a network involves
“optimizing” for the dissemination of knowledge, this is a tantalizing question
for computer scientists. In fact, much research has focused on modeling and
optimizing the spread of information throughout social networks. In computer
science, the problem of finding the k most influential people in a network
is referred to as the “maximum influence problem” (Kempe, Kleinberg, &
Tardos, 2003). Note that finding the “k most influential people” implies that
there exists no other subset of k people in the network such that if this
other group receives the same intervention training, that information will be
spread to more network members as compared with the originally selected k
individuals. Despite the difficulty of the problem, computer scientists studying
network effects have developed computational approaches for solving the
maximum influence problem, which we use in this work to propose best
practices for peer leader–based network interventions in the presence of
multimodal communication.
Using techniques from computer science to optimize the spread of infor-
mation throughout a network has had proven success in HIV prevention. For
example, a heuristic algorithm, HEALER, selects k homeless youth to be
trained in HIV prevention. In a real-world deployment of HEALER, they found

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 Elizabeth Bondi et al. 97

that using HEALER to select the peer leaders allowed for about 70% of the
non–peer leader homeless youth to receive the message about HIV prevention,
whereas the method typically used by officials at shelters allowed for 27%.
Not only did more individuals learn more about HIV prevention,? nearly 37%
of the non–peer leader participants got tested for HIV, whereas no participants
got tested when using the typical method (Yadav et al., 2017).
In this chapter, we first describe the survey methodology used to collect
the real-world YBW network data set. Then, we describe the mathematical
models of the communication network and its information diffusion process.
As the maximum influence problem is difficult, we describe some potential
simplifying assumptions that we can make to solve the problem efficiently,
and we evaluate the efficacy of multimodal and single communication mode
intervention trainings under these assumptions. From the results of our simula-
tion experiments using the real-world survey data, we propose a network-based
intervention approach where the best communication modality depends on the
underlying network structure, and peer leaders are selected depending on the
training modality.

Methods
YBW Quantitative Network Survey – Data Collection
As described in the preceding section, we evaluate the spread of information
using a real-world quantitative survey data set designed to probe the members
of a social network for their communication habits, both generally and in the
domain of sexuality and sexual health–related behaviors. The social network
structure was developed by surveying 78 YBW aged 18 to 25 and examining
sexual health communication at the dyadic level between YBW, though 5
YBW reported no connections with other participants in the study and were
not included in analysis. Recruitment was performed using respondent-driven
sampling performed from May 2016 to August 2016. Five seeds (initial
participants) were recruited from two beauty salons in Los Angeles County.
Qualifications for seed participation were age range (18–25 years, inclusive),
self-identification as a Black or African American woman, and residence in the
Los Angeles County area. During the survey, participants were incentivized
to invite at least three other YBW (i.e., friends, family, or acquaintances) via
text message to participate in the study, with referred individuals considered
eligible based on identical screening requirements. Eligible study participants
each answered the same survey and received $25 for completing the study, and
an additional $10 for referring at least two other YBW. Informed consent was
obtained from all study participants.

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 98 Spread of Sexual Health Information in Black Women

Each woman completed both a survey interview and a social network inter-
view. Survey interviews collected demographics, sexual risk behaviors, sexual
relationship information, and social media use. Social network interviews
asked women to list SNMs to produce a standard egocentric network using an
iPad application (Rice, 2010). The social network interview progressed in two
stages. First, names and demographic information of all of the participants’
network members were collected. Second, questions were asked regarding
the types of relationship and communication that participants have with
each SNM. The iPad application listed each provided name as an option
for each subsequent question, similar to categorical answers of a multiple-
choice question. Measures for this study were mode of communication and sex
communication. Mode of communication was assessed using two questions:
“Who do you talk to [on the phone, via text message, or in person]?” and
“Who do you talk with on [Facebook, Instagram, Snapchat, or Twitter]?” These
outcomes were binary with 1 equaling a selection and 0 equaling no selection.
Sex communication was assessed using the question, “Who do you talk with
about sex?” Persons identified as someone the YBW spoke to about sex were
coded as 1, and persons not identified were coded as 0.

Mathematically Modeling Information Diffusion

in the YBW Network
To examine the spread of information in the real-world network of YBW,
we define the abstract model of this network, which enables the application
of algorithms used to solve maximum influence problems. The input for a
maximum influence problem is a directed graph, G(N,E,C). The set of nodes N
is all of the people in the network, where each person is a node. The collection
of directed edges, E, includes all edges in the network, where each directed
edge is a one-way, person-to-person communication connection. For instance,
if nodes u and v communicate, then there may exist at most two directed
edges connecting them, with these edges denoted eu,v and ev,u . In the first
edge, labeled eu,v , node u communicates to node v, and in the second edge,
labeled ev,u , node v communicates to node u. Each directed edge will have
four associated binary values (txt, talk, call, soc) that denote if the first node
communicates to the second node about sexual health on each platform. The
use of directed edges for the graph G models the directionality of information
exchange relating to sexuality and sexual health, which we observed in the
real-world network surveyed. As an example, suppose node u provides sexual
health information to node v via social media, but node v does not feel

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 Elizabeth Bondi et al. 99

comfortable speaking to node u about sexual health using any platform. Stated
differently: node v may “receive” information about sexual health from node
u, but may or may not “give” any such information to node u. In this example,
the binary values associated to the edge eu,v would be (txt = 0, talk = 0, call = 0,
soc = 1), and the binary values associated to the ev,u would be (txt = 0, talk = 0,
call = 0, soc = 0). In terms of notation, the set of four binary values for each
edge is denoted as C.
With the network structure defined, the first natural follow-up question
is: How is the extent of the spread of information in a network measured?
Traditionally, interventionists must select the k people to be trained before
implementing the training program, thus the information diffusion process
must be accurately modeled and simulated prior to training to provide an
expected value of the number of people that will be influenced by the k
people selected. Within the domain of maximum influence, there are various
developed models that can be used to model the diffusion of information across
a network. The two most popular models are the linear threshold model and
the independent cascade model (Kempe et al. 2003). In both models, the initial
k people are marked as “active”, meaning they have learned or already know
the information to be spread throughout the network, and they are willing and
able to spread information to their connections. The goal is then to keep track
of how many other nodes become active.
In the linear threshold model, each node has some threshold for the number
of their immediate network connections that must be active in order for them
to become active. A node in a linear threshold model becomes active when its
node-specific threshold is surpassed. For instance, if a certain number of your
friends buy the newest FitBit, then you might also buy the newest FitBit. This
threshold is personal, meaning that it varies from individual to individual, and
it can potentially be further complicated by unequal “weights” or importance
of specific connections. In other words, some friends might influence your
Fitbit buying decision more heavily than others, and each friend’s influence
may vary depending on the subject matter of the particular decision.
In this work, we instead use the independent cascade model. With this
model, when a node learns some novel information and becomes active, the
node attempts to influence each of its neighbors with some probability of
success. Intuitively, the independent cascade model simulates the uncertainty
of information spread that we feel is inherent in our problem due to both the
indirect nature of social media exchanges and the highly personal/sensitive
domain of sexual health. For instance, suppose YBW A becomes active,
meaning that she has learned the new sexual health information helpful for
HIV prevention. YBW A can then try to activate all her connections, but this

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 100 Spread of Sexual Health Information in Black Women

activation may not always be effective. For instance, suppose interventionists

choose to do a social media–based training and intervention. After learning
about this new sexual health information (or, simply stated, becoming active),
YBW A decides to try to spread this message on social media. Perhaps
YBW A creates a post on her Facebook wall about this information. Each of
her Facebook connections might see this post, read it, and become activated
themselves – knowledgeable and also willing to share the message. But
each friend also might not see, read, or share the information, and there is
accordingly some probability associated with both outcomes (each connection
of YBW A becomes activated or does not). Using the independent cascade
model of information diffusion, we can employ algorithms developed for
solving the maximum influence problem to design optimal interventions for
multimodal communication networks of YBW.
We also must consider what simplifying assumptions we can make to
reduce the computational complexity of the problem, without resulting in
an oversimplification that would render the model unrealistic and useless
for social workers and interventionists. First, we assume that the network of
YBW is stable, meaning that new members are not joining or leaving the
network and that connections (edges) are neither created nor destroyed over
time. In addition, we assume that connections reported from the survey are
known to exist with absolute certainty. Both of these assumptions describe
“properties” of the network rather than its structure. These simplifying network
property assumptions are most likely valid on the whole for our specific
population of YBW, but they may not be appropriate for other network
populations. For instance, in the work of Yadav et al. (2016), networks of
homeless youth were known to be unstable owing to the transient nature
of such friendships and the geographic mobility of the members of this
population.
We also consider other simplifying assumptions about how information
spreads across our specific network of surveyed YBW. These assumptions
differ from the network stability and edge certainty assumptions mentioned
previously because these new assumptions result in the selection of the edge
weights to use for the network. In other words, these assumptions may change
the structure of the network without changing its overall properties. We call
these network structure assumptions. There are three such sets of assumptions
we consider in different experiments. In each experiment, we assume one set
of assumptions and then use an algorithm to solve the associated, simplified
maximum influence problem. We compare the results obtained when using
each set of assumptions to examine how information is spread throughout the
network for each set of associated edge weights. Below are the three sets of

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 Elizabeth Bondi et al. 101

network structure assumptions, in order of increasingly difficult (as in difficult

to determine people to select as peer leader) scenarios:

1. Any Platform: We first assume that the mode of communication used to

disseminate information is irrelevant. This experiment models the
real-world intervention scenario where the k leaders are trained to spread
sexual health information via all communication platforms (or that
training in one platform still spreads information via all communication
platforms). We assume that information spreads over the various platforms
with equal efficacy. In order to test this assumption, we solve the
maximum influence problem using all of the friendship/communication
edges, regardless of their binary (txt, talk, call, soc) values. Since we
include all possible edges in the network in this experiment, this test yields
the “theoretically maximum” amount of activated nodes that we could
hope to achieve with an intervention with k selected peer leaders, but we
note that this kind of multimodal communication training is not
necessarily realistic within the budgetary and time constraints of social
work interventions. This test, therefore, is most useful as a comparison.
That is, as communication modalities are eliminated from intervention
training efforts, we should compare these future results against the
network activation achieved using all edges or communication modalities.
2. One Platform Only: We next assume that we are restricted to training
selected peer leaders for the intervention on how to spread information
about sexual health using exactly one communication platform due to
budgetary and time constraints. We assume that information will only be
subsequently spread over this single platform, in order to see how we
perform should we face a more difficult scenario. Therefore, if an
individual in the network is not using that platform, that individual cannot
become activated. For instance, if YBW A does not own a phone, we
cannot influence her to buy one. We test this scenario by considering the
subnetwork for communication type c where we only include the directed
edges where the weight corresponding to c is one. We repeat this
evaluation for all four communication types in order to select among the
most effective communication training modality.
3. Sex Restricted, One Platform: Again, we assume that whatever
communication method is taught during the intervention is the only
communication method that will be used to spread the message. We also
assume again that we cannot influence two people to start communicating
over some platform if they were not at the time of the network survey. The
additional, new assumption we make is if two people are not already

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 102 Spread of Sexual Health Information in Black Women

talking about sex on this specific platform, then we cannot influence them
to start talking about sex on that platform. For example, perhaps two
people are very uncomfortable discussing sex with one another and are
unwilling to change their behavior no matter the intervention. These three
assumptions together are considered to model a worst-case scenario. We
test this by considering the subgraph for each communication class c
where we include only the directed edges where the weight for c on an
edge is 1 and for an edge denoted eu,v , where node u is already talking to
node v about sex. We note that edge directionality is particularly important
here.

Algorithms for Solving the Maximum Influence Problem

With our simplified network structures to evaluate, we now describe algorithms
that can be applied to solve the maximum influence problem. As stated earlier,
the goal of the maximum influence problem is to find the k most influential
people in the network. Using the independent cascade model, we can measure
the expected influence of each subset of k people. One strategy for solving
the maximum influence problem is that, for each subset of k people, we
could use the independent cascade model to determine the expected activated
number of people, and then choose the subset that allowed for the most
activation. Computer scientists generally refer to this approach as the “brute
force” algorithm because every possible solution is considered. Brute force
algorithms are rarely used in practice because most real-world problems have a
very large number of possible solutions, and it would take too long to evaluate
every possible case. For the maximum influence problem, if we were to use
this brute force approach, we would have to consider “n choose k” possible
solutions where n is the number of people in the network – that is, choosing k
people from the set of size n. So, for instance, if the network had 100 people
and 20 people were to be selected for the peer-led intervention, then there are
535,983,370,403,809,682,970 (or 5.35 × 1020 ) different subsets of 20 that the
brute force algorithm would need to evaluate. For comparison, this number of
cases is larger than the estimated number of grains of sand on Earth (which is
estimated at 7.5 × 1018 ). Thus, it is clear that the brute force algorithm is not a
practical or possible solution approach for this problem, even when the size of
the network is relatively small (about 100 people).
In fact, this “combinatorial” number of solutions is a frequent problem in
computer science. It is very often that brute force algorithms cannot evaluate
all possible solutions to produce a result during the lifetime of the computer

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 Elizabeth Bondi et al. 103

programmer. Maximum influence problems belong to a set of such problems,

labeled as NP-hard, which is a computer science term that classifies a problem
as one where the optimal solution cannot be determined in a reasonable amount
of time. When problems are NP-hard, computer scientists design algorithms
that can efficiently find an approximation to the optimal solution. Ideally, the
algorithm designers are also able to provide some approximation guarantees,
meaning that they can mathematically prove how close their solution is to the
optimal solution, within some clearly defined margin.
To solve the maximum influence problem, we use two approximation
algorithms for selecting k nodes for the intervention. In other words, these
algorithms are not guaranteed to select the exactly optimal k nodes to maximize
information diffusion; however, they will finish in a reasonable amount of time
and will yield provably good results. In the maximum influence literature, these
algorithms are referred to as the Greedy and HEALER algorithms. The Greedy
algorithm provides a (1-1/e) approximation guarantee, meaning the solution
generated by the Greedy algorithm will always be approximately 1/3 away from
the true optimal solution. Unfortunately, the HEALER algorithm has no proven
approximation guarantees.
The Greedy algorithm’s approximation guarantees come from the fact
that Greedy’s influence function is adaptive submodular, whereas HEALER’s
influence function is not. Submodularity is intuitively the idea of diminishing
returns from economics (for example, if trying to allocate grocery stores in
a new city, if two are next to one another, people will chose one or the
other; you may not get any benefit from the extra grocery store), and adaptive
submodularity takes into account observations from the past. Previous work
has proven that if an optimization problem is adaptive submodular, then an
adaptive Greedy algorithm will provide a near optimal solution (Golovin &
Krause, 2011). Though the HEALER algorithm has no proven approximation
guarantees, it was shown to outperform the Greedy algorithm in the specific
context outlined in Yadav et al. (2016). We utilize both algorithms in our
work to see if HEALER will outperform Greedy experimentally in this
context as well.
The Greedy algorithm is similar to the brute force approach but uses
heuristics to choose nodes for testing (rather than choosing all possible nodes).
One heuristic that can be used to choose nodes is “distance centrality,” which
assumes nodes that are “closest” to one another will most influence each other
(Kempe et al., 2003). Another assumption of the Greedy algorithm is that there
is no uncertainty in the edge existence in the network. We used a variation on
this algorithm for testing, which allows us to account for multiple interventions
over time, similarly to the HEALER algorithm.

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 104 Spread of Sexual Health Information in Black Women

The HEALER algorithm is quite different from the Greedy and brute force
algorithms in that it deals with uncertainty in the social network (Yadav
et al., 2016). To do this, HEALER uses a type of model called a Partially
Observable Markov Decision Process (POMDP). This model assumes that
we cannot tell exactly what is happening in the social network, but we
can observe the connections reported in surveys, for example, which makes
the network “partially observable.” In our work, we assume that we know
that edges exist in the surveyed network with absolute certainty, where
edge existence probabilities are either 1 or 0. HEALER also allows us to
consider the influence of multiple interventions over time, allowing for the
restructuring of the network at multiple survey time points. For example,
instead of choosing 12 people for a single intervention and modeling how
they might influence their peers at one time, we can choose four people for
an intervention, wait a month, and then choose another four people for an
intervention, where we may also learn new information about the network.
While we do not have a data set that allows us to learn new information
about the network after each intervention, this could be added in further field
studies.

Results and Discussion

To determine the optimal communication method for disseminating sexual
health information throughout a network of YBW, both the HEALER and
Greedy algorithms were applied to solving the maximum influence problem
for the surveyed real-world communication network of YBW, with different
edges included based on each of the aforementioned networks’ structure
assumptions. Each of these assumptions “removes” certain nodes from the
network if they do not participate in that method of communication or the
network of YBW, so the total number of nodes remaining after such node
removal is also presented. Figures 6.1–6.5 show the YBW network diagrams
generated using the survey data, where Figure 6.1 represents the full network
and Figures 6.2–6.5 represent the network when considering only single
communication modes.
For our experiments, we assumed there would be three rounds of interven-
tions, each at a different time point, with four YBW chosen as peer leaders at
each stage. For each experiment we report the number of people indirectly
influenced – the number of people who received sexual health information
without being directly influenced during an intervention. With 3 interventions

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 Elizabeth Bondi et al. 105

142
166 170

153 160

165 180 171

145
168 151 156
157
155 154 133 144
158
134
130 110
173
106 108
169 112

150 159 111 116 102

104 107
121
146 129
137 131
132

139 122
140 114
138
135 113 141 127
120
109
148 124 126
149 105
123 147 105
152 167
164 143
118 101

117

163
162

175 176
128
136 115
119

Figure 6.1 All connections in the network of 78 YBW, with 5 surveyed nodes
excluded who have no direct ties to other study participants. Connections are
included regardless of communication mode or discussion topic. As a result, there
are 106 total connections (edges) between participants (nodes).

and 4 peer leaders in each intervention, 12 individuals are directly influenced

by the peer leader training and therefore excluded from our reported results.
We also excluded from our analysis 5 singleton nodes without any self-reported
intra-network connections for any communication mode, yielding an analyzed
network size 73 instead of 78.
We performed 50 trials for each algorithm using the same parameters and
network structure assumptions. We generate results from multiple trials for
each algorithm because influence spread is modeled as a random process,
wherein a node becomes activated based on the result of a coin flip. In our
experiments, we set the probability of activating a node (equivalent to the
probability of getting a “heads” result in the coin flip, for example) to be
0.6, as in experiments with HEALER (Yadav et al., 2016). Since influence
spread is modeled probabilistically, the experimental results varied from one
individual trial to the next, even when using the same algorithm with an
identical network structure. Intuitively, the number of indirectly influenced
people may change from trial to trial because of the coin flip. Perhaps a less

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 106 Spread of Sexual Health Information in Black Women

142
166 170

153 160

165 180 171

145
168 156
151 157
154 133 144
155 158
134
130 110
173
106 108
169 112

150 159 116 102

104 107 111
121
146 129
137 131
132

139 122
140 114
138
113 141 127
135 120
109
148 124 126
149 105
123 105
152 167 147
164 143
118 101

117

163
162

175 176
128 136 115
119

Figure 6.2 In-person communication connections in the network, regardless of

discussion topic. Note that the large group on the right of the diagram of Figure
6.1 is fractured into three groups and one singleton node when considering
only in-person communication. Because we are only considering in-person
communication, there are now 83 edges. If we further constrain the network to
include connections only if the two participants talk about sex in-person, there are
73 edges.

intuitive result, however, is that the individual nodes selected as peer leaders
may also change, since these selections themselves also depend on the coin flip
results.
First, in Table 6.1, we report the number of indirectly influenced people
for each social network for the Greedy algorithm, considering the average and
standard deviation of the 50 trials for each experiment. Next, in Table 6.2 we
report the average results for the HEALER algorithm. The total number of
nodes in the network is denoted by N in the tables.
As expected, the full network including all communication methods leads
to the largest spread of influence for both the Greedy and HEALER algorithms,
since this network includes the most nodes and edges. This result shows that,
ideally, interventions should train peer leaders to spread information using
multiple modes of communication. When this is not possible due to resource
or time constraints, our simulation results using both algorithms show that
texting-based interventions would yield the most information diffusion for our

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 Elizabeth Bondi et al. 107

142
166 170

153 160

165 180 171

145
168 151 156
157
154 133 144
155 158
134
130 110
173
106 108
169 112

159 116 102

150 104 107 111
121
146 129
137 131
132

139 122
140 114
138 141
135 113 127
120
109
148 124 126
149 105
123
167 147 105
152
164 143
118 101

117

163
162

175 176
128 136 115
119

Figure 6.3 Phone communication connections, regardless of discussion topic.

Note that the large group on the right of the diagram of Figure 6.1 is fractured into
five groups and three single nodes when considering only phone communication.
There are 74 edges when only phone communications are considered, with only
68 edges when considering only phone communications about sex.

surveyed network of YBW. In particular, there is only a very slight difference in

the spread of information when using only texting as compared with using all
modes of communication. This is a promising result in favor of text message–
based interventions for networks of YBW, since simpler single communication
mode interventions would likely require less time and resources from social
workers and interventionists.
While our results suggest that text messaging is the best single mode of
communication for information diffusion, we note that the study subjects were
referred to participate via text messages from their peers. As a result of this
recruitment methodology, we advise the reader to consider that this particular
network may be biased toward text messaging as an overall preferred mode of
communication.
Our results also show that peer leader training designed to encourage the
discussion of sexual health among connections that do not typically discuss sex
would also be modestly beneficial, although texting still results in the largest
spread of information among networks where edges are removed if they do not
already discuss sexual health.

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 108 Spread of Sexual Health Information in Black Women

142
166 170

153 160

165 180 171

145

151 156
168 157
155 154 133 144
158
134
130 110
173 106 108
169 112

159 116 102

150 104 107 111
121
146 129
137 131
132

139 122
140 114
138 141
135 113 127
120
109
148 124 126
149 105
123
167 147 105
152
164 143
118 101

117

163
162

175 176
128 136 115
119

Figure 6.4 Social media communication connections, regardless of discussion

topic or social media platform. Note that the large group on the right of the
diagram of Figure 6.1 is fractured into seven groups and six single nodes
when considering only social media communication. There are 67 edges when
only social media communications are considered, with only 59 edges when
considering only social media communications about sex.

Additionally, the Greedy algorithm performs better than the HEALER

algorithm in our experiments. This is likely because HEALER is best used
when new information is gathered between intervention rounds, and when
there is uncertainty in the network. This was not the case in our survey data
set, where data was collected only at one point in time, and connections were
certain. However, we still include the results for HEALER in our discussion,
because a future data set could be collected for a population that would
utilize HEALER’s full potential for approximating maximum influence under
a dynamic and potentially uncertain network structure. For future studies
seeking to employ HEALER, Greedy should still be used for comparison so
that the best results can be achieved.
In Figure 6.6, we show the most commonly selected peer leader nodes
from among the 50 trials for the Greedy algorithm. We show the results
for texting-based interventions in Figure 6.6a, and the result for in-person
conversation-based interventions in Figure 6.6b. Comparing the nodes selected
by the Greedy algorithm in Figure 6.6 with those selected using the HEALER
algorithm in Figure 6.7, we see that the Greedy algorithm more frequently

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 Elizabeth Bondi et al. 109

Table 6.1. Greedy algorithm for maximum influence: Average number of

indirectly influenced people

Only Those Who

All Connections Discuss Sex Already

All Communication Methods 33.412 ± 0.420 (N = 73)

All Social Media 21.690 ± 0.277 (N = 64) 20.007 ± 0.259 (N = 60)
In Person 25.579 ± 0.340 (N = 72) 24.177 ± 0.381 (N = 65)
Phone 25.852 ± 0.276 (N = 67) 24.624 ± 0.171 (N = 67)
Texting 29.888 ± 0.512 (N = 72) 28.697 ± 0.507 (N = 64)

Note: n = 50 trials for each condition using the Greedy algorithm to choose peer
leaders for interventions.

142
166 170

153 160

165 180 171

145
156
168 151 157
155 154 133 144
158
134
130 110
173
106 108
169 112

159 116 102

150 104 107 111
121
146 129
137
131
132

139 122
140 114
138
113 141 127
135 120
109
148 124 126
149 105
123
152 147 105
143 167
164 118 101

117

163
162

175 176
128 136 115
119

Figure 6.5 Text messaging communication connections, regardless of discussion

topic. Note that the large group on the right of the diagram of Figure 6.1 is
fractured into just two groups and one single node when considering only social
media communication. Of all the communication methodologies, text messaging
leaves the greatest number of links of the diagram of Figure 6.1 intact. There are
95 edges when only text message communications are considered, with 81 edges
when considering only text message communications about sex.

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 110 Spread of Sexual Health Information in Black Women

Table 6.2. HEALER algorithm for maximum influence: Average number of

indirectly influenced people

Only Those Who

All Connections Discuss Sex Already

All Communication Methods 25.250 ± 1.739 (N = 73)

All Social Media 17.105 ± 1.560 (N = 64) 15.024 ± 1.272 (N = 60)
In Person 19.910 ± 1.493 (N = 72) 19.182 ± 1.361 (N = 65)
Phone 19.270 ± 1.327 (N = 67) 19.248 ± 1.227 (N = 67)
Texting 23.057 ± 1.872 (N = 72) 22.047 ± 1.679 (N = 64)

Notes: n = 50 trials for each condition using the HEALER algorithm to choose peer
leaders for interventions.

(a) 166
142
170

153 160

165
180 171 145

168 151
156
154 144
157 133
155 158
134 130
110
173
106 108
169 112

159 111 116 102

150 104 107
121
146 129
137 131
132
139 122

140 114
138 113 141
120 127
135
109
148 124 126
149 105
123 147 105
167
152 101
164 143
118

117

163
162

175 176

128 136 115

119

Figure 6.6 Dark gray nodes identify the 12 most commonly selected peer
leader nodes among the 50 trials of the Greedy algorithm for (a) texting based
interventions and (b) in-person interventions. Here we assume that nodes can be
influenced regardless of whether or not sex is already discussed. Note that the 12
nodes do not appear as the selected peer leaders during one trial, but rather are the
most commonly selected nodes across trials.

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 Elizabeth Bondi et al. 111

142
(b) 166 170

153 160

165
180 171 145

168 151
156
154 144
157 133
155 158
134 130
110
173
106 108
169 112

159 111 116 102

150 104 107
121
146 129
137 131
132
122

140 114
138 113 141
120 127
135
109
148 124 126
149 105
123
147 105
167
152 101
164 143
118

117

163
162

175 176

128 136 115

119

Figure 6.6 (cont.)

selects individuals from many different cliques. For example, Greedy selects
nodes 128 and 119 (Figure 6.6a), while HEALER does not (Figure 6.7a),
which might explain its greater influence spread as compared with HEALER
for our data set.
In Figure 6.7, we show the most commonly selected peer leader nodes
from among the 50 trials for the HEALER algorithm. We show the results
for texting-based interventions in Figure 6.7a, and the results for in-person
conversation-based interventions in Figure 6.7b.
The most striking difference between Figures 6.7a and 6.7b is that some
nodes are selected from among the medium-sized groups to the left-hand
side of the network diagram for the texting intervention, most likely because
the large clique on the right-hand side of the diagram can be influenced
by fewer selected peer leaders owing to the larger number of edges in the
texting communication network. In contrast, for the in-person communication
network, all but one of the most frequently selected nodes originate from the
large right-hand clique. This example demonstrates how the communication
modality may directly impact the selection of peer leaders in a network for
intervention training.
In Figure 6.8, we show the result of one iteration of the Greedy algorithm in
selecting individuals for the time series interventions using the edges from all

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 112 Spread of Sexual Health Information in Black Women

(a) 166
142
170

160
153
165 180 171
145
168 151 156
157
155 154 133 144
158
134
130 110
173
106 106
169 112
115 102
150 159 104 107 111
121
146
137 129
131

132
139 122
140 114
138
113 141 127
135 120
124 109 126
148
149
123 105 103
152 167 147
164 143
118 101
117

163
162

175 176
128 136 115
119

(b) 166
142
170

153 160
165 160 171
145
168 151 156 157
155 154 133 144
158
134
130 110
173
106 108
169 112
150 107 116 102
159 104 111
121
137 146 129
131

132 122
133
140 114
138
113 141 127
138 120
109
148 124 126
149
123 105 103
152 167 147
164 143
118 101
117
163
162

175 176
126 136 119 115

Figure 6.7 Dark gray nodes identify the 12 most commonly selected peer leader
nodes among the 50 trials of the HEALER algorithm for (a) texting based
interventions and (b) in-person interventions. Here we assume that nodes can be
influenced regardless of whether or not sex is already discussed. Note that the 12
nodes do not appear as the selected peer leaders during one trial, but rather are the
most commonly selected nodes across trials.

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 Elizabeth Bondi et al. 113

142
166 170

160
153
165 171
180
145
168
151 156
157
155 154 133
158 144
134
130 110
173
106 108
169 112

111 116 102

150 159 104 107
121
146 129
137 131
139 132 122

140
138 114
113 141
135 120 127
109
124 126
148
149 103
123 105
152 147
167
164 143
118 101
117

163
162

175 176
128
136
119 115

Figure 6.8 Dark gray nodes identify the selected peer leaders in the first time
series intervention of the Greedy algorithm, including all possible edges in the
network. Light gray nodes identify the selected peer leaders in the second time
series intervention, and black nodes identify the selected peer leaders in the third
(final) intervention. This is a result from one specific trial.

communication modes regardless of whether or not sex is discussed. Similarly,

in Figure 6.9, we also show a result of one iteration of the HEALER algorithm
in selecting individuals for the time series interventions. The dark gray nodes
are selected for the first intervention, the light gray nodes are selected for the
second intervention, and the black nodes are selected for the third intervention.
Note that in some cases, counterintuitive nodes can be selected (for example,
node 173). These cases typically have a slightly smaller reward than the
example provided in Figure 6.9, though they sometimes can have a higher
reward than the example provided in Figure 6.9. These variations are due to the
randomness in influence spread. Also note that the order of the nodes chosen
in each round in our case is irrelevant, as there is no additional information
after each round. We include multiple rounds in these experiments, since future
work could provide additional information after each round.

Conclusions and Future Work

In this work, we used real-world survey data from a social network of YBW
in combination with computational approaches for solving the maximum

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 114 Spread of Sexual Health Information in Black Women

170
142
166
160
153
171 145
165 180
168 156
151
157 133
155 154 144
134
130 110
173
106 108

169 112

111 116 102

150 159 104 107
121
146 129
137 131
132 122
139
140
138 114
113 141
135 120 127
109
124 126
148
149 103
123 105
152 147
167
164 143
118 101
117

163
162

175 176
128
136
119 115

Figure 6.9 Dark gray nodes identify the selected peer leaders in the first time
series intervention of the HEALER algorithm, including all possible edges in the
network. Light gray nodes identify the selected peer leaders in the second time
series intervention, and black nodes identify the selected peer leaders in the third
(final) intervention. This is a result from one specific trial.

influence problem. We demonstrated that in the communication modality of

a peer leader, social network–based intervention program for HIV reduction
can impact the extent of information spread and the best individuals to select
for training. We also compared the results of two known approximation
algorithms for solving the maximum influence problem. We considered several
different network structures, each of which allowed us to evaluate informa-
tion diffusion when making different simplifying assumptions about intra-
network communication. The results of our experiments show that while a
multimodal communication intervention training program would likely result
in the greatest spread of information, a texting-based intervention would yield
similar results for our surveyed network of YBW in the case of limited training
resources. Cost and benefit analysis could further be used to determine which
of the solutions (i.e., text-based or multimodal communication) would be most
advantageous for interventionists depending on the time and resources needed
for development, training, and implementation, and the overall reach of the
intervention and changes in behavioral outcomes.
Currently, many social network–based interventions are designed with
limited consideration on the best mode of dissemination for effective and

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 Elizabeth Bondi et al. 115

efficient diffusion of information and behavioral changes within targeted

social networks. We propose that for future peer-led intervention studies,
knowledge of the communication modalities and their prevalence within the
target network is critical to maximize the diffusion of information. Since our
experiments considered one specific network of YBW in the Los Angeles area,
with a data set collected specifically to understand communication among the
network members on the subject of sex and sexual health, the conclusion of
texting as the best intervention platform is specific to this network. However,
we hope that our results convince intervention researchers who are designing
future peer-led and/or social network–based intervention studies to consider
the effects of multimodal communication in their intervention designs and
dissemination protocols. With knowledge of the network structure and the
communication mode for each edge of the network, computer-based simula-
tions of social network–based interventions can be completed at virtually no
cost to assist in determining the best mode of dissemination (e.g., in-person,
social media, text, or a multimodal variation).
Within the computer science domain, our experiments considered a simpli-
fied network that did not exhibit highly complex behaviors or properties. In the
most general case, a network could be unstable, that is, with members joining
and leaving the network and connections being formed and destroyed over
time. Or, perhaps individuals could also influence other network members to
use new platforms for communication, which we assumed in our experiments
was not possible. One could also assume that information is spread over each
communication platform with some specific probability of success, which
could vary from connection to connection, from individual to individual,
or as a function of the communication modality. Additionally, perhaps the
peer leaders trained to spread information using a single communication
mode (and the subsequently activated individuals in the network) might also
disseminate information using communication modes other than the one
originally selected for training. In other words, perhaps information could be
modeled as moving “across platforms,” rather than being restricted to flow
over one channel. Furthermore, we could also consider the uncertainty on the
probability of influence on neighboring network members, which we simply
assumed was 0.6 for all edges in this work. These complex network properties
and modes of information diffusion increase the computational difficulty of
the maximum influence problem, rendering it significantly more difficult to
model mathematically. However, the inclusion of these network behaviors
or properties would be of interest to consider in future work, where novel
algorithms would likely need to be designed to address these complex network
properties.

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 116 Spread of Sexual Health Information in Black Women

Figure 6.10 Say individual A is a YBW who has been recruited and qualified to
participate in the study. Suppose individual A then listed SNMs B, C, and D as
her connections, but only individuals B and C were YBW eligible to participate
in the study. Perhaps A encouraged B and C to participate in the study, but only
B followed up and was enrolled. With our existing network analysis, the only
possible edges to be included in the network would be links between individuals
A and B (denoted by a solid black line), because C and D are not themselves
study participants (linked only by dotted gray lines). Now suppose that another
study participant, individual Z, also listed D as a top connection, but did not
list individual A. In other words, suppose that individuals A and Z do not know
each other but have one friend (D) in common. This type of connection, called
a “secondary” connection (denoted by a dashed black line), can be important for
modeling information diffusion throughout real-world networks. In this example,
if information flows from A to D and then from D to Z, then individual A
could influence Z, even though these two nodes are not directly connected
by an edge.

The social network survey methodology allowed for study participants to

list all of their top network connections – the people with whom they most
frequently communicated. This often included parents, significant others, and
friends, many of whom were not themselves surveyed panelists. For example,
see Figure 6.10. Since our analysis included only study participants, we
are most likely missing some secondary and tertiary connections involving
non-participants. Such added connections would lead to more information
diffusion throughout the network. Hence, our results should be considered as
a conservative estimate of indirect influence, and future work should endeavor
to include connections to individuals who are not study participants for a more
accurate model.
In summary, we found that a texting-based intervention would yield similar
results to a multimodal intervention for our surveyed network of YBW in
the case of limited training resources. To expand on this result in the future,
we suggest examining social networks longitudinally to account for structural
changes in networks of YBW and considering more complex networks with
characteristics such as instability and communication spread via modalities
other than those used for training, using a full network for analysis, including
secondary and tertiary connections, and taking multimodal communication

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 Elizabeth Bondi et al. 117

into account in recruitment and study design. These next steps will allow us
to define more accurate models of the possible influence we can spread in a
real YBW network.

References

Social Work Sources

Centers for Disease Control and Prevention. (2016). HIV Surveillance Report, 2015;
vol. 27. Retrieved December 15, 2016, from National Centers for HIV/AIDS,
Viral Hepatitis, STD, and TB Prevention: www.cdc.gov/hiv/library/reports/hiv-
surveillance.html
Convey, M. R., Dickson-Gomez, J., Weeks, M. R., & Li, J. (2010). Altruism and
peer-led HIV prevention targeting heroin and cocaine users. Qualitative Health
Research, 20(11), 1546–1557.
Davey-Rothwell, M. A., Tobin, K., Yang, C., Sun, C. J., & Latkin, C. A. (2011). Results
of a randomized controlled trial of a peer mentor HIV/STI prevention intervention
for women over an 18 month follow-up. AIDS and Behavior, 15(8), 1654–1663.
Jennings, J. M., Howard, S., & Perotte, C. L. (2014). Effects of a school-based sexuality
education program on peer educators: The Teen PEP model. Health Education
Research, 29(2), 319–329.
Kelly, J. A., St. Lawrence, J. S., Diaz, Y. E., Stevenson, L. Y., Hauth, A. C., Brasfield,
T. L., . . . & Andrew, M. E. (1991). HIV risk behavior reduction following
intervention with key opinion leaders of population: An experimental analysis.
American Journal of Public Health, 81(2), 168–171.
Lenhart, A., & Center, P. R (2015). Teens, social media & technology overview 2015.
Pew Research Center, 9. Retrieved October 13, 2016, from Internet, Science &
Tech: Report: www.pewinternet.org/2015/04/09/teens-social-media-technology-
2015/
Mahat, G., Scoloveno, M. A., De Leon, T., & Frenkel, J. (2008). Preliminary evidence
of an adolescent HIV/AIDS peer education program. Journal of Pediatric Nursing,
23(5), 358–363.
McKirnan, D. J., Tolou-Shams, M., & Courtenay-Quirk, C. (2010). The Treatment
Advocacy Program: A randomized controlled trial of a peer-led safer sex inter-
vention for HIV-infected men who have sex with men. Journal of Consulting and
Clinical Psychology, 78(6), 952.
Noar, S. M., & Willoughby, J. F. (2012). eHealth interventions for HIV prevention.
AIDS Care, 24(8), 945–952.
O’Grady, M. A., Wilson, K., & Harman, J. J. (2009). Preliminary findings from a brief,
peer-led safer sex intervention for college students living in residence halls. The
Journal of Primary Prevention, 30(6), 716.
Rice, E. (2010). The positive role of social networks and social networking technology
in the condom-using behaviors of homeless young people. Public Health Reports,
125(4), 588–595.
Schneider, J. A., Zhou, A. N., & Laumann, E. O. (2015). A new HIV prevention
network approach: sociometric peer change agent selection. Social Science &
Medicine, 125, 192–202.

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 118 Spread of Sexual Health Information in Black Women

Smith, A. (2014). African Americans and technology use: A demographic portrait.

Washington, DC: Pew Research Center. Retrieved from Pew Research Center:
http://www.pewinternet.org/2014/01/06/african-americans-and-technology-use/
Rogers, E. M. (2010). Diffusion of innovations. New York: Simon and Schuster.
Young, S. D., Rivers, C., & Lewis, B. (2014). Methods of using real-time social media
technologies for detection and remote monitoring of HIV outcomes. Preventive
Medicine, 63, 112–115.
Valente, T. W., & Pumpuang, P. (2007). Identifying opinion leaders to promote behavior
change. Health Education & Behavior, 34(6), 881–896.

Artificial Intelligence Sources

Borgs, C., Brautbar, M., Chayes, J., & Lucier, B. (2014, January). Maximizing social
influence in nearly optimal time. In Proceedings of the Twenty-Fifth Annual ACM-
SIAM Symposium on Discrete Algorithms (pp. 946–957). Society for Industrial
and Applied Mathematics.
Cohen, E., Delling, D., Pajor, T., & Werneck, R. F. (2014, November). Sketch-
based influence maximization and computation: Scaling up with guarantees.
In Proceedings of the 23rd ACM International Conference on Conference on
Information and Knowledge Management (pp. 629–638). ACM.
Golovin, D., & Krause, A. (2011). Adaptive submodularity: Theory and applications
in active learning and stochastic optimization. Journal of Artificial Intelligence
Research, 42, 427–486.
Kempe, D., Kleinberg, J., & Tardos, É. (2003, August). Maximizing the spread of
influence through a social network. In Proceedings of the Ninth ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining (pp. 137–
146). ACM.
Tang, Y., Xiao, X., & Shi, Y. (2014, June). Influence maximization: Near-optimal time
complexity meets practical efficiency. In Proceedings of the 2014 ACM SIGMOD
International Conference on Management of Data (pp. 75–86). ACM.
Yadav, A., Chan, H., Xin Jiang, A., Xu, H., Rice, E., & Tambe, M. (2016, May).
Using social networks to aid homeless shelters: Dynamic influence maximization
under uncertainty. In Proceedings of the 2016 International Conference on
Autonomous Agents & Multiagent Systems (pp. 740–748). International Founda-
tion for Autonomous Agents and Multiagent Systems.
Yadav, A., Soriano Marcolino, L., Rice, E., Petering, R., Winetrobe, H.,
Rhoades, H., . . . & Carmichael, H. (2015). Preventing HIV spread in homeless
populations using PSINET: emerging application case study. IAAI.
Yadav, A., Wilder, B., Rice, E., Petering, R., Craddock, J., Yoshioka-Maxwell, A.,
Hemler, M., Onasch-Vera, L., Tambe, M., Woo, D. (2017). Influence maximization
in the field: The arduous journey from emerging to deployed application. AAMAS.

https://doi.org/10.1017/9781108669016.008 Published online by Cambridge University Press

 7
Minimizing Violence in Homeless Youth
Ajitesh Srivastava, Robin Petering, and Michail Misyrlis

Introduction
There are an estimated 1.6–2.8 million youths experiencing homelessness in
the United States [1]. Although violence in the United States has steadily
decreased during the past decade, homeless youth (HY) remain dispropor-
tionately susceptible to violent victimization and perpetration [2]. These
youths experience all types of violence at higher rates than their housed
counterparts [3–5].
Interpersonal violence is defined as the intentional use of physical force
or power between two or more individuals resulting in injury, death, or
psychological harm [6]. This is typically the result of many contributing
factors including childhood experiences of trauma such as physical abuse
or neglect [7]. Many youths cite escaping a violent family environment as
a reason for becoming homeless [8–12]. This experience and exposure to
violence may contribute a propensity toward violence during adolescence or
young adulthood as the youth may reenact violent behaviors learned from
previous perpetrators of abuse [13]. Subsistence survival strategies, including
delinquent activities such as stealing, burglary, prostitution, and dealing drugs
to obtain money, food, or shelter [14–16], and exposure to perpetrators while
living on the streets increase likelihood of experiencing violence [8–10]. A
street-entrenched life increases exposure to violence, and many youths utilize
physical aggression to guarantee safety from future victimization [17–19].
Further, interpersonal violence occurs within a dyadic space, but it is also
contagious in a social network, diffusing across social space. A young person’s
risk of engaging in violence increases as the proportion of violent peers in his
or her network increases [3].

119

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 120 Minimizing Violence in Homeless Youth

The consequences of violence for homeless youth are severe. The proximal
consequence of physical injury is a primary health concern for HY [20–22],
who do not proactively seek health services [21]. Minor, treatable injuries
can often escalate into more severe health problems [20, 23, 24]. Violence
can also lead to nonphysical ailments as posttraumatic stress disorder [25–
27], depression [28, 29], and externalizing behavior (i.e., delinquency and
aggression; [27]. Interpersonal violence also limits a youths ability to suc-
cessfully exit homelessness by increasing the risk of of interaction with law
enforcement, arrest, and imprisonment [30–33], which can limit employment
and housing opportunities. Violence can also result in the temporary or
permanent termination of services at agencies designed to support and assist
HY in meeting needs or securing housing.
Reducing violence in the lives of homeless youth is imperative and will
contribute to a young person’s ability to safely and successfully exit the streets
and lead a long and productive life in society. However, public health and social
interventions to reduce violence within adolescent and young adult populations
are difficult [34], because, again, this phenomenon is complex, with many
intrinsic and extrinsic contributing factors.
Additionally, violence perpetuates violence and diffuses through a network
like a contagious disease [35]. Motivated by this contagious nature, a diffusion
model is ideal for modeling the spread of violence. Doing so can lead to
optimal intervention strategies under certain assumptions. To the best of our
knowledge, intervention strategies to reduce violence using diffusion models
have received very little attention in the literature [36, 37]. Violence is modeled
based on susceptibility and infectiousness in [36]. In [37], the idea of opposing
forces – “provocation” and “repression” – is used to model violence as two
diffusion processes. This is more accurate as it captures the nonprogressive
nature of violence, where an individual may switch between being “violent”
and “nonviolent.” However, it is a macroscopic approach, which disregards
the network structure. We propose the Uncertain Voter Model (UVM), which
is an extension of the Voter Model [38]. In the Voter Model, individuals
are influenced by a randomly selected neighbor.1 UVM allows for some
uncertainty in the knowledge of the neighborhood that may arise from an
individual being influenced by someone they did not explicitly state as their
“friends” during the survey to create the network. Our model also allows for
uncertainty in number of time-steps for which the model needs to be run to

1
We use the terms “neighbor” and “neighborhood” to refer to the links of a given individual in
the network and not their physical neighborhood

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 Ajitesh Srivastava, Robin Petering, and Michail Misyrlis 121

estimate the expected spread of violence. Specifically, our contributions are as

follows:

• We propose the Uncertain Voter Model for spread of violence, which can
capture its nonprogressive nature and take into account the uncertainty in a
neighborhood as well as uncertainty in the time period over which the
model is run.
• We formally define the Violence Minimization problem where the task is to
perform intervention with a finite resource, i.e., change the state of k
individuals so that total expected number of violent individuals is
minimized.
• We show that the Uncertain Voter Model can be reduced to the classic voter
model, and thus a Greedy algorithm forms the optimal solution to Violence
Minimization.
• We perform experiments on a real-life network of HY and find the nodes to
be selected for intervention.

Data Collection
A sample of 481 HY aged 18 to 25 years accessing services from two day-
service drop-in centers for HY in Hollywood and Santa Monica, CA, were
approached for study inclusion in October 2011 and February 2012. The
research team approached all youths who entered the service agencies during
the data collection period and invited them to participate in the study. The
selected agencies provided weekday services to eligible HY, including basic
needs, medical and mental health services, case management, and referrals and
connections to other programs such as housing services. Each youth signed a
voluntary consent form, and a consistent pair of research staff members was
responsible for all recruitment to prevent youths from completing the survey
multiple times during each data collection period per site. The overall response
rate was 80.1%; 19.9% of HY approached declined to participate, 6.2% did not
complete the full survey, and 2.6% completed the surveys at both sites three
months apart. Four participants were excluded because they were younger than
18 to limit the sample to late adolescence and early adulthood. The final sample
consisted of 366 participants. The institutional review board of University of
Southern California approved all procedures and waived parental consent for
minors without parents or guardians. All participants received $20 in cash or
gift cards as compensation for their time.

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 122 Minimizing Violence in Homeless Youth

The study consisted of two parts: a computerized self-administered sur-

vey and a social network interview. The computerized survey included an
audio-assisted version for participants with low literacy and was available in
English or Spanish. The computerized survey included approximately 200
questions and took an average of 1 hour to complete. The social network
interview was interactive, conducted by a trained research staff member. Social
network interviews took between 15 and 30 minutes to complete, depending
on each participant’s personal network size. For the social network interview,
each participant was asked to name anyone they interacted with in person,
on the phone, or through the internet in the previous month, prompted by
interviewers stating, “These might be friends; family; people you hang out
with/chill with/kick it with/ have conversations with; people you party with –
use drugs or alcohol; boyfriend/girlfriend; people you are having sex with;
baby mama/baby daddy; case worker; people from school; people from work;
old friends from home; people you talk to (on the phone, by email); people
from where you are staying (squatting with); people you see at this agency;
other people you know from the street.” Details describing the social network
data collection can be read in [3].

Dependent Variable
The variable of interest is violent behavior. Violent behavior was assessed by
recent participation in a physical fight. Participants were asked: “During the
past 12 months, how many times were you in a physical fight?” Eight ordinal
responses ranged from “zero times” to “over 12 times.” Due to the skew of the
original variable, responses were dichotomized similar to previous literature
on youth violence [5, 39] to distinguish between participants who had been
in no physical fights and participants who had been in at least one physical
fight during the previous year. This question was adopted from the Youth Risk
Behavior Survey [40] and did not distinguish between victims and perpetrators
of violence.

Model
To model the spread of violence, we model the network of homeless youth
as a graph G(V, E) where every individual is a node that can exist in one
of two states: violent or nonviolent. We have chosen to model violence as
a nonprogressive diffusion process, i.e, a node may switch its state, unlike
the progressive diffusion where once a node is violent, it cannot become

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 Ajitesh Srivastava, Robin Petering, and Michail Misyrlis 123

nonviolent again. Next, we provide a background on the Voter Model [38] on

which our model is based.

Voter Model
In the Voter Model [38], at every time step a node u picks an incoming
neighbor v at random with a probability p(v, u). The incoming probabilities

are normalized such that v p(v, u) = 1. Let xu,t represent the probability of
node u being violent at time t. According to the model,

xu,t = pv,u xv,t−1 . (7.1)
v

Let xt represent the state of all the nodes at time t, with ith element representing
the probability that vi ∈ V is violent at time t. Suppose matrix M represents the
transpose of the adjacency matrix of the weighted network, i.e., Mu,v = p(v, u).
Then
xt = Mxt−1 . (7.2)

It follows that
xt = M t x0 . (7.3)

Here x0 is the initial state of nodes, which is assumed to be known. Now we

wish to select k nodes out of those who are violent at t = 0 and turn them into
nonviolent so that the expected number of nodes that are violent at time t is
minimized. Define IX for X ⊆ V as the vector in which the i-th element is 1 if
vi ∈ X. Then the expected number of violent nodes at time t is given by
 
P(vi is violent at time) = xi,t = IVT xt (7.4)
i i

Formally we define the problem of Violence Minimization as follows.

Problem Definition (Violence Minimization) Given a weighted graph

G(V, E), an initial set of violent nodes S , a time frame t, and an integer k,
find T ⊆ S such that |T |= k, turning the nodes in T into nonviolent minimizes
the expected number of violent nodes after time t, i.e., IVT xt under voter model.

Uncertain Voter Model

A network formed through a survey may have missing edges due to the
uncertainty in a person’s ability to recall all “friends” they might be influenced
by. To capture this aspect, we propose the Uncertain Voter Model, where we

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 124 Minimizing Violence in Homeless Youth

assume that any node not directly connected to the node of interest may also
influence it. In this model, two mutually exclusive events happen: (1) with
probability θ a node randomly selects one incoming neighbor and adopts its
state; (2) with probability (1−θ) it selects a random node that is not its neighbor
in the network and adopts its state. Mathematically,
  1
xu,t = θ pv,u xv,t−1 + (1 − θ) xv,t−1 (7.5)
{v|p(v,u)>0} {v|p(v,u)=0}
|{v|pv,u = 0}|

If n is the total number of nodes and du is the number of incoming neighbors

of u, then |{v|pv,u = 0}|= n − du . Suppose we define,
⎧
⎪
⎪
⎨θpv,u if pv,u > 0
qθ (v, u) = ⎪
⎪
⎩ 1−θ if pv,u = 0 .
n−du

Now, Equation (7.5) can be rewritten as


xu,t = q(v, u)xv,t−1 or xt = Qθ xt−1 (7.6)
v

where [Qθ ]u,v = qθ (u, v), which reduces to the Voter Model (Equation 7.1),
where the transpose of the adjacency matrix is Qθ .

Greedy Minimization
Let x0 be the vector formed by turning some k nodes nonviolent, resulting in
the vector of probabilities xt at time t. Now, minimizing IVT xt is equivalent to
maximizing IVT (xt −xt ) = IVT Qtθ (x0 −x0 ), i.e., the problem reduces to maximizing

IVT Δxt = IVT Qtθ Δx0 = IVT Qtθ Iu (7.7)
{u|Δx0 (u)=1}

which can be optimized using Greedy strategy [38] as presented in

Algorithm 1.
The most expensive step of the algorithm is the computation of Qtθ that can
be computed in O(|V|2.3 log t).

Algorithm 1 Greedy algorithm to minimize violence

function MinViolence(G, S , θ, k, t)
Compute Qtθ for G
∀u ∈ S compute σ(u) = IV Qtθ Iu
Sort {σ(u)} in descending order and return top k.
end function

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 Ajitesh Srivastava, Robin Petering, and Michail Misyrlis 125

Figure 7.1 Visualization of the homeless youth network. The grey nodes represent
the violent nodes and the green ones represent the nonviolent ones. The black
nodes have an unknown state.

Uncertainty in Time
The Uncertain Voter Model requires t as a parameter that is unknown in real
life. While we may have a certain time period (days or weeks) over which
we want the intervention to work, finding a relation between that time period
and the parameter t is nontrivial, as it depends on how often the individuals
interact. To capture this uncertainty, we assume that time t takes a value τ with
probability P(t = τ). Now, we wish to minimize E(IV xt ) where the expectation
is taken over t. Therefore,


E(IVT xt ) = P(t = τ)IVT Qτθ x0
τ
⎛ ⎞
⎜⎜⎜ ⎟⎟
= IVT ⎜⎜⎝ P(t = τ)Qθ ⎟⎟⎟⎠ x0 .
τ
(7.8)
τ

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 126 Minimizing Violence in Homeless Youth

Notice from Equation (7.8) that a Greedy solution like Algorithm 1 still
applies.

Experiments
We construct the network obtained by the surveyed data, which consists of
369 nodes and 558 directed edges. Due to the lack of the knowledge of edge-
weights, we assume that all incoming links for a node are equally weighted. We
assume that the surveyed data is the current snapshot of the diffusion process
(x0 ). Out of the 369 nodes, 55.01% are “violent” (xu,0 = 1) and 42.55%
are “nonviolent” (xu,0 = 0). Data on the rest of 2.44% are missing and are
assumed to be equally likely to be of either state (xu,0 = 0.5). Based on this
“initial state,” we run Greedy Minimization for the Uncertain Voter Model. The
t parameter, for the sake of simplicity, was assumed to be uniformly distributed
between 1 to 10. The range was selected, keeping in mind that homeless youth
networks are dynamic, and so in practice, the intervention should be performed
in “waves,” making the the selection process short-term.
We experimented with different values for parameter θ = 1, 0.9, 0.8, 0.7, 0.6
and 0.5, i.e., increasing edge uncertainty. We refrain from using lower values
of θ as it would represent very low confidence in the collected data, i.e, it
would mean that a node is more likely to be influenced by one of the other
nodes it is not connected to. Table 7.1 presents the top 10 nodes (in terms of
PID assigned in the survey) chosen for intervention. Note that there are many
nodes such as PIDs 35, 47, 2007, 4, and 2156 that consistently appear in the
top 10. Some of these nodes are shown in Figure 7.2. It is nontrivial to find
them, as they are a mix of high-degree and low-degree nodes. We also varied
the value of t = 2, 4, 6, 8, 10, and 12. The value of θ was set to 0.75 because it

Table 7.1. Top 10 seeds for various values of θ output by Greedy

Minimization

θ Selected Seeds E(IVT xt )

1 35 2156 47 13 4 2115 2007 51 38 2086 160.9487

0.9 35 47 2156 4 13 2007 2115 51 38 2086 119.8357
0.8 35 47 2007 4 2156 2115 51 13 2086 38 117.3772
0.7 47 35 2007 4 51 2115 2156 2086 13 38 116.6939
0.6 47 2007 35 4 51 2115 2086 2038 2156 38 117.9434
0.5 47 2007 35 4 51 2086 2115 2038 38 2110 121.4455

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 Ajitesh Srivastava, Robin Petering, and Michail Misyrlis 127

Table 7.2. Top 10 seeds for various values of t output by Greedy

Minimization

t Selected Seeds E(IVT xt )

2 47 2007 2038 51 2086 2115 4 35 2110 38 147.2344

4 47 2007 35 4 51 2115 2086 2038 2110 38 138.0431
6 47 35 2007 4 51 2115 2156 2086 2110 13 130.1730
8 47 35 2007 4 2115 51 2156 13 2086 2110 123.1098
10 47 35 2007 4 2156 2115 51 13 2086 38 116.7971
12 47 35 2007 4 2156 2115 51 13 2086 38 111.0703

Figure 7.2 Visualization of some PIDs consistently selected by the Greedy

Algorithm for various parameters.

can be shown that it is equivalent to assuming that θ is randomly picked from

the interval [0.5, 1].
We have shown that the Greedy Algorithm described in Algorithm 1 is
optimal. However, to compare how much better it is from a different choice
of intervention strategy, we compare it against the following baselines:

• Degree:

We define the degree of a node based on the weighted graph as
dv = u pv,u . Then we select top k nodes.
• Betweenness Centrality: Top k nodes are selected based on the betweenness
centrality in the graph.

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 128 Minimizing Violence in Homeless Youth

Figure 7.3 Comparison of the baseline against the Greedy Algorithm for varying
intervention sizes under the Uncertain Voter Model.

Figure 7.3 shows the comparison for expected number of nodes that are violent
after time t = 5 and t = 10. The value of θ was set to 1 to generate these plots.
Other values for parameters t and θ show similar trends and, hence, have been
omitted. We observe that the Greedy Algorithm significantly outperforms both
baselines.

Demographics of Chosen Individuals. Finally, to study the demographics

of the individuals who are selected for intervention using Greedy Algorithm,
we performed the selection by setting θ = 0.75 (which is equivalent to
averaging for all θ ∈ [0.5, 1]), and for t = 5. Individuals were selected until
the expected violence was reduced to 25%. This resulted in the selection of
38 individuals. The demographics of the selected individuals are presented in

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 Ajitesh Srivastava, Robin Petering, and Michail Misyrlis 129

Table 7.3. Demographics of selected top 38 nodes

Top Violent Nodes Remaining Sample

n % n %

Total 38 10.38 328 89.62

Male 27 71.05 233 71.04
Age 21.54 2.01 21.05 1.64
Time Homeless1 4.75 2.89 3.65 2.83 *
White 16 42.11 117 35.67
Latino 4 10.53 50 15.25
Black 11 28.95 97 29.66
Mixed 7 18.42 48 14.68
LGBTQ 14 36.84 75 23.29
site 19 50.00 183 55.79
Child Abuse 14 36.84 130 39.63
Fight 38 100.00 167 51.23 ***
Injury from Fight 16 42.11 61 18.83 ***
highboth 10 26.32 54 16.46
highaware 6 15.79 59 17.99
highcontrol 6 15.79 53 16.16
dersboth 24.30 7.67 24.31 7.27
Degree1 4.76 3.64 1.90 1.89 ***
% Violent network1 0.70 0.33 0.42 0.41 ***
Core 17 44.74 51 15.55 ***
1
Indicates Mean and standard deviation
*p < 0.05, ***p < 0.001

Table 7.3. The basic demographics match the larger sample where there is no
indication of selection biases based on race, gender, or sexual orientation. The
selected nodes represent approximately 10% of the network. Those selected
have been homeless on average for longer than the remaining sample (4.75
years versus 3.65, p < 0.05). They are also more likely to have experienced
severe violence resulting in injury (p < 0.001). They have an average of 4.76
connections in the HY social network, which is significantly more than the
remaining sample average (1.90 connections, p < 0.001). Additionally, they
had a greater proportion of their network who were violent (0.70 versus 0.42,
p < 0.001) and are more likely to be in the core of the homeless youth network
(p < 0.001).

Discussion
We have proposed the Uncertain Voter Model in which, with probability θ,
a node selects one of its neighbors and with 1 − θ it selects one of the

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 130 Minimizing Violence in Homeless Youth

remaining nodes and adopts its state. We have shown that a Greedy Algorithm
is the optimal intervention strategy to minimize violence under this model.
The model captures uncertainty in time by assuming that the time over
which intervention is to be performed has a probability distribution over a
certain range. We show in our experiments that for sensible choice of these
parameters, the top individuals selected for intervention roughly remain the
same. However, for more accurate learning of the parameters, more data is
necessary. Moreover, with more data, a more complex model can be learned
that more accurately models the dynamics of violence diffusion by accounting
for personal characteristics. These extensions of our model are discussed in the
following.

Multiple Waves
We have performed all our experiments on one snapshot (wave) of data. Having
multiple waves of data may help obtain a more accurate modeling. Suppose the
two snapshots generate the vector of states x0 and xr . Then, according to UVM,
∃ θ, r, such that
xr = Qrθ x0 (7.9)

We can learn these parameters using maximum likelihood, i.e.,

arg maxθ P(x0 , xr |G, θ). More generally, given a sequence of observations:
< xi1 , xi2 , xi3 , · · · >,

θ = arg max P(< xi1 , xi2 , xi3 , · · · > |G, θ). (7.10)
θ

We performed a similar analysis to the single snapshot that was available

x0 , assuming that it is a steady state, i.e., x0 = Qθ x0 . Maximum likelihood
resulted in θ = 1. However, this is under a steady state assumption that
might not hold. Therefore, more snapshots are necessary for a more accurate
modeling. Multiple snapshots would also enable the analysis of how quickly
the diffusion process takes place. This would aid in understanding a mapping
between time parameter of the model and time elapsed in real life. Another
advantage of multiple snapshots would be to understand and take into account
the evolution of the network itself, i.e., rate of addition and removal of nodes
and links.
We also suggest that the intervention should be performed over short
periods of time and the effects should be observed. This new wave would help
in retraining the model and finding more accurate nodes for intervention in the
next iteration.

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 Ajitesh Srivastava, Robin Petering, and Michail Misyrlis 131

Utilizing Other Variables

We have only taken into account the state of the individuals (violent or
nonviolent), which came from whether they had been involved in a fight over
the past 12 months. Many more features for individual nodes are available
from the collected data. One feature of particular interest is Difficulty in
Emotion Regulation (DERS) [41]. Intuitively, an individual with high DERS
is likely to have a higher propensity for violence. Suppose α(u) is the
probability that a node u, given its DERS score, would prefer to be violent.
Mathematically,

α(u) j qθ (v, u)xv,t−1
xu,t =   . (7.11)
α(u) j qθ (v, u)xv,t−1 + (1 − α(u)) j qθ (v, u)(1 − xv,t−1 )
Assuming steady state for the available snapshot, we attempted to learn a
monotonic function α() using maximum likelihood. Inclusion of this function
made the error worse. In other words, α(u) = 0.5, ∀ u produced the minimum
error, which reduces this model to our Uncertain Voter Model. One reason
could be the fact that DERS scores and the state of violence/nonviolence
has a high correlation, and thus inclusion of DERS does not improve the
results. Again, having more snapshots may provide a more accurate learning
framework, and α(.) might play an important role. Moreover, we have observed
from the data that some nodes have been involved in violence more frequently
than others. With more snapshots we also wish to incorporate this informa-
tion into our model and form a more accurate representation of violence
diffusion that can take into account more information that is available in the
collected data.

Implications from Social Work Perspective

There are many implications for social work practice. Reducing violence
within homeless youth populations is imperative, as there are many conse-
quences. Even minimal incidents, such as those that do not result in an injury,
can severely impact a young person’s life. For example, verbally aggressive
conflict with another youth or a service provider can result in interruption
of supportive services, which impacts a youth’s ability to exit homelessness.
Additionally, violence perpetuates violence. A minimal reduction in violence
could have a global impact on the lives of young people. The use of
artificial intelligence and advanced analytics augments social work research
and practice to maximize efforts that may already be in place or provide a

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 132 Minimizing Violence in Homeless Youth

better understanding of real-world behavioral dynamics. What is unique about

the approach presented in this chapter is that artificial intelligence is being
used to determine how to disrupt the spread of violence within the network.
The violence minimization algorithm presented takes into consideration the
contagious nature of violence as well of the dynamic state of being violent.
Violence is different than other public health behaviors in that it is learned
and can be unlearned and then relearned. This property of violence requires
a different diffusion approach than prevention approaches that hinge on
information dissemination.
Future directions for applying this approach would be to use the violence
minimization algorithim to identify the key individuals in a network that
could be targeted for specific behavior change related to violence. Imagine
inviting key network nodes to particpate in a program to enhace intrinsic skills
when confronted with conflict, reducing impulsitivty and increased emotional
regulation. Such an intervention would attempt to eliminate violent nodes
in the larger network structure and therefore disrupt the spread of violence.
Implementing such an intervention will undoubtedly be challenging, as it
selects a group of individuals that are confirmed to be violent. It will be
necessary to think creatively, inclusively, and trauma-informed when designing
the specific intervention strategy to reduce intrinsic violent behaviors in these
individuals. However, as noted before, peer-based violence interventions are
being implemented with success. Despite the potential challenges, deploying
and testing an intervention program that utilizes this modeling technique
should be a priority given that violence-specific interventions for homeless
youth are currently lacking [34]. Future directions for this study will be to
explore the transmission of violence within a social network of homeless youth
using more precise metrics and watch changes overtime, without intervention,
to refine the algorithm modeling. Following that, developing and piloting an
intervention that incorporates the algorithim would be necessary. Regardless,
results from this chapter are extremely promising for reducing violence in
homeless youth populations as well as other groups that are at high risk of
experiencing violence.

References
[1] M. J. Terry, G. Bedi, and N. D. Patel. Healthcare needs of homeless youth in the
United States. Journal of Pediatric Sciences, 2(1):e17–e28, 2010.
[2] Department of Justice Office of Victims of Crime. 2013 National Crime Victims’
Rights Week Resource Guide: Section 6. Statistical Overviews, 2013.

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 Ajitesh Srivastava, Robin Petering, and Michail Misyrlis 133

[3] R. Petering, E. Rice, H. Rhoades, and H. Winetrobe. The social networks of

homeless youth experiencing intimate partner violence. Journal of Interpersonal
Violence, 29(12):2172–2191, 2014.
[4] J. A. Heerde, S. A. Hemphill, and K. E. Scholes-Balog. “Fighting” for survival: A
systematic review of physically violent behavior perpetrated and experienced by
homeless young people. Aggression and Violent Behavior, 19(1):50–66, 2014.
[5] D. K. Eaton, L. Kann, S. Kinchen, S. Shanklin, K. H. Flint, J. Hawkins,
W. A. Harris, R. Lowry, T. McManus, D. Chyen, et al. Youth risk behavior
surveillance – United States, 2011. Morbidity and Mortality Weekly Report.
Surveillance Summaries (Washington, DC: 2002), 61(4):1–162, 2012.
[6] E. G. Krug, J. A. Mercy, L. L. Dahlberg, and A. B. Zwi. The World Report on
Violence and Health. The Lancet, 360(9339): 1083–1088, 2002.
[7] S. M. Wolfe, P. A. Toro, and P. A. McCaskill. A comparison of homeless
and matched housed adolescents on family environment variables. Journal of
Research on Adolescence, 9(1):53–66, 1999.
[8] N. G. Milburn, M. J. Rotheram-Borus, E. Rice, S. Mallet, and D. Rosenthal.
Cross-national variations in behavioral profiles among homeless youth. American
Journal of Community Psychology, 37(1–2):63, 2006.
[9] M. J. Robertson and P. A. Toro. Homeless youth: Research, intervention, and
policy. In Practical Lessons: The 1998 National Symposium on Homelessness
Research. Washington, DC: US Department of Housing and Urban Development
and US Department of Health and Human Services, 1999.
[10] L. B. Whitbeck and D. R. Hoyt. Nowhere to Grow: Homeless and Runaway
Adolescents and Their Families. Transaction Publishers, 1999.
[11] L. B. Whitbeck, D. R. Hoyt, and K. A. Ackley. Abusive family backgrounds
and later victimization among runaway and homeless adolescents. Journal of
Research on Adolescence, 7(4):375–392, 1997.
[12] M. R. Zide and A. L. Cherry. A typology of runaway youths: An empirically based
definition. Child and Adolescent Social Work Journal, 9(2):155–168, 1992.
[13] T. I. Herrenkohl, C. Sousa, E. A. Tajima, R. C. Herrenkohl, and C. A. Moylan.
Intersection of child abuse and children’s exposure to domestic violence. Trauma,
Violence, & Abuse, 9(2): 84–99, 2008.
[14] S. W. Baron. Street youth, gender, financial strain, and crime: Exploring broidy
and agnew’s extension to general strain theory. Deviant Behavior, 28(3):273–302,
2007.
[15] D. M. Crawford, L. B. Whitbeck, and D. R. Hoyt. Propensity for violence
among homeless and runaway adolescents: An event history analysis. Crime &
Delinquency, 57(6):950–968, 2011.
[16] P. A. Toro, A. Dworsky, and P. J. Fowler. Homeless youth in the united states:
Recent research findings and intervention approaches. In National Symposium on
Homelessness Research, 2007.
[17] S. W. Baron. Street youth violence and victimization. Trauma, Violence, & Abuse,
4(1):22–44, 2003.
[18] S. W. Baron, L. W. Kennedy, and D. R. Forde. Male street youths’ conflict:
The role of background, subcultural, and situational factors. Justice Quarterly,
18(4):759–789, 2001.

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 134 Minimizing Violence in Homeless Youth

[19] S. Gaetz. Safe streets for whom? Homeless youth, social exclusion, and crim-
inal victimization. Canadian Journal of Criminology and Criminal Justice,
46(4):423–456, 2004.
[20] J. Ensign and J. Gittelsohn. Health and access to care: Perspectives of homeless
youth in Baltimore City, USA. Social Science & Medicine, 47(12):2087–2099,
1998.
[21] J. Ensign and A. Panke. Barriers and bridges to care: Voices of homeless female
adolescent youth in seattle, washington, usa. Journal of Advanced Nursing,
37(2):166–172, 2002.
[22] M. L. Forst, J. Harry, and P. A. Goddard. A health-profile comparison of
delinquent and homeless youths. Journal of Health Care for the Poor and
Underserved, 4(4):386–400, 1993.
[23] S. W. Hwang. Homelessness and health. Canadian Medical Association Journal,
164(2):229–233, 2001.
[24] J. Vanderleest. Medically uninsured and the homeless. In The Care of the
Uninsured in America, pages 153–160. Springer, 2010.
[25] K. Bender, S. Thompson, K. Ferguson, and L. Langenderfer. Substance use
predictors of victimization profiles among homeless youth: A latent class analysis.
Journal of Adolescence, 37(2):155–164, 2014.
[26] K. M. Fitzpatrick and J. P. Boldizar. The prevalence and consequences of exposure
to violence among African-American youth. Journal of the American Academy of
Child & Adolescent Psychiatry, 32(2):424–430, 1993.
[27] S. Overstreet and S. Braun. Exposure to community violence and post-traumatic
stress symptoms: Mediating factors. American Journal of Orthopsychiatry,
70(2):263, 2000.
[28] R. D. Latzman and R. R. Swisher. The interactive relationship among adolescent
violence, street violence, and depression. Journal of Community Psychology,
33(3):355–371, 2005.
[29] J. J. Mazza and S. Overstreet. Children and adolescents exposed to community
violence: A mental health perspective for school psychologists. School Psychol-
ogy Review, 29(1):86, 2000.
[30] X. Chen, L. Thrane, L. B. Whitbeck, and K. Johnson. Mental disorders, comorbid-
ity, and postrunaway arrests among homeless and runaway adolescents. Journal
of Research on Adolescence, 16(3):379–402, 2006.
[31] B. W. Miles and S. K. Okamoto. The social construction of deviant behavior in
homeless and runaway youth: Implications for practice. Child and Adolescent
Social Work Journal, 25(5):425, 2008.
[32] M. Schwartz, H. K. Sorensen, S. Ammerman, and E. Bard. Exploring the
relationship between homelessness and delinquency: A snapshot of a group
of homeless youth in San Jose, California. Child and Adolescent Social Work
Journal, 25(4):255, 2008.
[33] J. R. Yoder, K. Bender, S. J. Thompson, K. M. Ferguson, and B. Haffejee.
Explaining homeless youths’ criminal justice interactions: Childhood trauma or
surviving life on the streets? Community Mental Health Journal, 50(2):135–144,
2014.

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 Ajitesh Srivastava, Robin Petering, and Michail Misyrlis 135

[34] R. Petering, S. Wenzel, and H. Winetrobe. Systematic review of current intimate

partner violence prevention programs and applicability to homeless youth. Jour-
nal of the Society for Social Work and Research, 5(1):107–135, 2014.
[35] J. Fagan, D. L. Wilkinson, and G. Davies. Social contagion of violence. The
Cambridge Handbook of Violent Behavior, Daniel Flannery, A. Vazsonyi, I. Wald-
man, eds., Cambridge University Press, 2007; Columbia Public Law Research
Paper No. 06-126. Available at SSRN: https://ssrn.com/abstract=935104
[36] D. J. Myers. The diffusion of collective violence: Infectiousness, susceptibility,
and mass media networks 1. American Journal of Sociology, 106(1):173–208,
2000.
[37] D. J. Myers and P. E. Oliver. The opposing forces diffusion model: The initiation
and repression of collective violence. Dynamics of Asymmetric Conflict, 1(2):164–
189, 2008.
[38] E. Even-Dar and A. Shapira. A note on maximizing the spread of influence
in social networks. In International Workshop on Web and Internet Economics,
pages 281–286. Springer, 2007.
[39] J. Duong and C. Bradshaw. Associations between bullying and engaging in
aggressive and suicidal behaviors among sexual minority youth: the moderating
role of connectedness. The Journal of School Health, 84(10): 635–645, 2014.
[40] Centers for Disease Control and Prevention: 2011 State and Local Youth
Risk Behavior Survey. Available: http://www.cdc.gov/healthyyouth/yrbs/pdf/
questionnaire/2011_hs_questionnaire.pdf.2011. Accessed August 11, 2011
[41] K. L. Gratz and L. Roemer. Multidimensional assessment of emotion regula-
tion and dysregulation: Development, factor structure, and initial validation of
the Difficulties in Emotion Regulation Scale. Journal of Psychopathology and
Behavioral Assessment, 26(1): 41–54, 2004.

https://doi.org/10.1017/9781108669016.009 Published online by Cambridge University Press

 8
Artificial Intelligence for Improving Access
to Sexual Health Necessities for Youth
Experiencing Homelessness
Aida Rahmattalabi, Laura Onasch-Vera, Orlando Roybal,
Kien Nguyen, Luan Tran, Robin Petering,
Professor Eric Rice, and Professor Milind Tambe

Introduction
Homelessness and HIV Risk
Homelessness has been an issue in the United States for several decades and
is continuing to grow. In 2013, the U.S. Department of Education and the
U.S. Census found that there were approximately 1.6 million youth who were
experiencing homelessness. More than 500,000 of those youth were homeless
in California (Bassuk, DeCandia, Beach, and Berman, 2014). Homeless youth
are at a greater risk of being infected with HIV because they are more likely
to engage in risky behaviors such as inconsistent use of condoms, multiple
sexual partners, and engaging in survival sex (Barman Adhikari, Begun, Rice,
Yoshioka Maxwell, and Perez-Portillo, 2016).
When looking at homeless youth who access drop-in centers in Los
Angeles, one in ten were HIV positive, compared to 0.3% rate of their housed
counterparts. Additionally, about 40% of these youth do not know their status
because they do not participate in regular testing, meaning it is very likely that
the number of infected youth is substantially higher. Due to these findings a
group of social workers are working to create a “Smart Kiosk" to address this
public health epidemic.

Our Innovation: “Smart Kiosks”

The “Smart Kiosks” would allow youth to take their sexual health into their
own hands. These kiosks provide free rapid HIV tests, condoms, and feminine
hygiene products. Feminine hygiene products were included because the team
discovered that one out three people experiencing homelessness in California

136

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 Aida Rahmattalabi et al. 137

are women and thus wanted to tailor their innovation to that increasing
population (Evans, 2016).
The locations of the “Smart Kiosks” are very important in providing youth
experiencing homelessness with the power to access these services in their own
terms. For this purpose, our first step was to locate an area or city where youth
experiencing homelessness frequented. We were able to narrow down our
search to Venice, California. Our goal is to identify a specific location that is
safe and accessible for youth experiencing homelessness. Due to limited fund-
ing and that these kiosks are stationary, it is important we find the best location
before actual deployment. This problem led to the collaboration with AI.
We model the problem of identifying the “Smart Kiosks” locations as a
variant of the well-known k-center problem. Given a metric space d, the
k-center problem finds, among n locations, the best k locations to place
facilities that are as close as possible to the rest of the points.
Our problem is also very similar to the k-center problem except that we are
faced with several other requirements. First, an important constraint is that not
all places in the area are eligible for placing the kiosks. The only eligible places
are those that are city-owned, with access to utility. Secondly, the population
distribution is a key factor in deciding the right places for the kiosks. A good
estimation of the this distribution will help find locations that are accessible
to larger population. Therefore, we will leverage data on where these youth
are most likely to reside and we will estimate the distribution and density of
homeless youth in the area. Finally, we would like to consider other existing
centers that provide the same resources to these youth. This is particularly
important because it is undesirable to place a kiosk near an existing service
provider, which results in that kiosk not being utilized and not reaching the
youth who might need these resources the most.

Problem Definition
In the following sections, we will first give a concrete description of the
problem, then we discuss the required parameters, and how we use existing
data to estimate these parameters. Next, we present the algorithm to solve the
problem. Finally, we will conclude with experiments and results.

Weighted K-Center Problem

The k-center problem is a well-studied problem within operations research
and computer science, which has different variants (Dyer and Frieze, 1985;

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 138 Sexual Health Necessities for Homeless Youth

Kariv and Hakimi, 1979; Hochbaum and Shmoys, 1985; Plesnik, 1987). Also
known as facility location problem, the problem in general is how one should
select locations for new facilities or centers – “Smart Kiosks” in this work –
to minimize a measure of distance to those locations, under some specific
constraints.
A variant of this problem, which is particularly relevant to this work, is the
weighted k-center, where the weights are used to represent the homeless youth
population estimate, used as a proxy to measure demand. In this problem, the
youth experiencing homelessness are not spread evenly in the area, and as a
result, different sectors have different demands. Clearly, highly populated areas
have higher weights as compared to sparsely populated areas.
Given a network G(V, E, ω), where V is the set of points, E is the set of
weighted edges (a measure of cost between nodes) that form a complete metric
space d, and ω is the set of all the nodes’ weights, the weighted k-center
problem is defined as selecting a subset S of the nodes in the graph, S ⊂ V, to
minimize the maximum distance from all kiosks or shelters to the population of
the homeless youth. The weights ω are the population density at each location.
In order to formalize all the above assumptions, let V  be the set of all
eligible points to place these “Smart Kiosks.” Also, let V  be the set of
existing centers and kiosks. The objective of the optimization problem can
be formulated as:

Minimize Max Min [ω(v) ∗ d(v, r)] (8.1)

K=(V  ∪V”), |V  |=k v∈V r∈K

where d(v, r) ∈ E is the cost associated with traveling from node v to node r.
This objective summarizes as: For an arbitrary point with population density
ω(v), Min [w(v) ∗ d(v, r)] yields the kiosk r with minimum weighted distance
r∈K
from point v. Then for all points in V, the maximum distance to the closest
kiosk will be minimized in order to find the best set of K kiosks.

Datasets
We use a dataset containing information of 118 youth experiencing homeless-
ness in Venice, California. The data set includes survey data that asks the
participants to indicate where they slept last night, where they spend time
with friends, and their favorite locations. The dataset was collected in 2012.
We also use a dataset containing information of existing services, including
their locations and the type of services they provide, e.g., HIV testing, needle
exchange, mental health service, drop-in center. The regions in Venice that are
eligible to put kiosks are also collected.

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 Aida Rahmattalabi et al. 139

Figure 8.1a Homeless youth geographic data, Venice area

Figure 8.1a shows the reported locations for where the youth spent last
night and where they usually spend time with friends. Figures 8.1b and 8.1c
show the locations by type. As seen in the figures, Venice beach region is an
attractive place for this population, which confirms the necessity and potential
for installing the “Smart Kiosks.”

Area Division
As our target area, we only consider Venice in Los Angeles, CA, as shown in
Figure 8.2a. This area is about 3 square miles with a fairly large population of
homeless youth. This area is divided into a grid with cells of size 100 m × 100
m, as shown in Figure 8.2b. Further discussion will be based on this grid.

Model Parameters
We modeled this problem using the notion of points, distances, and weights.
In this section, we explain in more details how to collect such data and what
we mean by abstract points.

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 140 Sexual Health Necessities for Homeless Youth

Figure 8.1b Homeless youth sleeping locations, Venice area

Points V
As discussed before, points in V can be the centers of the 100 m × 100 m cells
in the grid of Venice area. The eligible cells to install the kiosks can only be
the cells that are owned by the city and have accessibility to utility. Therefore,
we need to specify such cells beforehand. Also, note that the optimization is
performed to maximize access for the entire region, and this restriction only
applies to the candidate cells to place the kiosks.

Existing Centers V 
We are also interested in finding solutions that work best in the presence of
other competing centers. If we fail to do so, chances are the proposed solutions
is near an existing center and therefore will not be utilized as anticipated, which
is indeed a waste of resources. In this problem, we assume there is a set of
nodes that contain existing centers (denoted by V  ), and |V  |= m is the number
of exiting centers.

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 Aida Rahmattalabi et al. 141

Figure 8.1c Homeless youth hangout locations, Venice area

Distances d(v, r)
For d, the distance between pairs of points in V, Euclidean distance, Manhattan
distance, or walking distance can be used. While Euclidean distance and
Manhattan distance can be calculated based on coordinates of the points, cal-
culating walking distance is more challenging. Thus, external online services,
e.g., Google Distance API, must be used to achieve this. In this work, we
primarily use the Manhattan distance as it is provides a good proxy to the actual
traveling distance. Given two points with (x, y) coordinates as p1 = (x1 , y1 ) and
p2 = (x2 , y2 ), Manhattan distance (denoted by d) is calculated by:

d(p1 , p2 ) = |x1 − x2 |+|y1 − y2|. (8.2)

Weights W(v)
The point’s weights, or the normalized population density of homeless youth
residing in each cell, can be determined from existing survey data. This part is
explained in more detail in the “Population Estimation” subsection.

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 142 Sexual Health Necessities for Homeless Youth

Figure 8.2a Target area: Venice, Los Angeles, CA

Proposed Approach
The k-center problem is shown to be NP-hard, and therefore it is hard to find
the optimal solution for equation 8.1, and approximation algorithms are often
used to find near-optimal solutions.
Approximation algorithms for the basic k-center problem are well studied
in Hochbaum and Shmoys (1985), Gonzalez (1985), and Plesnik (1987). These
schemes present a natural approximation factor of 2. (This means that the

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 Aida Rahmattalabi et al. 143

Figure 8.2b Venice area, divided into a grid of 100m × 100m cells

solution of the approximate algorithm cannot perform worse than 2 times the
value of the optimal solution.) Several approximation algorithms are known for
the interesting generalizations of the basic k-center problem, including ones
that add costs or weights (Hochbaum and Shmoys, 1986; Plesnik, 1987). From
the literature, the weighted k-center problems in Dyer and Frieze (1985) and
Plesnik (1987) are closest to the present work, as they consider weights on
points in their models. In particular, we develop an algorithm similar to the
one proposed in Dyer and Frieze (1985), who present a greedy-based approach

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 144 Sexual Health Necessities for Homeless Youth

Algorithm 8.1 Greedy Algorithm for weighted k-center problem

1: i←1
2: Set D(v) = min{ω(v) ∗ d(v, v j )}, ∀v ∈ V ∀ j ∈ M
3: repeat
4: Determine vi+1 by D(vi+1 ) = maxv∈V D(v)
5: Set D(v) = min{ω(v) ∗ d(v, vi+1 ), D(v)}, ∀v ∈ V
6: i=i+1
7: until (i < K)

shown to provide approximate solutions no worse than min(3, 1+α) factor of

the optimal solution.
Algorithm 8.1 is also a Greedy-based approach similar to the one presented
in Dyer and Frieze (1985). However, unlike Dyer and Frieze (1985) that starts
by selecting the first point (kiosk) with the maximum weight, here we choose
the one with maximum weighted distance from the existing kiosks. And this
procedure is continued until the desired number of points are selected.
As a result, the approximation guarantee of the original algorithm must
be studied to make sure it is still valid and extends to our problem. To this
end, in the following section, we discuss how we can show that the same
guarantee holds.

Approximation Guarantees of the Greedy Approach

In this section, we prove how the approximation guarantee for the weighted
k-center problem can also be extended for the k-center variant proposed in this
work, which assumes m centers are already chosen. The difference between
the two Greedy approaches is that in Algorithm 8.1, we skip the first step of
selecting the first node with maximum weight. However, we can still show
that the same approximation guarantee holds, and to do so we simply assume
that the already chosen nodes (existing centers) have very high weight values.
This is a legit assumption, since these nodes are already chosen, and it will
not affect the final solution to the problem. In fact, if we run the original
Greedy algorithm on the new problem, with modified weights of the nodes
associated with the existing centers and assuming no existing centers, the
algorithm chooses those centers as the first m centers, and from that point on,
it will select the additional points that are in fact the solutions to our problem.
Therefore, we showed that the solution of Algorithm 8.1 for weighted k-center
with existing centers is the same as the solution of the algorithm in Dyer and

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 Aida Rahmattalabi et al. 145

Frieze (1985) for the basic weighted k-center problem. Based on the argument
made, we see that the Greedy approach provides the same approximation of
min{3, 1 + α}, where α is the maximum ratio of weights between the points
in V. Based on the assumption made, α also takes a high value, and therefore,
factor of 3 is the approximation guarantee for the weighted k-center problem
with existing centers.

Population Estimation
Since the objective is to provide a better access to these kiosks, it is important
to factor in the distribution of homeless youth in the area. Therefore, estimating
the number of homeless youth in a given cell is a crucial part of this research.
This problem is considered as density estimation problem (Silverman, 1986).
Let z be a discrete random variable, where z = (X, Y) indicates that a person
resides at cell (X, Y). More formally, the probability mass function Rao (2009):

fz (x, y) = Pr(z = (x, y))

gives the probability that a person is at cell (x, y). Density estimation is the
process of estimating the function fz , given the observed data of homeless
youth locations.
In this study, we consider three different density estimations: histogram
based distribution estimation, mixture of Gaussian estimation, and finally we
compare with a uniform distribution estimation that assumes the population
is uniformly distributed over the area. In other words, in uniform distribution
estimation, cells in the region have the same probability of a homeless youth
being present,
1
fz (x, y) = ,
N
with N being the total number of cells. In this approach, the estimated number
of homeless youth in each cell is the same and it leads the optimization
to only consider the distances between cells and kiosks. This estimation
approach can be used for the regions that have homeless youth distributed
approximately uniformly over cells. In the second approach, histogram based
distribution estimation, the counts of homeless youth in cells are used to
estimate the density function fz (x, y). Let count1 , count2 , . . . , countN denote the
number of homeless youth in cells (x1 , y1 ), (x2 , y2 ) . . . , (xN , yN ), respectively.
The probability that a person stays at cell xi , yi , 1 ≤ i ≤ N is:
counti
fz (xi , yi ) = N .
i=1 counti

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 146 Sexual Health Necessities for Homeless Youth

This approach is better than uniform distribution estimation when all the
counts are collected. But in reality, there is a lot of missing data. Figure
8.3a shows the counts of homeless youth over cells. As we can see in this
figure, there are areas that have high number of homeless youth and also
a lot of missing data. In this approach, for cell xi , yi in which the count of
number of homeless youth is missing, we assume that count is 0 and therefore
fz (xi , yi ) = 0. However, imputing the missing data is one critical problem.
There are a lot of studies on this problem: temporal prediction (Shi et al., 2016),
which exploits the temporal relationship of value at one location at multiple
times; spatial interpolation (Lam, 1983), which exploits the relation between
the values at nearby locations; and others. With the mixture of Gaussian
estimation approach, we try to estimate the missing information by spatial
interpolation of the observed data. We assume the distribution of homeless

Figure 8.3a Homeless population distribution: Counts of homeless youth

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 Aida Rahmattalabi et al. 147

youth is a mixture of Gaussian so that the homeless youth are distributed

smoothly over cells. The probability that a person is present in the cell xi , yi
is a linear combination of q two dimensional Gaussian distribution density
functions,
 
q
exp( 12 (x − μ)T −1 (x − μ))
fz (x , y ) =
i i
wi   ,
i=1 (2π)2 | |


with qi wi = 1 where x is a two-dimensional real-valued column vector

representing the coordinate of a cell, | | is the determinant of the covariance

matrix , μ is the vector mean, and wi is the coefficient determining the weight
of component i. With only one Gaussian distribution, the mean value gets the
highest probability. Therefore, we choose the q cells with highest counts to be
the means of the k components. Figure 8.3b shows the heat map of estimated

Figure 8.3b Homeless population distribution: Estimated mixture of Gaussian

distributions

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 148 Sexual Health Necessities for Homeless Youth

homeless youth distribution with q = 50. The means of the Gaussians are
locations
 withthe top 50 counts of youth and the covariance matrix is equal to:
 30 5
= . The weights for the components are set proportionally to the
5 30
number of homeless youth at the center of each component. As we can see in
this figure, the missing data is imputed smoothly over the entire area.

Results
To demonstrate the effectiveness of our algorithm, we compare it with a
baseline algorithm. The baseline algorithm selects nodes consecutively in the
most populated cells. Figure 8.4 compares the maximum weighted distance to
centers for three different population distributions and the case of installing
two new centers (K = 2). As indicated, the Greedy algorithm outperforms the
baseline in all three cases with a maximum improvement of 40%. Also, Figure
8.5 shows the output of the Greedy algorithm on the map of the Venice area. In
the discussion section, we explain how these suggested spots are reasonable.

Greedy Baseline
1

0.75

0.5

0.25

0
Uniform Distribution Probability Histogram Gaussian Distribution

Figure 8.4 Normalized maximum distance to centers for different population

distributions for the case of selecting two centers. In all three cases, the Greedy
algorithm outperforms the baseline algorithm.

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 Aida Rahmattalabi et al. 149

Figure 8.5 The two candidate locations for placing kiosks under Gaussian
population distribution assumption.

In another experiment, we looked at how the algorithm scales with the

number of centers. Figures 8.6–8.8 indicate the results for the three population
distribution models when the number of centers is increased from K = 1 to
K = 5. As seen, the Greedy algorithm still performs significantly better than
the baseline. Interestingly, with Gaussian distribution model, as we increase
the number of centers, the marginal improvement from the Greedy algorithm
also increases. This indicates that with more information about the population,
it is increasingly more difficult to perform well, and more careful optimization
is required.

Discussion and Conclusions

In this study, we specifically focused on placing Smart Kiosks in the Venice
area in Los Angeles, CA. We viewed this problem as an optimization task
where we aimed to minimize the maximum traveling distance to the Smart
Kiosks. This problem is very important to our target population, as it can

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 150 Sexual Health Necessities for Homeless Youth

Greedy Baseline
1

0.75

0.5

0.25

0
k=1 k=2 k=4 k=5

Figure 8.6 Normalized maximum distance to centers for uniform distribution for
different number of centers.
Greedy Baseline
1

0.75

0.5

0.25

0
k=1 k=2 k=4 k=5

Figure 8.7 Normalized maximum distance to centers for probability histogram

distribution for different number of centers.

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 Aida Rahmattalabi et al. 151

Greedy Baseline
1

0.75

0.5

0.25

0
k=1 k=2 k=4 k=5

Figure 8.8 Normalized maximum distance to centers for Gaussian distribution for
different number of centers.

greatly impact the usability of the kiosks for such a transient and vulnerable
group. We formulated this problem using a weighted variant of the well-known
k-center problem and we used a Greedy-based approach to choose the kiosk
locations. We compared the results of the Greedy algorithm to a baseline
algorithm, and we observed that the Greedy approach outperforms the baseline
(up to 40% improvement). As an example, in the case of placing two new
kiosks, the Greedy algorithm suggested locations that are in central areas of
Venice. One is a location on the Main Street, a central street in Venice located
in an area that would capture foot traffic of the youth who go between seeking
services more inland and going back to the Venice Beach Boardwalk and skate
park. By examining the proposed locations, we can observe that the algorithm
has not only identified areas that are viable through using data, but it is also
selecting areas that are meaningful and understandable when learning more
about the patterns of youth experiencing homelessness. Therefore, by the help
of artificial intelligence, we were able to provide places to deploy these kiosks
that not only make it more accessible to the youth but also have the potential
to reach and serve a greater number of youth experiencing homelessness.
Through this partnership between computer science and social work we have

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 152 Sexual Health Necessities for Homeless Youth

shown that complex social problems can be addressed with innovative ideas
that are created through the dual knowledge of social work and computer
science. There will need to be further work to assess if such technology is
appropriate and if it would have an impact on the HIV/AIDS epidemic. We are
hopeful to see this innovation continue with the continued support of social
work and computer science experts.

References
Barman Adhikari A., Begun S., Rice E., Yoshioka Maxwell A., and Perez-Portillo A.
2016. Sociometric network structure and its association with methamphetamine
use norms among homeless youth. Social Science Research, 58, 292–308.
Bassuk E. L., DeCandia C. J., Beach C. A., and Berman F. 2014. America’s
youngest outcasts: A report card on child homelessness. The National
Center on Family Homelessness at American Institutes for Research. https://
www.air.org/sites/default/files/downloads/report/Americas-Youngest-Outcasts-
Child-Homelessness-Nov2014.pdf.
Dyer, M. E, and Frieze, A. M. 1985. A simple heuristic for the p-centre problem.
Operations Research Letters, 3(6), 285–288.
Evans, Erica. 2016. L. A. County will try to house 100 homeless youths in 100 days.
Los Angeles Times, July 26.
Gonzalez, T. F. 1985. Clustering to minimize the maximum intercluster distance.
Theoretical Computer Science, 38, 293–306.
Hochbaum, D. S., and Shmoys, D. B. 1985. A best possible heuristic for the k-center
problem. Mathematics of Operations Research, 10(2), 180–184.
Hochbaum, D. S., and Shmoys, D. B. 1986. A unified approach to approximation
algorithms for bottleneck problems. Journal of the ACM (JACM), 33(3), 533–550.
Kariv, O., and Hakimi, S. L. 1979. An algorithmic approach to network location
problems. I: The p-centers. SIAM Journal on Applied Mathematics, 37(3), 513–
538.
Lam, N. S.-N. 1983. Spatial interpolation methods: A review. The American Cartogra-
pher, 10(2), 129–150.
Plesnik, J. 1987. A heuristic for the p-center problems in graphs. Discrete Applied
Mathematics, 17(3), 263–268.
Rao, S. S. 2009. Engineering Optimization: Theory and Practice, 4th ed. Hoboken, NJ:
John Wiley and Sons.
Shi, W., Zhu, Y., Yu, P. S., Huang, T., Wang, C., Mao, Y., and Chen, Y. 2016. Temporal
dynamic matrix factorization for missing data prediction in large scale coevolving
time series. IEEE Access, 4, 6719–6732.
Silverman, B. W. 1986. Density Estimation for Statistics and Data Analysis. London:
Chapman & Hall.

https://doi.org/10.1017/9781108669016.010 Published online by Cambridge University Press

 9
Know-Stress
Predictive Modeling of Stress among Diabetes Patients
under Varying Conditions

Subhasree Sengupta, Kexin Yu, and Behnam Zahiri

Introduction
Around 422 million people worldwide had diabetes mellitus in 2014 (World
Health Organization, 2016). The global diabetes prevalence rate has increased
from 4.7% in 1980 to 8.5% now (World Health Organization, 2016). Type
2 diabetes (T2D) diminishes patients’ quality of life. Diabetes patients are
more likely to suffer from visual and hearing impairments, overweight issues,
osteoarthritis, stroke, and possess two- to threefold increased risk of physical
disability than people not suffering from diabetes (Gregg et al., 2000; Kirkman
et al., 2012; Schwartz et al., 2002; Volpato et al., 2002; Volpato, Leveille,
Blaum, Fried, & Guralnik, 2005).
T2D has been widely recognized as a chronic disease caused by unhealthy
lifestyle. However, in recent years, psychological factors were brought to
researchers’ attention because physical factors could not fully explain the
development and progression of T2D. Compared to nondiabetic individuals,
diabetes patients experience chronic allostatic load, which manifests as
impaired post-stress recovery and blunted stress reactivity (Carvalho
et al., 2015; Steptoe et al., 2014).
The prevalence of diabetes is higher among low-income groups and ethnic
minorities (Lanting, Joung, Mackenbach, Lamberts, & Bootsma, 2005). For-
mer research presents the trend that stress level increases as the socioeconomic
status decreases (Orpana, Lemyre, & Kelly, 2007). Occupational and financial
stressors, frequently experienced by low-income groups, have been shown
to be significantly associated with the development and progression of T2D
(Morikawa et al., 2005). Social stressors – for example, social rejection – have
also been indicated to be correlated with physiologic stress response (Slavich,
O’Donovan, Epel, & Kemeny, 2010). Considerable amount of empirical

153

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 154 Know-Stress

evidence implies the importance of delivering stress relief intervention for low-
income T2D patients.
In existing studies, social scientists employ statistical analyses to identify
the association between stress and the development of T2D on a group level
(Kelly & Ismail, 2015). Machine learning algorithms to make prediction of
stress levels in individual cases becomes possible.
The purpose of this chapter is to predict stress levels of low-income T2D
patients using machine learning–based classification techniques, aspiring to
build models that can be utilized to help identify vulnerable individuals and
thus provide timely interventions.

Data Description
A Brief Background of the Data Set Used
This project employs secondary data from the Diabetes–Depression Care-
management Adoption Trial (DCAT). The purpose of the DCAT program was
to reduce health disparities of low-income population by optimizing depression
screening, treatment, follow-up, outcomes, and cost saving using technology
(Wu et al., 2014). In collaboration with Los Angeles County Department of
Health Services (DHS) Ambulatory Care Network (ACN), DCAT intervention
was implemented from 2010 to 2013 among 1,400 low-income diabetes
patients in Los Angeles County (Wu et al., 2014). The participants were pre-
dominantly Hispanic/Latino Immigrants (89%). The DCAT program collected
baseline and 6-, 12-, and 18-month follow-up data (Wu et al., 2014). This
current study analyzes only the baseline data.

Description of Variables Presented in the Dataset

The DCAT baseline data was collected through surveys, interviews, and
medical records. Interviews provide socio-demographic information (e.g., age,
gender, ethnicity) as well as self-reported health information. Depression
symptoms are assessed with Patient Health Questionnaire-9 (PHQ-9) and
Symptom Checklist-20 (SCL-20). Dysthymia is measured using two standard
questions from Diagnostic and Statistical Manual of Mental Disorders (DSM)-
IV. Functional impairment was measured using Sheehan Disability Scale.
Medical Outcomes Study Short-Form Health Survey (MOS SF-12), and Physi-
cal and Mental component summaries (PCS, MCS) were used to assess Health-
related quality of life. Pain impact was measured using SF-12. Anxiety was
measured with Brief Symptom Inventory (BSI). Diabetes distress is measured

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 Subhasree Sengupta, Kexin Yu, and Behnam Zahiri 155

Table 9.1. Description of the variables in the data set

Measurement Scales Used/Description

Variables Category of the Variables

Social demographics Age, gender, marital status, racial/ethnic group,

birthplace, time living in the United States,
language, and education
Depression PHQ-9, SCL-20
Dysthymia Two standard question from DSM-IV
Functional impairment Sheehan Disability Scale
Health-related quality of life Medical outcome study short-form health survey
(MOS SF-12), physical and mental component
summaries (PCS, MCS)
Chronic pain Pain present most of the items for 6 months or
longer
Pain impact Short Form-12
Anxiety Brief symptom inventory (BSI)
Patient satisfaction with diabetes A 5-point Likert scale from “very dissatisfied” to
care and emotional care “very satisfied.”
Diabetes symptoms Whitty 9-item questionnaire
Diabetes distress Diabetes distress scale
Body Mass Index Self-reported weight and height
Medical history Including hospitalization, clinic and emergency
department visits, pharmacy pick-up, and
International Statistical Classification of Diseases
and Related Health Problems (ICD)-9 codes for
diagnostic diseases, as well as the lab results of
glycated hemoglobin (A1C), micro-albumin, and a
lipid panel during the same period.
Health service utilization Previous outpatient (medical and mental health),
inpatient, and emergency room service utilization.
Administrative cost Using data from DHS outpatient and pharmacy
billing records, electronic emergency room and
hospitalization billing cost, and mental health care
use data from medical records.

using a two-item diabetes distress scale. Diabetes symptoms are assessed

using Whitty-9 item questionnaire. The DCAT research team also acquired
medical record data that included medical history, health service utilization,
and administrative costs. All these variables were used as predictors of stress
in our data set. Table 9.1 presents a summary of these predictor variables.
Socioeconomic stress is the outcome this project aims to predict, which
is measured with 12 items that describe a list of stressful life issues, including

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 156 Know-Stress

work problems, unemployment problems, financial problems, marital conflicts,

family conflict, parenting concerns, caregiving problems, cultural conflicts,
legal problems, immigration problems, serious illness or loss of a close person,
and community violence concerns. The participants give answers using a scale
from 0 to 10, which represents stress level ranging from “no stress at all”
to “the most stress you can imagine.” In subsequent sections we provide a
description of how to convert the 12-item measurement into a single variable
using a technique called factor analysis.

Aim of Our Analysis

We address three fundamental objectives in our analysis:
1. Predicted variable creation: As mentioned, our desired outcome variable is
expressed as a combination of 12 separate outcome variables. We present
an empirical process of aggregating these into a single outcome variable.
2. Determining optimal cutoff: We want to predict if the stress level of the
patient falls into the high stress level category or into the low stress level
category. Thus, since we have only two possible categories, we want our
stress predictor variable to be a binary variable (‘1’ – high stress and ‘0’ –
low stress). However, after the initial aggregation process as mentioned in
(1), the final aggregated variable is a continuous variable whereas we
needed a categorical variable. Thus, we also present a mechanism creating
this binary classification scheme from our aggregated predictor variable by
determining the most optimal “cutoff score,” i.e. for any value above the
cutoff we will consider the level of the associated patient to fall into the
high stress category, and vice versa for any value below the cutoff.
3. Feature selection: Our data set has a large number of predictor variables,
so collecting this information might need a lot of time and effort;
furthermore, it may not always be possible to have all this information.
Hence the most important goal for us was to identify a small subset of
predictor variables without affecting the overall performance of the model
as compared to the case when all the predictor variables are present. To
achieve this, we experiment with different data sets formed by grouping
our predictor variables into four categories. Then we create a dataset for
each of these four categories having predictor variables pertaining to only
that category to predict socioeconomic stress. We also conduct a feature
selection experiment using a machine learning algorithm called Extremely
Randomized Trees (Geurts, Ernst, & Wehenkel, 2006), which we will
briefly explain in subsequent sections.

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 Subhasree Sengupta, Kexin Yu, and Behnam Zahiri 157

More Detailed Description of the Aims of the Project

Predicted Variable Creation
In the data set, there are 12 indicators of socioeconomic stress. Since we want
a single indicator variable for stress, we create a single variable for indicating
stress level using the following aggregation technique. First, we conduct factor
analysis (Thompson, 2004) to find latent factors in the 12 variables. We
found that there was only a single dominant latent factor and hence we could
aggregate these 12 variables into a single variable. Using this knowledge, we
know that a single variable is sufficient to represent the combined effect of
these 12 variables. We compute the single variable as the sum of the 12 stress
variables to create the single continuous variable, which we shall refer to as
“stress_sum.” Thus, we use the “stress_sum” value to express the combined
effect of the 12 different stress indicator variables.

Determining Optimal Cutoff

As mentioned earlier, we aim to be able to distinguish high–stressed patients
from low–stressed patients. Thus, we want to create a binary classification
scheme to divide our continuous “stress_sum” variable into the two required
categories. The way we achieve this is by creating several cutoff values. All
the individuals whose stress level was below the cutoff value were assigned to
the low–stress category, and the values above were assigned to the high–stress
category.
To find the optimal cutoff, we want to determine the cutoff value that
performs best among our experimental models. To choose the cutoff values,
we first sort the “stress_sum” value in ascending order for all records. Next,
we choose the values at 20th percentile, 25th percentile, 33rd percentile, and
finally 50th percentile from the aforementioned ordering. The corresponding
values were 15, 19, 25, and 39, respectively. We did not choose cutoff values
beyond this point, as doing so would lead to severe class imbalance (the
proportion of low-stress patients would significantly outnumber the proportion
of high-stress patients) thus affecting the prediction ability of our machine
learning models.

Feature Selection and Various Data Sets

To choose the ideal set of features (predictor variables), we started by selecting
different subsets of the entire data set.

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 158 Know-Stress

1. Cleansed data set: One major drawback of our data set is that there are
several missing values and so the first data set we made was by removing
all these missing values. In total we had 130 features and 1,340 patient
records in this cleansed data set. We removed about 35% of the data. This
data set acts as the baseline against which we compare the performances of
the data sets described below.
2. Data sets based on categories of predictor variables: There can be several
ways in which the features or predictor variables can be divided into
categories. However, we divided the data set into four major categories,
namely demographic, biological (health-related variables like weight,
other ailments, etc.), psychological, and social variables. The intuition
behind this classification of predictor variables is the fact that sometimes,
due to financial or time constraints, it might not be feasible to run the
entire study to collect all the information we had in our data set. Hence, if
we can have access to only some categories of information, can we still get
a model with comparable performance to model using the data set in (1)?
3. We used a machine learning model called extremely randomized trees,
which is a tree-based ensemble method for learning the Top 25 features.
We chose 25 as per consultation with domain experts. The goal was to
compare the predictor variables selected here using this method with those
in the data sets in (2) and to get an intuitive reason behind the features
selected by the machine learning model.

Experimental Setup
For each of the data sets mentioned on the preceding list, we randomly divide
the data into 70% for training and 30% for testing. Training implies the we
use 70% of the data in each data set to learn the machine learning models and
then evaluate the learned models using the test data sets. The metrics used for
evaluation are as described in the metrics section that follows.

Models of Classification
Our problem is a classification task (Caruana & Niculescu-Mizil, 2006): i.e.,
using the machine learning algorithms described in this section, we want to
classify each patient in the data sets into either the high-stress or the low-stress
category.
An important term used in the context of classification is decision bound-
ary – i.e., a separation point. The idea of the decision boundary is basically that

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 Subhasree Sengupta, Kexin Yu, and Behnam Zahiri 159

it serves as a threshold such that any value above the threshold is classified as
“True” (or high stress in our application) and anything below is classified as
“False” (or low stress in our application). This separation can be expressed as a
line or hyperplane in case of linear decision boundaries, but often a line cannot
serve as an efficient separator. So, depending on the distribution of the data,
we might need to revert to nonlinear separators.
Given this, we use three widely used machine learning models of classifi-
cation. Their popularity stems from several reasons, such as efficient software
that enables the model to scale very well with the increasing size of the data set;
interpretability – ease of understanding what the model is accomplishing; etc.
Very briefly, the mathematical intuition behind classification is expressed as
follows:

1. Let us assume we have a vector <P> of predictors, where P = < p1 , p2 , . . . >.

2. We want to learn a function f(w,p), a weighted combination of P, such that
the error between the computed functional value and the true value in the
training data set is minimized.
3. Then we test and evaluate this computed function using the test data set.

Neural Networks: Neural networks have recently gained widespread pop-

ularity in health care, as reported in Lisboa & Taktak (2006). Neural net-
works have demonstrated to be very useful for classification, since they can
approximate both linear and nonlinear decision boundaries very well, making
the neural networks very powerful and efficient. Further, with the growing
size of data sets, neural network–based classification can be very effective,
and on this basis has gained tremendous popularity as a standard image
classification/speech recognition tool.
The inspiration behind this method is the neural architecture found in the
human brain. Like the human brain, the different layers act as multiple layers
of processing or functional transformations (also called hidden layers) on the
initial input before the final output layer is reached.
Briefly the working of a neural network can be explained as follows:

1. Each layer consists of a set of neurons that serve the input for that layer.
2. We then take a weighted combination of the inputs coming through these
neurons.
3. Some form of activation function is applied.
4. Now we take the output of (3) to become inputs of the next layer, and this
follows until the output layer.

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 160 Know-Stress

5. Finally, the output obtained from the final layer is compared against the
actual output; the error is computed and the weights all the way back to the
input layer are adjusted. This process is also called back propagation.
6. This entire process can be repeated many times to better train the model –
each full cycle (steps 1–5 are called an epoch.)

The neural network model we use has the following characteristics:

1. Neurons in the input layer equal the number of predictor variables.

2. Two hidden layers, with 20 neurons per hidden layer.
3. Two neurons in the output layer, since we have two classes.
4. We use the “Relu” activation function for connecting the input layer to
the first hidden layer and for connecting the first hidden layer to the
second one.
5. The final connection from the second hidden layer to the output layer will
use a SoftMax activation function, since this is a multi-class
classification case.

We use the keras wrapper for implementing the architecture of the neural
network with Tensor flow as the backend for training the weights.

Decision Tree
Using a decision tree–based aid has gained a lot of importance in mining
of health care data (Koh & Tan, 2011). The ease of interpretability of this
method makes it a popular choice among machine learning methods. For a
concrete understanding of how a decision tree is constructed, please refer to
Quinlan (1987).
Aside from adding validity to the results obtained by the neural network,
another purpose of the decision tree method is to get an idea about which
predictor variables significantly affect the predictive capability of the learned
model, thereby giving us an intuition into which features we necessarily need
to keep when selecting the top 25 features from the dataset.

Logistic Regression
Logistic regression is the common technique used in a multi-class classifica-
tion used when the outcome variable is categorical (binary in this application).
Thus, this technique will serve more as a baseline or benchmark for us to
compare with the machine learning classifiers as described earlier.

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 Subhasree Sengupta, Kexin Yu, and Behnam Zahiri 161

Feature Selection
The top 25 features were selected by a machine learning model called
extremely randomized trees (Geurts, Ernst, & Wehenkel, 2006). It is a tree-
based ensemble method. An ensemble is a collection of many decision trees.
Besides accuracy, the main strength of the resulting algorithm is computational
efficiency. This technique gives us the importance value for each feature
(predictor) in obtaining the output. We then sort the features by this value in
descending order and select the top 25 features.

Metrics of Evaluation
Some important terminology:

1. True positives (TP): Those patients in the test set who had “high” stress
and were predicted to have high stress.
2. False positives (FP): Those patients in the test set who had “low” stress
and were predicted to have high stress.
3. True negatives (TN): Those patients in the test set who had “low” stress
and were predicted to have low stress.
4. False negatives (FN): Those patients in the test set who had “high” stress
and were predicted to have low stress.

The metrics used are:

1. Accuracy: (Σ TP + ΣTN)/(Σ TP + ΣTN + ΣFN + ΣFP)

2. Recall: Σ TP/(Σ TP + Σ FN) {Over all the actual “high” samples how
many “high” samples did the model correctly “recall”}
3. Precision: ΣTP/(Σ TP + Σ FP) {Over all the “high” samples the model
identifid how many were actually “high”}
4. Confusion matrix A matrix representation of TP, FP, TN, FN.
This tabular representation provides more of a visual aid to understand the
effectiveness of our models. For example, here we find that the rate of true

Table 9.2. A sample confusion matrix

Predicted Low Predicted High

Actual Low 189 (TN) 10 (FP)

Actual High 5 (FN) 198 (TP)

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 162 Know-Stress

positives and true negatives is high, which means we have learned a very
good model.
5. AUC: Area under the ROC (Receiver Operating Characteristic).
AUC value gives the area under the curve (where the x-axis is the false
positive rate and the y-axis is the true positive rate), and since we want to
increase the true positive rate, we want a higher AUC value. That indicates
that the larger the area under the plotted curve, the better our model’s
performance will be.

Experimental Results and Discussion

A total of 84 different models were experimented on. Here we present and
discuss the results obtained from the neural network experiments. We also
discuss the results from performing feature selection as described in the
previous section. The results and trends obtained from the decision tree and
logistic regression are very close numerically to the results of the neural

Table 9.3. Results for the cleansed data set

Accuracy Precision Recall Confusion Matrix AUC

Cutoff = 15 0.72 0.718 0.67 171 49 0.791

61 125
Cutoff = 19 0.736 0.713 0.8 135 66 0.813
41 164
Cutoff = 25 0.682 0.607 0.842 122 100 0.792
29 155
Cutoff = 39 0.73 0.7318 0.682 166 48 0.801
61 131

Table 9.4. Results for data set with only demographic info

Accuracy Precision Recall Confusion Matrix AUC

Cutoff = 15 0.685 0.708 0.6108 160 51 0.764

79 124
Cutoff = 19 0.714 0.583 0.608 209 62 0.774
56 87
Cutoff = 25 0.705 0.458 0.541 232 71 0.699
51 60
Cutoff = 39 0.874 0.5 0.08 358 4 0.77
48 4

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 Subhasree Sengupta, Kexin Yu, and Behnam Zahiri 163

network experiments, which means they reflect the same trends and hence have
not been explicitly presented here.
An important observation here is that we find with increasing the cutoff
value there is a significant drop in the recall score obtained. This can further

Table 9.5. Results for data set with only biological info

Accuracy Precision Recall Confusion Matrix AUC

Cutoff = 15 0.504 0.486 0.974 20 204 0.704

5 193
Cutoff = 19 0.644 0.53 0.631 168 89 0.708
56 87
Cutoff = 25 0.734 0.566 0.146 293 13 0.679
99 17
Cutoff = 39 0.817 0.222 0.255 333 42 0.701
35 12

Table 9.6. Results for data set with demographic and biological info

Accuracy Precision Recall Confusion Matrix AUC

Cutoff = 15 0.725 0.774 0.587 189 34 0.777

82 117
Cutoff = 19 0.741 0.72 0.471 241 28 0.794
81 72
Cutoff = 25 0.748 0.459 0.583 260 66 0.786
40 56
Cutoff = 39 0.843 0.345 0.387 337 36 0.794
30 19

Table 9.7. Results for data set with demographic, biological and
psychological info

Accuracy Precision Recall Confusion Matrix AUC

Cutoff = 15 0.73 0.69 0.74 164 63 0.79

51 144
Cutoff = 19 0.71 0.76 0.35 248 17 0.81
102 55
Cutoff = 25 0.77 0.73 0.395 280 18 0.779
75 49
Cutoff = 39 0.88 0.4 0.127 336 9 0.78
41 6

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 164 Know-Stress

Table 9.8. Results for data set with demographic, biological, and
psychological and social info

Accuracy Precision Recall Confusion Matrix AUC

Cutoff = 15 0.72 0.74 0.59 191 39 0.79

51 144
Cutoff = 19 0.73 0.76 0.37 253 18 0.81
95 56
Cutoff = 25 0.78 0.62 0.41 287 27 0.80
64 44
Cutoff = 39 0.89 0.66 0.04 375 1 0.85
44 2

Table 9.9. Results for feature selection

Accuracy Precision Recall Confusion Matrix AUC

Cutoff = 15 0.73 0.67 0.89 112 89 0.815

22 183
Cutoff = 19 0.7 0.77 0.51 186 28 0.76
94 98
Cutoff = 25 0.73 0.74 0.67 166 45 0.81
63 132
Cutoff = 39 0.75 0.7 0.83 144 69 0.84
32 161

be observed from the confusion matrix; the number of false negatives is

significantly higher as compared to the number of true positives. The chief
cause of this is class balance, i.e., there are many more data points labeled
as “low stress” than “high stress.” This trend can also be seen in many of the
results presented below.

Discussion on the Results

Optimal Cutoff Range
From the results presented in the previous section it appears that pretty much
all the different cutoffs give a reasonably good model performance when
the cleansed data set is used. When the experiments are performed on the
demographic or biological data sets, or combinations of these, we find that
choosing a higher cutoff value (>20) severely affects the recall score of the

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 Subhasree Sengupta, Kexin Yu, and Behnam Zahiri 165

model and the true positive rate drastically falls, which is not a desirable
outcome. Overall, taking all the metrics into consideration, based on the
results, a cutoff score of 15 seems to be perform best. In terms of a range
for the cutoff value, the range of 15–19 performs best.

Model Stability
One significant observation about the experiments described earlier in the
chapter is that when the data set is manually partitioned into the demographic,
biological, psychological, and social data sets, experimenting on these or
a combination of these data sets severely affects model performance. For
example, in the combined demographic, biological, psychological, and social
data set we find that at the cutoff of 39 the recall score drops to 0.04 and the
ratio of FP:TP is 22:1, which is not desired at all. This is a bit counterintuitive,
as based on our feature selection results where most of the selected features
are psychological, we would presume that the model would perform better.
A probable reason for this can be that some form of over-fitting is taking
place. We also find that, rather than experimenting with only demographic or
biological data, using a combination of both reflects improvements in model
performance, but adding psychological and social variables does not boost the
performance – in fact, quite the opposite.

Feature Selection
We perform feature selection using a machine learning model called extremely
randomized trees. Using this we can identify the top 25 features in the data
set. The selected features are a mixture of biological, psychological, and some
demographic features. Interestingly, we find that the features located at the top
two levels of the decision tree as applied to the demographic and biological
data sets also appear in the features selected by the machine learning algorithm
for feature selection. This helps interpret the machine learning–based feature
selection method.
We find that there is not as much fluctuation in model performance as we
have seen when trying the data sets with distinct categories created manually
(demographic, biological, psychological, and social). Thus, using this machine
learning approach would be better/more stable than the manual category–based
data-partitioning method we have presented.

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 166 Know-Stress

Briefly explaining the features selected using this method: in terms of

demographic features, we had some items such as age and gender; in terms
of biological features, we had onset age of diabetes, BMI, and measures of
any form of disability; in terms of psychological features, we had measures of
depression and anxiety; and for social factors, we had some measures related
to economic status.
We also understand that using only demographic or only biological or
only psychological or only social features may not be enough to build a
powerful and efficient classification model. However, the machine learning–
based feature learning experiment provides a guide on how to select these
features.

Conclusion
We present a rigorous machine learning-based analysis with varying data
sets. We present results based on three popular classification models (neural
networks, decision trees, logistic regression). We address the challenge of
transforming our continuous predicted outcome variable in a categorical
variable with two categories (high versus low). To achieve this, we experiment
with several “cutoff scores” to divide the values into the two categories. We
present a comparison of various cutoffs used on these different data sets and
find an optimal cutoff range to be used. We reason that the optimal cutoff
should be in the range of 15–19 based on the experiments and across all data
sets. We also conduct feature selection to select top 25 features in the data set
in view of situations where gathering all the extended list of variables in the
given data set might be infeasible and present the results. The selected features
are predominantly psychological variables, some biological variables, and a
few demographic variables. We further validate this feature selection process
by comparing it with results obtained from manual splits of the data based on
demographics, biological, social, and psychological categories. In conclusion,
our extensive analysis provides important insights and a foundation of using
computational tools from artificial intelligence to answer challenging social
work questions.

Future Work
There are several ways to extend upon the analysis techniques presented in
this chapter. An interesting direction of further exploration will be to add

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 Subhasree Sengupta, Kexin Yu, and Behnam Zahiri 167

in the layer of time series analysis to the current investigation. This can
be achieved by gathering data at different time periods and then predicting
stress using that information along with historical observations. Further, the
process of feature selection can be used for each data set representing the
four categories as previously mentioned. In that regard, more categories of
the predictor variables can be explored than those we have used. Also, more
sophisticated machine learning techniques can be used to answer the question
about tackling missing data.

References
Caruana, R., & Niculescu-Mizil, A. (2006, June). An empirical comparison of super-
vised learning algorithms. In Proceedings of the 23rd International Conference on
Machine Learning (pp. 161–168). Association for Computing Machinery.
Carvalho, L. A., Urbanova, L., Hamer, M., Hackett, R. A., Lazzarino, A. I., & Steptoe,
A. (2015). Blunted glucocorticoid and mineralocorticoid sensitivity to stress in
people with diabetes. Psychoneuroendocrinology, 51, 209–218. https://doi.org/
10.1016/j.psyneuen.2014.09.023
Geurts, P., Ernst, D., & Wehenkel, L. (2006). Extremely randomized trees. Machine
Learning, 63(1), 3–42.
Gregg, E. W., Beckles, G. L., Williamson, D. F., Leveille, S. G., Langlois, J. A.,
Engelgau, M. M., & Narayan, K. M. (2000). Diabetes and physical disability
among older US adults. Diabetes Care, 23(9), 1272–1277.
Kelly, S. J., & Ismail, M. (2015). Stress and type 2 diabetes: A review of how stress
contributes to the development of type 2 diabetes. Annual Review of Public Health,
36, 441–462. https://doi.org/10.1146/annurev-publhealth-031914-122921.
Kirkman, M. S., Briscoe, V. J., Clark, N., Florez, H., Haas, L. B., Halter, J. B., . . . Swift,
C. S. (2012). Diabetes in older adults. Diabetes Care, 35(12), 2650–2664. https://
doi.org/10.2337/dc12-1801.
Koh, H. C., & Tan, G. (2011). Data mining applications in healthcare. Journal of
Healthcare Information Management, 19(2), 65.
Lanting, L. C., Joung, I. M. A., Mackenbach, J. P., Lamberts, S. W. J., & Bootsma, A.
H. (2005). Ethnic differences in mortality, end-stage complications, and quality of
care among diabetic patients. Diabetes Care, 28(9), 2280–2288. https://doi.org/10
.2337/diacare.28.9.2280.
Lisboa, P. J., & Taktak, A. F. (2006). The use of artificial neural networks in decision
support in cancer: A systematic review. Neural Networks, 19(4), 408–415.
Morikawa, Y., Nakagawa, H., Miura, K., Soyama, Y., Ishizaki, M.,
Kido, T., . . . Nogawa, K. (2005). Shift work and the risk of diabetes mellitus among
Japanese male factory workers. Scandinavian Journal of Work, Environment &
Health, 31(3), 179–183.
Orpana, H. M., Lemyre, L., & Kelly, S. (2007). Do stressors explain the association
between income and declines in self-rated health? A longitudinal analysis of the
National Population Health Survey. International Journal of Behavioral Medicine,
14(1), 40–47.
Quinlan, J. R. (1987). Simplifying decision trees. International Journal of Man-
Machine Studies, 27(3): 221. https://doi.org/10.1016/S0020-7373(87)80053-6.

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 168 Know-Stress

Schwartz, A. V., Hillier, T. A., Sellmeyer, D. E., Resnick, H. E., Gregg, E., Ensrud, K.
E., . . . Cummings, S. R. (2002). Older women with diabetes have a higher risk of
falls. Diabetes Care, 25(10), 1749–1754.
Slavich, G. M., O’Donovan, A., Epel, E. S., & Kemeny, M. E. (2010). Black sheep
get the blues: A psychobiological model of social rejection and depression.
Neuroscience & Biobehavioral Reviews, 35(1), 39–45.
Steptoe, A., Hackett, R. A., Lazzarino, A. I., Bostock, S., La Marca, R., Carvalho,
L. A., & Hamer, M. (2014). Disruption of multisystem responses to stress
in type 2 diabetes: Investigating the dynamics of allostatic load. PNAS Pro-
ceedings of the National Academy of Sciences of the United States of Amer-
ica, 111(44), 15693–15698. https://doi.org/http://dx.doi.org.libproxy1.usc.edu/10
.1073/pnas.1410401111
Thompson, B. (2004). Exploratory and confirmatory factor analysis: Understand-
ing concepts and applications. Washington, DC: American Psychological
Association.
Volpato, S., Blaum, C., Resnick, H., Ferrucci, L., Fried, L. P., & Guralnik, J. M. (2002).
Comorbidities and impairments explaining the association between diabetes and
lower extremity disability. Diabetes Care, 25(4), 678–683. https://doi.org/10
.2337/diacare.25.4.678
Volpato, S., Leveille, S. G., Blaum, C., Fried, L. P., & Guralnik, J. M. (2005). Risk
factors for falls in older disabled women with diabetes: The women’s health
and aging study. The Journals of Gerontology Series A: Biological Sciences and
Medical Sciences, 60(12), 1539–1545.
World Health Organization Global report on diabetes. (2016). Retrieved January 21,
2017 from http://www.who.int/diabetes/global-report/en/.
Wu, S., Ell, K., Gross-Schulman, S. G., Sklaroff, L. M., Katon, W. J.,
Nezu, A. M., . . . Guterman, J. J. (2014). Technology-facilitated depression care
management among predominantly Latino diabetes patients within a public safety
net care system: Comparative effectiveness trial design. Contemporary Clinical
Trials, 37(2), 342–354.

https://doi.org/10.1017/9781108669016.011 Published online by Cambridge University Press

 10
A Multidisciplinary Study on the Relationship
between Foster Care Attributes and
Posttraumatic Stress Disorder Symptoms
on Foster Youth
Amanda Yoshioka-Maxwell1 , Shahrzad Gholami, Emily Sheng,
Mary Hemler, Tanachat Nilanon, and Ali Jalal-Kamali

Introduction
Nearly 2 million unaccompanied youth aged 13–24 years experience
homelessness in the United States each year (National Alliance to End
Homelessness, 2012). Almost 40% of these youth have spent time in foster
care (Dworsky et al., 2012). Homeless former foster youth are a vulnerable
population receiving relatively little attention in the scientific literature,
compared to both homeless youth and youth with foster care experiences.
However, 30% of all homeless adults report a foster care history, compared
to 4% of the general public (Courtney, Piliavin, Grogan-Kaylor, & Nesmith,
2001; Reilly, 2003; Roman & Wolfe, 1995), and research has shown that
between 11% and 36% of the foster youth population experience homelessness
(Dworsky et al., 2012).
Among both homeless youth and former foster youth, rates of trauma expe-
riences and PTSD symptomatology are significantly higher than those of the
general public (Hudson & Nandy, 2012). Studies have indicated that trauma
experiences common among homeless youth and former foster youth impact
the development of PTSD symptoms, and that experiences in the child welfare
system can exacerbate the impact of trauma on youths’ mental health (Bruskas,
2008; Pecora, White, Jackson, & Wiggins, 2009; Stewart et al., 2004). In exam-
ination of issues such as homelessness, foster care, and PTSD, social sciences
have historically utilized statistical modeling, such as regression models, to
examine the relationship between these variables. Statistically significant mod-
els have been produced as a result, indicating associations between foster care
experiences, homelessness experiences, and PTSD symptoms. And while these
models have produced valuable information for intervention and treatment

1
Amanda Yoshioka-Maxwell and Shahrzad Gholami are joint first authors.

169

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 170 Relationship between Foster Care Attributes and PTSD

programs aimed at treating PTSD symptoms, statistical modeling assumes

that cause-and-effect relationships remain unchanged, and it is based on the
collection of limited data, making it difficult to extrapolate findings more
generally. Additionally, for predicting outcomes such as PTSD symptoms,
statistical models may not be the most effective modeling strategy based on the
types of data largely accessible in social sciences (Lette et al., 1994). Toward
this end, modeling based on the use of artificial intelligence (AI) through
machine learning may be effectively used to predict outcomes that typical
regression models cannot. Given the limited use of AI in many social sciences,
and given the need to continue to explore the impact of foster care experiences
on PTSD symptoms among homeless former foster youth, this analysis seeks
to explore how effectively a number of AI modeling techniques can predict
PTSD symptoms, compared to more traditional statistical models.

Literature Review
PTSD Symptoms among Youth with Homelessness
and Foster Care Histories
As children and youth are placed into foster care primarily as a result of
substantiated physical abuse, sexual abuse, or neglect, youth with child welfare
experience are often at elevated risk for symptoms associated with trauma
exposure (Badeau & Gesiriech, 2004). In a study assessing alumni of the
foster care system, 30% of respondents met lifetime diagnostic criteria for
PTSD, while only 7.6% of a general population sample with comparable
demographics met such criteria (Pecora et al., 2009). Similar disparities have
been found in other studies, with two studies on adolescents aged 17–18 years
preparing to exit the child welfare system reporting PTSD lifetime prevalences
of 14% and 16% among participants (Keller, Salazar, & Courtney, 2010;
McMillen et al., 2005). In contrast, research has demonstrated that the
lifetime prevalence of PTSD among 17- and 18-year-olds in the general
population is approximately 7% (Merikangas et al., 2010). In addition, a
study involving a large sample of lower-income adults in an urban setting
aged 19–23 years reported a lifetime prevalence of 7%; transition-age youth
with foster care history therefore appear to experience lifetime prevalence
rates approximately twice that of same-age peers in the general population
(Breslau et al., 2004).
Research has indicated that youth exiting foster care are at substantially
greater risk for homelessness than their peers without foster care history

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 Amanda Yoshioka-Maxwell et al. 171

(Dworsky, Napolitano, & Courtney, 2013), and homeless adults have nearly
eight times the likelihood of having foster care history than the housed
general public (Reilly, 2003). It has been suggested that many youths leave
home to escape abuse; a recent study demonstrated that up to 85% of
homeless youth may have experienced either physical or sexual abuse prior
to becoming homeless, with 42% of these youth having reported experiencing
both (Keeshan & Campbell, 2011). However, while many of these youth leave
home to escape abuse, experiencing homelessness puts them at higher risk for
victimization, with evidence demonstrating that homeless youth experience
disproportionately high rates of robbery, assault, and sexual assault (Thrane,
Hoyt, Whitbeck, & Yoder, 2006). The increased rates of victimization that
homeless youth experience are associated with significantly higher rates of
PTSD, even among the youth who witnessed an assault but did not experience
one (Bender, Ferguson, Thompson, & Langenderfer, 2014).
Although the experience of a traumatic event is a necessary prerequi-
site for a diagnosis of PTSD, not all individuals who experience one or
multiple traumatic events meet the criteria for the disorder. In addition,
research has demonstrated that while experiencing childhood abuses and
street victimizations was independently associated with negative mental health
outcomes, their interaction was not (Bender, Brown, Thompson, Ferguson, &
Langenderfer, 2015). Contrary to prior research demonstrating that individuals
who experience several abuse types, followed by victimizations, frequently
have poorer mental health outcomes (Ford, Elhai, Connor, & Freuh, 2010),
recent research has suggested that experiencing street victimization did not
exacerbate the relationship between childhood abuse and PTSD (Bender
et al., 2015). As traditionally used hierarchical logistic regression models have
failed to provide insight into the cumulative effect of traumatic experiences on
the presence or absence of PTSD, AI modeling techniques may offer insights
that cannot be revealed by linear modeling (Bender et al., 2015).

Machine Learning within the Social Sciences

Machine learning techniques coupled with data mining have been used in
different studies to assess PTSD symptoms. Random forest classifiers were
used for prediction purposes based on different data sets including psychiatric
interview, psychiatric scales (i.e., Clinician-administered PTSD scale, positive
and negative syndrome scale, Hamilton anxiety scale, and Hamilton depression
scale) and a combination of both of them. This study was conducted on 102
inpatients, i.e., 51 with a diagnosis of PTSD and 51 with psychiatric diagnoses

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 172 Relationship between Foster Care Attributes and PTSD

other than PTSD. Applicability of data–mining methods was demonstrated,

and group of comorbid diagnoses, including neurotic, stress-related, and
somatoform disorders, surfaced as important (Marinić et al., 2007). Also,
another rsearch study was conducted using a sample of 1,517 treatment-
seeking military veterans to develop a truncated assessment protocol based
on the Clinician-administered PTSD scale. Decision tree analysis was used
to analyze the most important predictor variables among all items in the
Clinician-administered PTSD scale to diagnose PTSD (Stewart, Tuerk, Met-
zger, Davidson, & Young, 2016). Although decision tree–based models were
shown to be accurate in the existing literature, their importance in analyzing
PTSD symptoms for homeless youth based on their foster care attributes is still
uninvestigated.

Data Augmentation
Generally, machine learning techniques work best with more data
(Domingos, 2012). Data augmentation is the collection of techniques used
to increase the amount of data points, such that machine learning techniques
can train better predictive models. These techniques have become popular
in computer vision, speech recognition, and, more recently, natural language
processing. In computer vision, data augmentation for image recognition
involves applying a transformation to an image, such as a rotation, translation,
or flip. The new image depicts the same objects and object relationships as the
old image, but the new image is a new data point for the model (Hauberg et al.,
2015; Wong, Gatt, Stamatescu, & McDonnell, 2016; Zhang & LeCun, 2016).
In speech recognition, examples of transformations include the addition of
artificial noise and the manipulation of speech signals (Hannun et al., 2014);
in natural language processing, researchers have replaced words and phrases
with synonyms (Zhang & LeCun, 2016).
Data augmentation techniques are not generally used in social science
research; as social science research outcomes directly affect the well-being of
human subjects, analysis from real data is typically used. As such, there is not
much prior work on exploring augmentation techniques on human data. In our
work, we augment our data sets in order to show the viability of our machine
learning techniques if we collected more human data; we do not use data
augmentation as a technique to directly influence intervention and treatment
policies. In the intersection of AI and social work, it remains unclear if data
augmentation is more of a strength or a weakness; on one hand, it could help
improve the effectiveness of AI techniques, but on the other hand, the results
may not be meaningful from a social work perspective. We experiment with

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 Amanda Yoshioka-Maxwell et al. 173

data augmentation in our work to investigate how it could be effective to both

disciplines.

Data Samples and Procedures

Two data sets will be utilized for these analyses. The first data set utilized the
YouthNet data set, a convenience sample of 352 homeless former foster youth
(aged 13–25) from three drop-in centers in Hollywood, Venice, and Santa
Monica, CA, collected as part of a panel study of 1,046 homeless youth. Any
client receiving services at a participating agency during the data collection
periods was eligible to participate. General inclusion criteria for participation
in the study included self-identifying as homeless (e.g., sleeping on the streets,
in an emergency shelter, at immediate risk of being homeless such as couch-
surfing, about to be evicted) and being 13–25 years old. Recruitment was
conducted for approximately 1 month at each site; during that time period,
recruiters were present at the agency to approach youth for the duration of
service provision hours. Youth new to the agency first completed the agency’s
intake process before beginning the study to ensure they met the eligibility
requirements for the agency (and thus the study). Two research staff members
were responsible for all recruitment to ensure youth did not complete the
survey multiple times during each data collection period per site.
Signed voluntary informed consent was obtained from each participant,
with the caveat that child abuse and suicidal and homicidal intentions would
be reported. Informed consent was obtained from youth 18 years of age or
older and informed assent was obtained from youth 13 to 17 years old. The
affiliated institutional review board waived parental consent because homeless
youth younger than 18 were unaccompanied minors who may not have had a
parent or adult guardian from whom to obtain consent. Interviewers received
approximately 40 hours of training, including lectures, role-playing, mock
surveys, ethics training, and emergency procedures. The study consisted of
two parts: a computerized self-administered survey and a social network
interview. The former included an audio-assisted version for participants with
low literacy, and both parts of the survey could be completed in English or
Spanish. All participants received $20 in cash or gift cards as compensation
for their time. The institutional review board approved all survey items and
procedures.
The second set (FCHIV) of data utilized a sample of 155 homeless former
foster youth, utilizing the same drop-in center and data collection procedures as
the YouthNet Study. In this second study, however, additional questions were
added to the YouthNet questionnaire. This augmented Foster Care Experiences

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 174 Relationship between Foster Care Attributes and PTSD

Assessment included quantitative measures created from qualitative interviews

and questions related to foster care experience. The sample collected was
modeled after the YouthNet Study in terms of age range, data collection site,
and participant demographics. The only exception is the distribution of race
among two populations, which slightly differed between the two datasets.

Data: Selected Variables and Outcomes for PTSD Prediction

To predict PTSD symptoms, the most relevant variables in two datasets are
chosen as follows:
YouthNet Data Set:

1. Age
2. Gender
3. Sexual Orientation
4. Race
5. Age at placement in foster care
6. Number of foster care placement
7. Time of homelessness

FCHIV Data Set:

1. Age
2. Gender
3. Sextual Orientation
4. Race
5. Age at placement in foster care
6. Number of foster care placement
7. Foster care placement type
8. Reason for being placed in foster care
9. Their overall opinion of foster care experience
10. Asking whether they were respected while in foster care
11. Asking whether they felt supported while in foster care

The outcome variables assessing for PTSD (included in both data sets) are:
In the past 30 days,

1. Have you had nightmares about a situation or thought about that when
they did not want to?
2. Have you tried hard not to think about a situation or went out of your way
to avoid situations that reminded you of it?
3. Were you constantly on guard, watchful, or easily startled?

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 Amanda Yoshioka-Maxwell et al. 175

4. Have you felt numb or detached from others, activities, or their

surroundings?

The main outcomes of study could be either the total sum of PTSD symptoms
experienced or the dichotomized version, i.e., presence or absence of any
PTSD symptom.

Methods
Data Preprocessing for Models
Data Sets: To process the data for our models, we filter out individuals
without PTSD predictor values, leaving a total of 449 individuals from the
YouthNet data set and the FCHIV data set combined, 133 from FCHIV and 316
from YouthNet. We standardize the features “Age at placement in foster care”
and “Number of foster care placement” so that the values are binned for both
data sets. As mentioned in the previous section, YouthNet Data Set includes
7 predictor features and FCHIV Data Set includes 11 predictor features. So,
when these two data sets are combined, some predictor feature values are
missing for data samples inherited from YouthNet Data Set.
In the first group of experiments, we combined and preprocessed YouthNet
and FCHIV data set, which contains 449 samples. In this case, we enable the
existence of missing values for missing predictor features for data samples
inherited from YouthNet Data Set, i.e., when some features are present in
one data set but not in the other, we use the value “−1” to represent these
missing values. In these experiments, 25% of the combined data is held out
as the test set to evaluate the results of learning on different training sets. The
training data sets are constructed based on (1) 75% of the FCHIV+YouthNet
combined data sets, (2) only the YouthNet data set, (3) only the FCHIV data
set, (4) variations of the augmented data set, and (5) 75% of the FCHIV data
set in conjunction with subsets of the augmented data sets (we experimented
with FCHIV combined with different subsets of augmented data as opposed
to FCHIV+YouthNet combined with different subsets of augmented data,
because FCHIV has more non-missing foster care-specific features that we
thought might be interesting to study if we had more data).
In the second group of experiments, we want to see if certain features are
more important predictors of PTSD than others. We run a series of experiments
where we remove different subsets of features. The intuition behind these
experiments is that if removing feature X results in a dramatic decrease of
model score, then we know feature X must be an important predictor of PTSD.

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 176 Relationship between Foster Care Attributes and PTSD

In the third group of experiments, we want to examine the FCHIV and

YouthNet data sets separately. The motivation is that if the two data sets were
collected under slightly different contexts and thus have different distributions
of features, it may not be the best idea to combine everything together. In
other words, we train models only on one data set at a time and also test
only on the corresponding test sets. We also take a common subset of features
between FCHIV and YouthNet, create separate data sets for FCHIV-subset and
YouthNet-subset, and train and test separate models. The common subset of
features used in this group of experiments are gender, age, sexual orientation,
race, entry age into foster care, and number of foster care placements. One
final data set variation we experiment with is combining FCHIV-subset and
YouthNet-subset and training one model on the combined data set.

Augmented Data Sets: We perform data augmentation based on the dis-

tributions of values for each feature and the correlations between features.
Specifically, we first decide the number of new individual data points we want
to generate. We constrain the generation so that half of the individuals will
have PTSD symptoms and half will not. Some machine learning techniques
are more sensitive to data imbalance, so we ensure our data generation does
not skew the population of people with and without PTSD symptoms.
Before generating new data, we calculate the distributions of values for all
features based on a simple ratio of observed values for each feature. We also
calculate the kappa coefficient (Cohen, 1960) between each pair of categorical
features (all data set predictor features excluding age). The kappa coefficient
measures the correlation of two categorical variables above chance.
The data augmentation algorithm we use is depicted in Figure 10.1. To
generate data for a sample, we first choose if the sample has the presence or
absence of PTSD symptoms; we assign a binary 1 or 0 to the dichotomized
PTSD outcome variable. Then, we randomly select the next feature X to
assign a value to. To choose a value X-value for feature X, we generate a
random real number between 0 and 1. If the number is less than or equal to
1
|Previously selected f eatures| + 1 , then we choose an X-value from the distribution of
feature X in the preprocessed data set of 449 samples. Otherwise, we look
at all the features that have already been selected for the sample and get the
correlation scores between all previously selected features and feature X. We
1
use |Previously selected f eatures| + 1 as a basic threshold to decide whether to pick
the X-value from feature X’s distribution or based on the correlation with
previously selected features; in our work, we are interested in observing the
general effect data augmentation and not necessarily searching for the best data

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 Amanda Yoshioka-Maxwell et al. 177

Figure 10.1 Data augmentation algorithm used to generate a new sample

augmentation method. An interesting point of experimentation could involve

tuning the threshold value as a hyperparameter.
If “Age” is a previously selected feature, we skip its correlation score
because we only calculate kappa correlations for pairs of categorical features.
If “Age” is the feature of interest, i.e., feature X, then we simply randomly
choose the X-value from the distribution of feature X as described earlier.
We normalize the correlation scores across all previously selected features,
generate another random real number between 0 and 1, and use the number to
select which previously selected feature value Y-value to condition on when
choosing the X-value. To choose an X-value given the previously selected
Y-value, we take all samples in the preprocessed dataset that have feature
Y value of Y-value, and collect a list of the X-values of the aforementioned
samples. We experiment with two different approaches to select an X-value
from the list of X-values. In what we call the Average Correlation approach,
we choose the X-value by taking an average over the list of X-values. In our
Random Correlation approach, we choose a random X-value from the list
of X-values. After generating an X-value for feature X, we then repeat the
process until we have generated values for all features and thus created a new
sample. To generate a value for the predictor PTSD variable that is the sum
of the four individual PTSD variables, we constrain the value to be 0 if the
dichotomized PTSD variable is 0, and otherwise randomly generate a value
from 1 to 4.
For this work, we generate 551 new samples with the Average Correlation
method and another 551 new samples with the Random Correlation method.
This gives us two separate sets of augmented data to experiment with, which
can be used separately or in conjunction with the preprocessed real set of 449
samples.

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 178 Relationship between Foster Care Attributes and PTSD

Modeling Techniques and Evaluations

Evaluation Techniques
Prior to discussing the different machine learning techniques studied in this
chapter, we briefly mention the evaluation metrics we used. According to
existing machine learning literature, we considered precision, recall, and
F1 score to evaluate the prediction results, quantitatively. In more details,
precision is the fraction of the true positive labeled instances compared to all
positive labeled classes predicted by model. Recall is the fraction of the true
positive labeled instances compared to all actual positive labeled instances in
test set. F1 score is a score to measure accuracy, which considers both precision
and recall. More precisely, F1 score is a harmonic mean of precision and recall
defined as F1 = 2. (precision * recall)/(precision + recall). We study various
machine learning techniques including Logistic Regression as the baseline,
Neural Networks, Decision Tree, and Bayesian Networks. In addition to the
evaluation metrics, we also assess the branching rules generated by Decision
Tree model from a social science viewpoint to develop insight about the most
important factors affecting PTSD experiences for former foster care youths.

Model 0 (baseline): Logistic Regression

We use logistic regression as a baseline technique to predict PTSD from our
sample features. Logistic regression is similar to linear regression, though
the dependent variable is categorical instead of continuous. This technique
allows us to analyze which sample features influence the PTSD variables.
The precision, recall, and F1 results are shown in Table 10.1. From the model
weights we also observed that the support people received and being placed
in a shelter had the largest influence on the PTSD variables. Cox (1958)
first proposed this technique, and it has been popular as a general machine
learning method. We use this technique as a baseline for our other techniques
because we believe more sophisticated methods will let us better examine the
relationship between foster care variables and PTSD variables.
In Table 10.1, we observe that for logistic regression, using FCHIV alone
as the training set best predicts the presence of PTSD in the test set, which
is composed of a subset of FCHIV and a subset of YouthNet. We also see
that the augmented data set for training does no better than FCHIV alone,
which makes sense, as the augmented data set was created according to feature
distributions and correlations in the FCHIV and YouthNet data sets. Table 10.1
also shows that FCHIV and a subset of the augmented data set does no better
than just the augmented data set. We also see that the dichotomized PTSD
prediction models always outperform the multi-class PTSD prediction models

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 Amanda Yoshioka-Maxwell et al. 179

Table 10.1. Logistic regression results for models trained on different

training data and tested on the same YouthNet+FCHIV test data (first group
of experiments)

Dichotomized Multi-class
PTSD prediction PTSD prediction
Training data Precision Recall F1 Precision Recall F1

FCHIV 0.65 0.82 0.72 0.34 0.17 0.18

YouthNet 0.57 0.64 0.60 0.19 0.13 0.12
FCHIV+YouthNet 0.55 0.43 0.48 0.23 0.14 0.16
Aug_Rand_Corr-552 0.65 0.39 0.49 0.20 0.22 0.19
Aug_Avg_Corr-552 0.60 0.82 0.69 0.29 0.17 0.16
FCHIV + Aug_Avg_Corr-552 0.58 0.82 0.68 0.22 0.15 0.12
FCHIV + Aug_Avg_Corr-100 0.59 0.81 0.68 0.24 0.16 0.14
FCHIV + Aug_Avg_Corr-50 0.51 0.60 0.55 0.26 0.5 0.21

by a significant amount. The low scores of the multi-class PTSD prediction

models could be due to the fact that the multiple “classes” of PTSD are the
number of predefined symptoms that individuals report having experienced,
and that the number of data samples we have is relatively low.
We examine the effects of different features on the logistic regression model
more closely in Table 10.2. Taking different subsets of features for all data
samples in the combined FCHIV and YouthNet data sets, the best model results
are when we use only the foster care–specific features. In general, the scores
we get from withholding different subsets of features do not vary by significant
amounts, which suggests that we have an insufficient amount of data to be able
to meaningfully generalize from the logistic regression model.
We perform additional experiments with models that are both trained and
tested on subsets of the same data, e.g., trained on a subset of FCHIV and tested
on a subset of FCHIV, to see if models would better be able to predict PTSD
for samples from the same data collection; the results are shown in Table 10.3.
Although the different test data makes it difficult to compare these additional
models, the results do not indicate particularly high precision, recall, and F1
scores for the respective data sets.
Here we present the logistic regression model as a baseline model. We
expect that with more sophisticated machine learning techniques that are more
tuned to this prediction task, we can get better results. Regardless, if the spar-
sity of data presents a problem in the simpler logistic regression models, we
acknowledge that it might have a larger impact on the more complex models.

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 180 Relationship between Foster Care Attributes and PTSD

Table 10.2. Results for models trained on FCHIV+YouthNet training set with
withheld features

Dichotomized PTSD prediction

Withheld features Precision Recall F1

None 0.55 0.43 0.48

(Used all [‘gender’, ‘age’,‘sexori’, ‘raceall’]
and foster care questions)
[‘sexori’]
(Used only [‘gender’, ‘age’, ‘raceall’] and 0.57 0.54 0.55
foster care questions)
[‘gender’, ‘sexori’]
(Used only [‘age’, ‘raceall’] and foster care 0.54 0.52 0.53
questions)
All foster care questions
(Used only [‘gender’, ‘age’, ‘sexori’, 0.54 0.45 0.49
‘raceall’])
All foster care questions + sexori
(Used only [‘gender’, ‘age’, ‘raceall’]) 0.58 0.55 0.56
[‘gender’, ‘age’, ‘sexori’, ‘raceall’]
(Used only foster care questions)

Table 10.3. Logistic regression results for models trained on different training
data and tested on corresponding test data. FCHIV subset and YouthNet
subset use a subset of features for all samples (third group of experiments)

Dichotomized PTSD prediction

Training data Features used Precision Recall F1

FCHIV All 0.78 0.58 0.67

Subset 0.79 0.46 0.58
YouthNet All 0.65 0.71 0.68
Subset 0.59 0.57 0.58
Combined All 0.55 0.43 0.48
Subset 0.63 0.54 0.58

Model 1: Neural Networks

A neural network is a powerful machine learning model that can learn a map-
ping from inputs to outputs using objective function. For binary classification
outputs, the most commonly used objective function is binary cross-entropy,
which can be considered a continuous extension of prediction accuracy. It
has been shown that neural networks are universal function approximators
(Csáji, 2001), meaning that they can learn any mapping present in the data

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 Amanda Yoshioka-Maxwell et al. 181

given enough data and trainable parameters. Here, we will not only utilize
neural networks for our prediction task but also use them as a tool to
discover feature importance. A base neural network will be trained with all
the features available in the data set. Model complexity of this base model
has to be adjusted such that it has a reasonable prediction performance. Then,
comparative neural networks will be trained, but with features that are subsets
of original set of features. By comparing their prediction performance, we can
discover the importance of each feature or a combination of features.
To illustrate the process, let us examine Figure 10.2. The base model to the
left is trained with all the features available in the data set. The comparative
model to the right is trained with only features Xa and Xb . If the comparative
model’s prediction performance is significantly worse than the base model’s,
we infer that feature Xc contributes significantly to the model’s prediction
performance. It should be noted, however, that this does not imply a direct
causation link.
As a rule of thumb, the size of the data set should be comparable to or
larger than the number of trainable parameters. Due to the small size of this
dataset, we have to design the architecture of the neural networks carefully,
making sure that we are frugal with the number of trainable parameters.
After experimentation, we selected a model that has four hidden layers with
rectified linear activation. The output layer has exactly one node with sigmoid
activation. We experimented with dropout technique to reduce overfitting and
have found that our data set is too small to effectively utilize dropout in our
model.

Y Y

Neural Neural
Network Network

Xa Xb Xc Xa Xb Xc

Base model Comparative model

Figure 10.2 Base model versus comparative model

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 182 Relationship between Foster Care Attributes and PTSD

Table 10.4. Neural network results for models trained on different

training data and tested on the same YouthNet+FCHIV test data

Dichotomized PTSD prediction

Training data Precision Recall F1

FCHIV 0.61 0.68 0.65

YouthNet 0.58 0.62 0.60
FCHIV+YouthNet 0.61 0.81 0.70
Aug_Rand_Corr-552 0.58 0.85 0.69
Aug_Avg_Corr-552 0.56 0.72 0.63
FCHIV+Aug_Avg_Corr-552 0.65 0.33 0.44
FCHIV+Aug_Avg_Corr-100 0.67 0.42 0.51
FCHIV+Aug_Avg_Corr-50 0.67 0.35 0.47

Table 10.5. Base model and comparative model results, using

YouthNet+FCHIV as training data

Dichotomized PTSD prediction

Withheld features Accuracy Precision Recall F1

None 0.59 0.61 0.81 0.70

[‘sexori’] 0.58 0.60 0.82 0.69
[‘gender’, ‘sexori’] 0.60 0.60 0.94 0.73
All foster care questions 0.58 0.58 1.00 0.73
(Used only [‘gender’, ‘age’, ‘sexori’, ‘raceall’])
All foster care questions + sexori 0.58 0.58 1.00 0.73
(Used only [‘gender’, ‘age’, ‘raceall’])
[‘gender’, ‘age’, ‘sexori’, ‘raceall’] 0.60 0.60 0.94 0.73
(Used only foster care questions)

From experimental results, we conclude that for this classification task,

having extra augmented data does not help improve prediction accuracy.
The results in Table 10.5 suggest that the neural network learns very poorly
in this data set. Using only gender, age, and race as features, we can obtain
the same prediction performance as using all the available features. Due to
encouraging results obtained from training decision trees on each data set
separately, we decided to also train neural a network on each data set separately
as well. Table 10.6 reports the results.
The results in Table 10.6 improve markedly over previous experiments.
This suggests that the data from the two data sets (YouthNet and FCHIV)
was sampled from two significantly different populations. Traditional machine

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 Amanda Yoshioka-Maxwell et al. 183

Table 10.6. Neural network results when trained and tested

on each data set separately

Dichotomized PTSD prediction

Training data Features used Precision Recall F1

FCHIV All 0.72 0.96 0.82

Subset 0.64 0.58 0.61
YouthNet All 0.62 0.98 0.76
Subset 0.57 0.85 0.68
Combined All 0.61 0.81 0.70
Subset 0.64 0.99 0.78

learning methods will struggle to learn if they are trained on merged data
containing such discrepancies. This issue is well known in the machine
learning community and has been studied under the name “domain adaptation.”
Interested readers can take a look at a recent work of interest published by
Ganin and colleagues (2015). We left the incorporation of domain adaptation
in our model for future work.

Model 2: Decision Tree

A decision tree model is a well-known technique for inductive learning that is
comprised of a sequence of branching operations based on the comparisons
of existing quantities of specific attributes or features of the data samples.
Decision tree classifies instances of data samples by sorting them down the
tree to some leaf node, which provides the classification of instances. Each
node in the tree represents a test of some attribute of the instance, and each
branch descending from that node corresponds to one of the possible values
of that attribute. An instance is classified by starting at the root node of the
tree, testing the attribute specified by this node, then following the tree branch
corresponding to the value of the attribute in the given example. This process
is then repeated for the subtree rooted at the new node. Figure 10.3 illustrates
a typical learned decision tree. This decision tree classifies a day according
to whether it is suitable for playing tennis (Mitchell, 1997). In this study, we
learn a decision tree model to predict the presence of PTSD symptoms based
on the different variations of the datasets mentioned earlier. The decision tree’s
attributes are the aforementioned shared variables of the data sets.
Table 10.7 demonstrates the results for the decision tree learned based on
the setup for the first group of the experiments. The main outcome is the
dichotomized version of PTSD symptoms. Experiment results demonstrate

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 184 Relationship between Foster Care Attributes and PTSD

Table 10.7. Decision tree results for models trained on

different training data and tested on the same
YouthNet+FCHIV test data (first group of experiments)

Dichotomized PTSD prediction

Train data Precision Recall F1

FCHIV 0.65 0.39 0.49

YouthNet 0.62 0.6 0.6
FCHIV+YouthNet 0.62 0.75 0.68
Aug_Rand_Corr-552 0.57 0.54 0.55
Aug_Avg_Corr-552 0.56 0.66 0.61
FCHIV +Aug_Avg_Corr-552 0.55 0.61 0.58
FCHIV + Aug_Avg_Corr-100 0.61 0.51 0.55
FCHIV + Aug_Avg_Corr-50 0.62 0.6 0.6

OUTLOOK

Sunny Overcast Rain

Yes
HUMIDITY WIND

High Normal High Normal

No Yes No Yes

Figure 10.3 A typical learned decision tree

that augmented data sets are dominated by the original data sets in their
performance quality. Another important point about Table 10.4 is that the
model learned based on the smaller and more specific data set, FCHIV, is not
a good representative for a test set that consists of 25% of the combined data
set. Most of the specific variables are missing in the combined samples, and
that may lead to poor results. Learning based on the YouthNet data set and
the combined data set results in nearly similar results; however, the combined
version outperforms all other training schemes. This observation led us to
conduct the third group of experiments, where we learned and evaluated each
data set separately to address the intrinsic differences between two data sets.
It is worth noting that, not only the prediction results are not sufficiently
significant, but also the large amount of missing data renders the application
of decision tree rules difficult to justify. Figure 10.4 demonstrates the first

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 Amanda Yoshioka-Maxwell et al. 185

FC-support

< 2.5 > = 2.5

Years_homeless Yes

< 1.5 > = 1.5

FC-entryage Years_homeless

<0 >=0 < 12 >=12

Yes No Yes No

Figure 10.4 The learned decision tree for predicting PTSD symptoms

three layers of the decision tree learned. The root node represents the foster
care support attribute of the data set. The first rule indicates that if foster care
support attribute is larger than 2.5, there is a high probability of experiencing
PTSD – which is counterintuitive. As discussed earlier, the YouthNet data
set has a larger number of data samples and a smaller number of variables.
Consequently, combining the data sets while enabling missing variables leads
to a high level of uncertainty and missing values for major portion of the data
points. The presence of a huge imbalance in the distribution of the samples
over different categories demonstrates that handling missing data requires a
more attentive treatment.
To minimize the adverse effects of missing values and challenges due to
combining two data sets, we conducted the third group of experiments dis-
cussed in the data processing subsection. According to Table 10.8, evaluation
results improve when we study each data set independently. Subset means that
we only use gender, age, sexual orientation, race, entry age into foster care,
and number of foster care placements to make prediction, rather than using all
available features. For both FCHIV and YouthNet, when we reduce number
of the variables, accuracy is reduced. Since some of the more specific vari-
ables, which had a predicting potential, were removed, accuracy consequently
declined. To facilitate comparison, the row for Combined-All is basically
borrowed from the previous group of experiments, as it is impossible to disable
the missing values and hold all of the available features simultaneously while
merging two different size data sets. Since the majority of the data samples
are from YouthNet data set, the combined data set is mostly similar to the

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 186 Relationship between Foster Care Attributes and PTSD

Table 10.8. Decision tree results for models trained on different training data
and tested on corresponding test data. FCHIV subset and YouthNet subset use
a subset of features for all samples (third group of experiments).

Dichotomized PTSD prediction

Training data Features used Precision Recall F1

FCHIV All 0.88 0.96 0.92

Subset 0.79 0.92 0.85
YouthNet All 0.82 0.76 0.79
Subset 0.75 0.85 0.8
Combined All 0.62 0.75 0.68
Subset 0.78 0.86 0.82

YouthNet data. So supplementing the YouthNet with data from FCHIV results
in improvement in predictions compared to individual YouthNet dataset. This
combined version for the subset of features is demonstrated in the last row
of the table, which is showing a nearly average performance between two
individual data sets.
Results of decision tree model presented in Table 10.8, outperforms all
other models. So in addition to the standard metrics for evaluation, we also
present the explanation of decision trees learned. Since decision tree is a
rule-based model, it provides the opportunity to develop insight about the
most important underlying factors in prediction of outcomes. We examine the
decision trees rules for the most general data set, which is a combination
of YouthNet and FCHIV for the common subset of features, and the most
specific data set in terms of foster care attributes, which is FCHIV including
all of its variables, in order to figure out the most important variables for
predicting PTSD.
For the general data set, age at entry into foster care, number of foster
care placements, and gender were found to be the most important factors,
which confirms that foster care attributes play a more important role in
prediction of experiencing PTSD compared to only demographics. Several
traced rules for discovering PTSD symptoms for the general data set is
outlined below:

• If you have less than 3 placements AND are 22 years old or older
• If you have more than 3 placements AND you are lesbian, gay, or bisexual
• If you have more than 3 placements AND you were placed in foster care
before the age of 2 AND you’re under 19 years old

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 Amanda Yoshioka-Maxwell et al. 187

• If you have more than 3 placements AND you were placed in foster care
after the age of 3 AND you are female or transgender
• If you have more than 3 placements AND you are heterosexual or
questioning AND you were placed in foster care over the age of 6 years
• If you have more than 3 placements AND you are heterosexual or
questioning AND you were placed in foster care over the age of 6 years
AND you had less than 5 placements
• If you have more than 3 placements AND you are heterosexual or
questioning AND you were placed in foster care under the age of 5 years
AND you are currently under 19 years old
• If you have more than 3 placements AND you are heterosexual or
questioning AND you were placed in foster care at 6 years old or older
AND you had more than 3 placements AND you are questioning

For the specific data set, gender, fc-reason-neglect and fc-opinion were found
to be the most important factors. Several traced rules for discovering PTSD
symptoms for the specific data set are as follows:

• If you were placed in foster care at birth

• If you were placed for parental drug problems
• If you were placed for parental drug problems AND you were not placed in
a juvenile detention camp
• If you were placed for and “other” reason
• If you have an opinion of FC that is not entirely positive AND you were
placed earlier than 4 years old
• If you have an opinion of FC that is not entirely positive AND you were
placed older than 4 years old AND you are female or transgender
• If you have an opinion of FC that is not entirely positive AND you were
placed older than 4 years old AND you are male AND you are under 21.5
years old

Model 3: Bayesian Networks

A Bayesian network (BN) is a probabilistic graphical model that represents a
set of variables and their conditional dependencies via a directed acyclic graph
(DAG). For example, suppose that there are two events that could cause grass to
be wet: either the sprinkler is on or it is raining. Also, suppose that the rain has a
direct effect on the use of the sprinkler (namely that when it rains, the sprinkler
is usually not turned on). Then the situation can be modeled using a Bayesian
network (shown to the right). All three variables have two possible values, T

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 188 Relationship between Foster Care Attributes and PTSD

Sprinkler Rain
Rain T F T F
F 0.4 0.6 0.2 235
Sprinkler Rain
T 0.01 0.99

Grass Wet

Grass Wet
Sprinkler Rain T F
F F 0.0 1.0
F T 0.8 0.2
T F 0.9 0.1
T T 0.99 0.01

Figure 10.5 Example of a Bayesian network

(for true) and F (for false). Figure 10.5 represents the network structure of
dependencies and the corresponding probabilities.
Based on the network of the dependencies above we can fit that to the data
and obtain a set of coefficients in order to build a linear predictive model
containing the variables of the network. Such a model then can be used for
inferences and and predictions.
Bayesian networks are very intuitive for humans to understand, and one
of the main uses of Bayesian network is to find the causality through the
dependences in the network.
There are two ways to find the dependency network structure:
1. The network dependencies can be defined by the domain export based on
their knowledge of the data and their dependencies of the variables.
2. The dependency network can be learned from the data based on the
conditional probabilities. There are different algorithms to use for learning
such network.
For our current data we used both data sets separately and also combined from
the augmented data set. We tried different multiple learning algorithms for
extracting the links from the datasets. The methods used are:
1. Grow shrink and two variations of it
2. Tabu
3. Hill climbing
The best results were from hill climbing algorithm for all data sets.
However, in all cases, the PTSD variable had no dependency links. That means

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 Amanda Yoshioka-Maxwell et al. 189

that we cannot use our variables to predict the PTSD variable with the Bayesian
network dependencies. The reason is that since there are no links at all to
the PTSD variable, it is conditionally independent of all other variables, and
it reduces to just a probability distribution, since all predictive coefficients
are zero.
Nonetheless, learning the links from different data sets can exploit other
dependencies. Specifically the links that are common among networks from
different data sets indicate a very strong dependency among such variables.
Figures 10.6–10.8 present the dependency links learned for different data
sets using the hill climbing method.
As one could imagine, there is a very strong network among variables
related to foster care and age in the FCHIV data set. As it is shown in all of the
networks, we can see that the foster care age of entry has a direct effect on the
number of places for foster care, in all three networks. So even though based
on this network we cannot make any predictions for the PTSD variable, we
can find the natural dependencies among variables based on the data. Also, the
domain knowledge is quite important for the links learned in the network. For
example, even though the variables age and gender have a dependency from
conditional probability perspective, based on the domain we know that they are

Figure 10.6 The learned network from the FCHIV dataset

Figure 10.7 The learned network from the YouthNet dataset

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 190 Relationship between Foster Care Attributes and PTSD

Figure 10.8 The learned network from the Augment_avgcorr dataset

completely independent, which leads us to consider the links of dependency

only based on the domain knowledge.
Looking at the data itself, it is observable that the data sets are quite
small and the data are very sparse. Both of those qualities cause the BN
learning algorithms to suffer a lot and have a difficult time finding the proper
dependencies. Given that the data were collected prior to analysis, it would be
helpful if future data collection efforts would be mindful of those two qualities
for a BN analysis.

Discussion and Conclusions

Artificial intelligence (AI) space models have a great deal to offer social
workers focusing on behavioral and mental health outcomes. For populations
that can be hard to reach, such as homeless youth, and difficult to sample
in large numbers, the predictive modeling capabilities of AI may be able to
answer some difficult questions that social workers have struggled to answer
using traditional statistical models. In an attempt to determine if AI models
may better predict PTSD outcomes among a sample of former foster youth, a
series of experiments were run with models such as neural networks, Bayesian
networks, and decision trees. Not only were we able to improve the prediction
accuracy by using rule-based techniques (i.e., decision tree), but also these
methods provided some extremely valuable insight for research within this
population. Data augmentation, while not a predictive model, can be extremely
useful for generating additional data samples, which can be later used for
prediction. Among the specific predictive models tested, neural networks were
not found to be effective for predicting PTSD symptomatology when compared
to the logistic regression model. These findings were most likely the result of
small sample size. However, future work with larger size may prove the use of

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 Amanda Yoshioka-Maxwell et al. 191

neural networks to be helpful in predicting outcomes for populations such as

homeless former foster youth.
Bayesian networks were also not effective at finding dependency links
between foster care characteristics and PTSD symptoms. However, when
examining both data sets, dependency links emerged between foster care
entry age and number of placements, indicating a relationship between the
two. Thus, while this model was not able to predict PTSD symptoms from
a sample of former foster youth, it was able to direct our attention to other
variables, within the arena of foster care variables, that could be further
explored. Future work with Bayesian networks should consider the impact that
sparse networks have on the results. Finally, a series of decision trees were
used to explore the predictive nature of foster care attributes where PTSD
symptoms were concerned. Models for both data sets were run separately
as well as combined. Final models utilized only those variables common
to both data sets, as they outperformed all other models. Results indicate
several important features of these data. For the most general model, utilizing
both data sets and only those variables common to both data sets, it was
determined that the age at which a youth entered the foster care system,
number of placements in foster care, and gender were the most important
variables for predicting PTSD outcomes among homeless former foster youth.
For the more specific model, utilizing only the foster care data set with all
available variables, it was determined that gender, being placed into foster
care due to neglect, and one’s opinion of foster care (positive or negative)
were the most important variables for predicting PTSD symptoms among this
population. Overall, the decision tree results outperformed the other models,
with an F1 score 20% larger. More specifically, decision tree models and neural
networks outperformed logistic regression in terms of accuracy of predictions
and standard evaluation metrics (e.g., F1 score), but decision tree modeling
was more helpful in terms of explanation compared to neural networks.
Potential hindrances to the performance of decision trees may be differences
between the data sets. While the smaller FCHIV data set intended to model
its gender and race demographics, differences, such as race, do vary between
the data. These differences make comparison between the data much more
difficult.
These results highlight several important aspects of this population. Model-
ing based on both data sets and more general variables yielded highly different
results compared to the more specific data set with a larger number of variables
but smaller sample size. Thus, depending on the individuals of the population
being considered, different models should be used. Alternatively, modeling
techniques that explicitly handle population difference can be explored, such

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 192 Relationship between Foster Care Attributes and PTSD

as domain adaptation. Additionally, results indicate important information for

working with youth with PTSD symptoms. For the more general model, in
predicting PTSD symptoms, youth’s age of foster care placement, number of
placements, and gender should be considered when attempting to intervene
in the development of PTSD symptoms. For the more specific model, gender,
reason for foster care placement, and opinion of foster care should be consid-
ered. These results could prove to be extremely important in the development
of screening tools and interventions aimed at reducing the burden of PTSD
symptoms, as it directs practitioners on specific characteristics and experiences
that would likely impact the development of PTSD. Future work combining
social work constructs and AI methods should consider the importance of the
type of data collected and needed, both to capture all necessary aspects of
a population, but also to meet the specific data requirements needed for AI
predictive models. Nevertheless, the future of AI and social work has proven
to be full of potential, leveraging the powerful predictive techniques of AI to
solve real-world social work problems.

References
Badeau, S., & Gesiriech, S. (2004). A child’s journey through the child welfare system.
Washington, DC: The Pew Commission on Children in Foster Care.
Bender, K., Brown, S. M., Thompson, S. J., Ferguson, K. M., & Langenderfer, L.
(2015). Multiple victimizations before and after leaving home associated with
PTSD, depression, and substance use disorder among homeless youth. Child
Maltreatment, 20(2), 115–124.
Bender, K., Thompson, S. L., Ferguson, K. M., Yoder, J., & Kern, L. (2014).
Trauma among street-involved young adults. Journal of Emotional and Behavioral
Disorders, 22, 53–64.
Breslau, N., Wilcox, H. C., Storr, C. L., Lucia, V. C., & Anthony, J. C. (2004). Trauma
exposure and posttraumatic stress disorder: A study of youths in urban American.
Journal of Urban Health: Bulletin of the New York Academy of Medicine, 81(4),
530–544.
Bruskas, D. (2008). Children in foster care: A vulnerable population at risk. Journal of
Child and Adolescent Psychiatric Nursing, 21(2), 70–77.
Cohen, J. (1960). A coefficient of agreement for nominal scales. Educational and
Psychological Measurement, 20(1), 37–46. doi:10.1177/001316446002000104.
Courtney, M. E., Piliavin, I., Grogan-Kaylor, A., & Nesmith, A. (2001). Foster youth
transitions to adulthood: A longitudinal view of youth leaving care. Child Welfare,
80(6), 685.
Cox, D. R. (1958). The regression analysis of binary sequences (with discussion).
Journal of the Royal Statistical Society B. 20, 215–242.
Domingos, P. (2012). A few useful things to know about machine learning. Communi-
cations of the ACM, 55(10), 78–87.
Dworsky, A., Dillman, K. N., Dion, R. M., Coffee-Borden, B., & Rosenau, M. (2012).
Housing for youth aging out of foster care: A review of the literature and program
typology. Washington, DC: U.S. Department of Housing and Urban Development.

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 Amanda Yoshioka-Maxwell et al. 193

Ford, J. D., Elhai, J. D., Connor, D. F., & Freuh, C. B. (2010). Poly-victimization and
risk of posttraumatic, depressive, and substance use disorders and involvement in
delinquency in a national sample of adolescents. Journal of Adolescent Health, 46,
545–552.
Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F.,
Marchand, M., & Lempitsky, V. (2015). Domain-adversarial training of neural
networks. Journal of Machine Learning Research, 17, 1–35.
Hannun, A., Case, C., Casper, J., Catanzaro, B., Diamos, G., Elsen, E., . . . Ng, A. Y.
(2014). Deep speech: Scaling up end-to-end speech recognition. arXiv preprint
arXiv:1412.5567.
Hauberg, S., Freifeld, O., Larsen, A. B. L., Fisher III, J. W., Hansen, L. K., Pudwell,
L., . . . Pechenot, B. (2015). Dreaming more data: Class-dependent distributions
over diffeomorphisms for learned data augmentation. arXiv preprint arXiv:
1510.02795.
Hudson, A., & Nandy, K. (2012). Comparisons of substance abuse, high-risk sexual
behavior and depressive symptoms among homeless youth with and without a
history of foster care placement. Contemporary Nurse, 42(2), 178–186.
Keller, T. E., Salazar, A. M., & Courtney, M. E. (2010). Prevalence and timing of
diagnosable mental health, alcohol, and substance use problems among older
adolescents in the child welfare system. Children and Youth Services Review,
32(4), 626–634.
Lette, J., Colletti, B. W., Cerino, M., McNamara, D., Eybalin, M. C., Lavasseur, A.,
& Nattel, S. (1994). Artificial intelligence versus logistic regression statistical
modeling to predict cardiac complications after noncardiac surgery. Clinical
Cardiology, 17(11), 609–614.
Marinić, I., Supek, F., Kovičć, Z., Rukavina, L., Jendričko, T., & Kozarić-Kovčć, D.
(2007). Posttraumatic stress disorder: Diagnostic dat analysis by data mining
methodology. Croatian Medical Journal, 48(2), 185–197.
Merikangas, K. R., He, J. Burstein, M., Swanson, S. A., Avenevoli, S.,
Cui, L., . . . Swendsen (2010). Lifetime prevalence of mental disorders in U.S. ado-
lescents: Results from the National Comorbidity Survey Replication – Adolescent
Supplement (NCS-A). Journal of the Academy of Child & Adolescent Psychiatry,
49(10), 980–989.
McMillen, J. C., Zima, B. T., Scott, L. D., Auslander, W. F., Munson, M. R., Ollie,
M. T., Spitznagel, E. L. (2005). Prevalence of psychiatric disorders among older
youths in the foster care system. Journal of the American Academy of Child &
Adolescent Psychiatry, 44, 88–95.
Mitchell, T. M. (1997). Machine learning. Burr Ridge, IL: McGraw-Hill Education.
National Alliance to End Homelessness (2012). An emerging framework for end-
ing unaccompanied youth homelessness. http://www.endhomelessness.org/library/
entry/an-emerging-framework-for-ending-unaccompanied-youth-homelessness.
Pecora, P. J., White, C. R., Jackson, L. J., & Wiggins, T. (2009). Mental health of current
and former recipients of foster care: A review of recent studies in the USA. Child &
Family Social Work, 14(2), 132–146.
Reilly T. (2003). Transition for care: Status and out of youth who age out of foster care.
Child Welfare. 82(6), 727–748.
Roman, N. P., & Wolfe, P. (1995). Web of failure: The relationship between foster care
and homelessness. Washington, DC: National Alliance to End Homelessness.
Stewart, A. J., Steiman, M., Cauce, A. M., Cochran, B. N., Whitbeck, L. B., & Hoyt,
D. R. (2004). Victimization and posttraumatic stress disorder among homeless

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 194 Relationship between Foster Care Attributes and PTSD

adolescents. Journal of the American Academy of Child & Adolescent Psychiatry,

43(3), 325–331.
Stewart, R. W., Tuerk, P. W., Metzger, I. W., Davidson, T. M., & Young, J. (2016).
A decision-tree approach to the assessment of posttraumatic stress disorder:
Engineering empirically rigorous and ecologically valid assessment measures.
Psychological Services, 13(1), 1.
Wong, S. C., Gatt, A., Stamatescu, V., & McDonnell, M. D. (2016, November).
Understanding data augmentation for classification: when to warp? In Digital
Image Computing: Techniques and Applications (DICTA), 2016 International
Conference on (pp. 1–6). IEEE.
Zhang, X., & LeCun, Y. (2016). Text understanding from scratch. arXiv preprint
arXiv:1502.01710.

https://doi.org/10.1017/9781108669016.012 Published online by Cambridge University Press

 11
Artificial Intelligence to Predict Intimate
Partner Violence Perpetration
Robin Petering, Mee-Young Um, Nazanin Alipourfard, Nazgol Tavabi,
Rajni Kumari, and Setareh Nasihati Gilani

Introduction
One-third of homeless youth (HY) experience some sort of physical intimate
partner violence (IPV), and frequently young people experience bidirectional
partner violence. One in four homeless youth have reported being both a
victim and a perpetrator in their current or most recent romantic relationship
[1]. Programs that address youth IPV frequently focus on victim and survivor
support. Although supporting victims/survivors is necessary, offering only this
type of support is not only paternalizing, with the underlying assumption
that it is the victim’s fault or responsibility to prevent IPV; it also does not
effectively break the cycle of violence [2]. IPV prevention programs should
focus on reducing perpetration. Although “batterer interventions” exist [3, 4],
they intervene after the incident of violence has occurred. These programs are
typically court ordered, meaning the perpetrator participates in the intervention
after there is an arrest or conviction. Additionally, “batterer interventions”
have been shown to be ineffective and potentially do more harm than good
in reducing IPV [4, 5]. The current study proposes a method to predict the
occurrence of adolescent and young adult IPV perpetration using machine
learning. Using data from homeless youth, an IPV perpetration “triage tool”
will be built that can be implemented in the field to identify young people who
are at high risk for engaging in violence perpetration. This tool will encourage
targeted services that will buffer the likelihood of engaging in violence,
such as mindfulness training to reduce impulsivity or general education in
healthy conflict resolution skills. In order to solve the aforementioned problem,
different ML techniques will be used to predict the chance of perpetration for
an individual and also most effective factors will be identified using statistical
approaches and later compared with primary expectations from social work
perspectives.

195

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 196 Predict Intimate Partner Violence Perpetration

Problem Definition
Violence is a complex phenomenon that impacts adolescents and young adults
across the United States. It occurs in multiple ways including interpersonal
violence, intimate partner violence, gang violence, and gun violence. Homeless
youth experience all types of violence at higher rates than their housed
counterparts [6–8]. This is typically the result of many contributing factors
including childhood experiences of trauma, subsistence survival strategies, and
exposure to perpetrators while living on the streets [9–11]. The consequences
of violence are severe: besides the proximal consequence of severe injury
or death, violence can also cause nonphysical ailments such as posttrau-
matic stress disorder, depression and externalizing behavior, delinquency, and
aggression. It is clear that reducing violence in the lives of homeless youth
is imperative and will contribute to a young person’s ability to safely and
successfully exit the streets and lead a long a productive life in society.
However, public health and social interventions to reduce violence within
adolescent and young adult populations are difficult [12], because, again,
this phenomenon is complex, with many intrinsic and extrinsic contributing
factors.
The public health model recognizes three levels of prevention: primary,
secondary, and tertiary [13]. Primary prevention is designed to reduce inci-
dence by preventing the first occurrence of an event. Secondary prevention
is designed to decrease the prevalence of a problem after its onset and
often includes interventions targeting populations at greatest risk of harm.
Tertiary prevention occurs once a problem is clearly evident and causing
harm [13, 14].
There are primary IPV prevention mechanisms that are curriculum based,
which focus on teaching awareness about violence, promoting healthful
behavior, and teaching conflict resolution skills. However, the IPV problem
among homeless populations is more complex, given that the youth are not
usually going to school and are not a regular part of any community to get
access to those primary intervention techniques [12].
Moreover, some primary intervention systems use simplistic and generic
methods that might not be particularly effective in the long run. There are
several community programs to help the IPV victims, but not many programs
are out there that are focused on identifying and counseling the perpetrators
and batterers. Often, perpetrators learn their violent behavior by witnessing
or being exposed to domestic violence during their formative years. The
personal background and upbringing of each perpetrator play a key role in
their future violent behaviors. Hence, these might go unnoticed in traditional

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 Robin Petering et al. 197

intervention techniques, which highlights the need for newer and efficient ways
to intervene. Most of the intervention and prevention programs for IPV for the
homeless have limited funding, and it is important that we use the allocated
resources in the most efficient way. AI-supported methods are expected to
provide targeted interventions that are more effective.
Currently, there are several organizations that are working on many of the
world’s hardest problems: combating child exploitation, disrupting illicit net-
works, delivering humanitarian aid in the wake of conflict and natural disasters,
and more. When we talk about AI here, it spans the core AI, machine learning,
and data mining using machine learning methods. AI and data mining helps
revolutionize the way we use data in pursuit of helping the society. Further,
unlike many other proposals to improve society, machine learning tools can be
easily scaled to larger demographics depending on the requirements.

Data Set Description

As part of a longitudinal study of Los Angeles area homeless youth, drop-
in service–seeking youth completed a self-administered questionnaire. The
presented results are from the third panel of data collection (sample size; N =
452). The Revised Conflict Tactics Scale (CTS2) was used to assess physical
IPV perpetration. The sample was limited to youth who answered the questions
related to IPV (99 youth either did not answer the corresponding questions or
were never in a relationship, which narrows down the sample size to 353). The
research team approached all youths who entered the service agencies during
the data collection period and invited them to participate in the study. The
selected agencies provided weekday services to eligible HY, including basic
needs, medical and mental health services, case management, and referrals and
connections to other programs such as housing services. Each youth signed a
voluntary consent form, and a consistent pair of research staff members was
responsible for all recruitment to prevent youths from completing the survey
multiple times during each data collection period per site. The questionnaire
asks the respondent about their personal life, their interactions with other
people, where and how they live, the quality of their relationship and sexual
life if they have any partner, etc. These data are unique in that they include the
Revised Conflict Tactics Scale [15], which was to assess physical IPV perpe-
tration in each participant’s most recent intimate relationship. The CTS2 is the
most widely used instrument in research on interpersonal violence and includes
data on perpetration and victimization across various domains of violence (i.e.,
physical, emotional, relational, sexual, and threatening). The current study’s

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 198 Predict Intimate Partner Violence Perpetration

Table 11.1. Description of data set variables

Age Age at the time of interview

Exchange Sex Exchanged sex for money or other items
Children Number of children
Suicide Attempt Suicide attempt in previous 12 months
Homeless Age Age of first homelessness
Jail Ever been in jail
Juggalo Identifies as a Juggalo (fan of musical group ICP)
Male Identifies as male
LGBQ Sexual minority
White Of Caucasian race/ethnicity
Literal Homeless Is literally homeless (sleeps on street, park, car, etc.)
Weapon Carried a weapon in previous month
Violence Engaged in interpersonal violence (physical fight) in
previous 12 months
Gang Is current or former gang member
PTSD Has symptoms of PTSD
Depressed Has symptoms of depression
Community Violence Witnessed community violence during childhood
Sexual Abuse Experienced childhood sex abuse
Physical Abuse Experienced some form of physical child abuse
(physical and witness family IPV)
Lonely 1 How often do you feel that you lack companionship?
Lonely 2 How often do you feel left out?
Lonely 3 How often do you feel isolated from others?
Hard Drug Use Hard drug use in past 30 days
Foster Care Ever in foster care
Job Currently has a job
IPV perpetration In a violent relationship or not?
(The outcome variable we want to predict)

primary outcome variable was physical perpetration that includes a range of

items from “I slapped my partner” (minor) to “I kicked my partner” (severe).
The original data set had more than 1000 variables. The number of variables
in comparison with the number of participants is too large for using ML
techniques. Therefore, unrelated variables were removed from the data set
for analyses, which resulted in decreasing the number of potential predictor
variables to 26. The variable name and description are presented in Table 11.1.

Data Analysis
The data analysis occurred in several stages. First, we did an analysis based
on what is found in previous social science literature on IPV perpetration.
Second, we did a p-value and LASSO technique analysis. Each stage in this

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 Robin Petering et al. 199

analysis phase will be used to identify the most important features that could
be included in a IPV perpetration triage tool.

Theoretical Quantitative Analysis

Before performing machine learning analyses, we established a baseline
analysis as a comparison. We identified variables in the current data set that
correspond with risk factors and predictors that have been found in previous
literature. Age, race, and gender were included in the baseline analysis as
statistical control variables. For adolescents and youth, IPV experiences are not
predicted by gender, which is contrary to adult experiences of IPV [16]. The
most common predictor for youth IPV perpetration is experiences of childhood
maltreatment or abuse. A large body of literature has suggested that exposure to
maltreatment in childhood is related to some form of IPV in later life [17–21].
The link between maltreatment and IPV is often referred to as intergenerational
transmission of violence [22, 23], which is derived from Bandura’s (1979)
social learning theory. This framework suggests that children exposed to
violence, either as a victim or a witness, are more likely to use violence
as a tactic for conflict resolution because this has been modeled in intimate
relationships during childhood. A statistical analysis was performed including:
Age, White, Male, Community Violence, Sexual Abuse, and Physical Abuse.

Statistical Analysis
Multiple methods were used to uncover the most important features. P-value
is described here; AI algorithm techniques, such as SVM and random forest,
are described in the Methodology section later in the chapter.

P-value
The p-value is used in the context of null hypothesis testing in order to quantify
the idea of statistical significance of evidence [24]. We calculated this value for
each feature to achieve a ranking of the most influential ones.

LASSO
A method that is useful in determining the most important features is LASSO
[25]. LASSO is a regression analysis method that performs both variable
selection and regularization in order to enhance prediction accuracy and
interpretability of the statistical model it produces. In other words, it will give
us the variables that are most influential on the target variable. Since we have

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 200 Predict Intimate Partner Violence Perpetration

many variables, using this method will help us reduce the number of features
to consider in the final model and in our triage tool at the end, since it is not
feasible to consider all the features that we have now.

Support Vector Machines (SVM)

SVM is a well-known predictive model. Assuming features form the dimen-
sions in space (in this way each data point would be a point in our space),
SVM finds the hyperplane or a set of hyperplanes that best separates the data
points of different classes. Best separator for an SVM model is the one with
the largest margin (distance of the hyperplane to the nearest data point of any
class). Having the trained model, a previously unseen data point is mapped into
the same space, and based on its position relative to the hyperplane, it is marked
as positive or negative (in a binary classification). For some data sets, different
classes are not linearly separable, so to generate nonlinear classifiers, kernels
are introduced. Different kernels and also different parameters for each kernel
are among parameters we tune to get the best model. The best SVM model
for our data set is an rbf kernel. Though this model is clearly a predictive
model, its other features lead us to try it to find the most effective features. The
hyperplane constructed by SVM gives a weight to each feature in space. These
weights or coefficients can be used to determine which features had a bigger
impact in separating the positive and negative values and which features were
irrelevant [26].

Random Forest
Random forest is a popular algorithm for feature ranking. In the structure of
decision trees, leaves represent class labels, or “IPV_perpetration” in our data,
and branches represent conjunctions of features that lead to those class labels.

Final Rankings
In Table 11.2, seven most effective features based on each algorithm are given.
LASSO and T -test are described in the previous subsections; SVM and Ran-
dom Forests are described in the Methodology section. The repetitive features
in algorithms are used as important features. With this approach, the most
important features are: “PTSD”, “depressed”, “violence”, “hard_drug_use”,
“lonely3”, “exchange_sex”.
After removing “ptsd” from variables set, we run the ranking algorithms
again on the data. By removing “ptsd,” other variables can increase their power
in the data. Based on the results, most important features are: “depressed”,
“hard_drug_use”, “violence”, “gang”, “physical_abuse”, “lonely 3”.

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 Robin Petering et al. 201

Table 11.2. Ranking of the features in the decreasing order of importance

SVM Random Forest T-test LASSO

1 lonely2 depressed ptsd violence

2 hard_drug_use ptsd depressed ptsd
3 ptsd homeless_age hard_drug_use depressed
4 exchange_sex Age violence hard_drug_use
5 juggalo violence lonely3 gang
6 lonely3 weapon physical_abuse exchange_sex
7 sexual_abuse hard_drug_use suicide_attempt LGBQ

Table 11.3. Ranking of the features, after excluding PTSD, in the decreasing
order of importance

SVM Random Forest T-test LASSO

1 hard_drug_use depressed depressed depressed

2 exchang_sex homeless_age hard_drug_use hard_drug_use
3 gang violence violence violence
4 lonely2 Age lonely3 gang
5 juggalo hard_drug_use physical_abuse physical_abuse
6 lonely3 lonely3 suicide_attempt LGBQ

Methodology
In this section, we will describe the final features we used in our ML
techniques, our baseline model with which we will compare our results, and a
short description of all the ML algorithms that we used.

Features
As discussed in the section describing the data set, we have 26 features
in our data set. In the data analysis part, we were trying to find the most
important variables that we have. Based on the results, we found out that
“ptsd” is a major contributor to determining the target value. But in fact, this
is a controversial issue, because the direct correlation of this variable with
perpetration is unclear. There is an immense research body that shows that
PTSD is a direct consequence of violence [27], and without longitudinal data

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 202 Predict Intimate Partner Violence Perpetration

there is no way to confirm this is not what is happening. Therefore, PTSD was
excluded from the predictive modeling. In order to see the importance of this
variable and to determine if we can perform as good as we did when we are not
considering this variable at all, we are defining three feature sets with which to
explore our ML algorithms. The first feature set includes all the features. The
second one is just considering the most important ones based on the results of
our data analysis. And the third feature set is using the most important features
when we remove the “ptsd” from the initial feature set and then rank the other
features based on their importance. Results of running our algorithms on these
three distinct feature sets are reported in the “Results” section.

Baseline
Due to lack of prior work in the domain of our problem, we define literature-
based baseline. We used logistic regression algorithm for the following vari-
ables: age, white, male, community_violence , sexual_abuse, physical_abuse.

Learning Algorithms
Predictive models use known results to develop (or train) a model that can be
used to predict values for unseen data. Modeling provides results in the form of
predictions that represent a probability of the target variable based on estimated
significance from a set of input variables. Using the previously described
features, we tried several approaches by using different ML algorithms to
build our classification models. Some of the supervised ML system that
we explored were: Logistic Regression, Support Vector Machines, Random
Forest, Neural Network, Deep SVM. We trained and tested these models using
K-fold (fivefold) cross-validation, and in order to evaluate the models we used
metrics such as precision, recall, F1-score, ROC AUC (Area Under Curve) and
accuracy.

Neural Networks
Artificial neural networks are a computational model used in several com-
putational techniques in computer science and other research disciplines.
They are analogous to biological neurons where connections between neurons
carry activation signals of varying strengths. If the combined strength of
the incoming signals is large enough, the neuron becomes activated and is
passed on to the next layer [28]. We chose neural networks for our model
because they have an ability to generalize and respond to unexpected inputs

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 Robin Petering et al. 203

and patterns. But at the same time neural nets usually need massive amounts
of data to train, which might be the reason we are not getting the best results.
For our model we are using 10 layers and 10 hidden nodes.
The structure of a decision tree gives a good overview of important features;
thus, beside classification, we use random forest for feature ranking.

Logistic Regression
Logistic regression is a linear classifier that finds the hyperplane that can best
describe the data (by mapping the data points onto that hyperplane) and then
separates the data points by a threshold on that hyperplane, which again best
separates the data points of different classes [29].

Deep Support Vector Machine

The next model we used for classification is two-class locally deep support
vector machine. It is a supervised learning method and is available as part of
the Microsoft Azure Toolkit. It creates a two-class classification model and
uses nonlinear support vector machine (SVM) optimized to achieve higher
efficiency in cases of larger training sets [30]. Use of a local kernel function
enables the model to learn arbitrary local feature embeddings, including high-
dimensional, sparse, and computationally deep features that introduce non-
linearities into the model. The kernel function that is used for mapping
data points to feature space is designed to reduce the training time while
maintaining the classification accuracy. LD-SVMs are best used when we have
a complex data such that linear models (viz. logistic regression) perform poorly
and data is very complex.

Results
We performed our experiments with several algorithms. To standarize the
experiments we used 10-fold cross-validation of handling the test and training
sets. The following metrics were used to determine which algorithms worked
best for our experiments.

1. Accuracy – Accuracy is the proportion of true results to the total number

of examined data.
2. Precision – Precision is the probability measure that a retrieved
information is relevant.
3. Recall – Recall is the probability measure relevant information is retrieved
in a search.

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 204 Predict Intimate Partner Violence Perpetration

Table 11.4. Results of prediction of classification algorithms for all variables

(feature set 1)

Deep Neural Random Logistic SVM(RBF

SVM Net Forest Regression kernel) Baseline

Accuracy 0.65 0.62 0.66 0.61 0.66 0.55

Precision 0.5 0.47 0.46 0.42 0.48 0.36
Recall 0.48 0.45 0.34 0.56 0.6 0.52
F1 0.48 0.44 0.44 0.48 0.53 0.42
ROC_AUC 0.64 0.6 0.64 0.62 0.69 0.54

4. F1-score – A measure that combines precision and recall in the harmonic

mean of precision and recall, the traditional F-measure or balanced
F-score:
(2*(Precision*Recall)/(precision+recall)
5. ROC AUC – Area under the curve

Since we have imbalanced data, the best measures for comparing different
algorithms are F1-measure and ROC AUC. As shown in Table 11.4, both
F1-score and ROC AUC measure of the SVM classifier (with RBF kernel)
have the highest value and produces the best results, so this is the model we
propose for our problem.
For defining the most important factors, we find factors that are more
common in our four rankings available in Table 11.2. The most important
variables are “ptsd”, “depressed”, “violence”, “hard_drug_use”, “lonely3”,
and “exchange_sex”.
As mentioned before, after removing “ptsd” from features, we then
extracted the most important features from the rest of the features. Hence
we got a new feature set: “depressed”, “hard_drug_use”, “violence”,
“gang”, “physical_abuse”, “lonely3”. Table 11.6 shows the results using this
feature set.
To summarize the data presented in Table 11.4 and Figure 11.1, Table 11.5
and Figure 11.2, and Table 11.6 and Figure 11.3, we observe that SVM gave
us the best results in two of our feature sets. When using all the variables as
our feature set, we got the highest ROC and F1-score (ROC = 0.69, F1 = 0.53)
using SVM with RBF kernel. SVM with linear kernel gave us the best ROC
and F1-score when we used the significant variables (after removing PTSD
from our initial variables) (ROC = 0.67, F1 = 0.51). When exploring with
the significant variables including PTSD, we got the highest ROC using the
logistic regression method (ROC = 0.69 ). In this case, SVM with RBF kernel

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 Robin Petering et al. 205

Table 11.5. Results of prediction of classification algorithms for important

variables including PTSD (feature set 2)

Deep Neural Random Logistic SVM(RBF

SVM Net Forest Regression kernel) Baseline

Accuracy 0.69 0.71 0.62 0.66 0.58 0.55

Precision 0.49 0.6 0.41 0.48 0.43 0.36
Recall 0.36 0.42 0.43 0.63 0.9 0.52
F1 0.41 0.47 0.42 0.54 0.58 0.42
ROC_AUC 0.64 0.71 0.67 0.69 0.67 0.54

Figure 11.1 Comparison of different algorithms for feature set 1

gave us the highest F1-score of 0.58. In general, we didn’t get great results in
terms of F1 measure and ROC curve. However, these might have resulted from
our limited and sparse data set. We had too many variables compared to the
size of our data set; we also had a lot of missing values for some variables.

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 206 Predict Intimate Partner Violence Perpetration

Table 11.6. Results of prediction of classification algorithms for important

variables excluding PTSD (feature set 3)

Deep Neural Random Logistic SVM(Linear

SVM Net Forest Regression kernel) Baseline

Accuracy 0.66 0.66 0.61 0.63 0.63 0.55

Precision 0.43 0.5 0.41 0.45 0.46 0.36
Recall 0.26 0.36 0.41 0.58 0.58 0.52
F1 0.3 0.4 0.41 0.51 0.51 0.42
ROC_AUC 0.67 0.64 0.63 0.66 0.67 0.54

Figure 11.2 Comparison of different algorithms for feature set 2

Conclusion
Violence experienced during adolescence is related to experiences of violence
in adulthood, as a result of developing ineffective coping strategies in situations
of conflict within relationships [31]. Additionally, it is known that children
who witness intimate partner violence between their parents are at increased
risk of becoming perpetrators themselves. There is a broad understanding
that violence is not only learned, but there is a “cycle of violence,” meaning
violent victimization only perpetuates more violence interpersonally and

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 Robin Petering et al. 207

Figure 11.3 Comparison of different algorithms for feature set 3

intergenerationally. To break this cycle of violence, it is imperative that we

shift the narrative from “victim” to “perpetrator.” Although victim support
services are necessary, solely focusing on victim services perpetuates an
underlying “victim-blaming,” misogynistic, and oppressive service system. As
noted before, perpetrator programs exist but they are ineffective and happen
after the incident of violence. Additionally, by switching the narrative from
“victim” to “perpetrator” we are not proposing a “perpetrator-blaming” model.
To that end, our study aimed to build a triage tool that could help us predict
perpetrator behaviors among homeless youth populations before they start
showing such behaviors. Our study results indicated that experiencing physical
abuse during childhood, being a current or former gang member, and engaging
in a recent physical fight were significantly associated with IPV perpetration.
Machine learning techniques allowed us to gain better insights into the data
set, which traditional techniques used in social sciences could not have alone.
IPV perpetration is the result of negative early-life experiences and other
life stressors. Holding the individual accountable does not take into consider-
ation a life-course or systems model that is the framework that defines social
work research and practice. To stop the cycle of violence, primary prevention
services are necessary. IPV can affect youth regardless of race, culture, or

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 208 Predict Intimate Partner Violence Perpetration

socioeconomic status. However, certain groups, including homeless youth,

are at greater risk of experiencing IPV when compared to the general youth
population (O’Keefe, 1997). Healthy relationship and conflict resolution skills
building programs should be offered widely but in particular within at-risk
populations. Given limited resources, which is the unfortunate reality of social
service settings, a tool that could triage those that are at risk for perpetrating
violence toward their partner could be extremely useful in reducing overall
violence in adolescent and youth populations. This triage tool could be used
as brief screening or assessment and provide targeted supportive services that
would improve an individual’s skills in coping with stress and conflict in a
relationship. It is important that predictive models for social good come from a
supportive approach rather than punitive one. The predictive model and triage
tool that we are building will not claim that an individual is a perpetrator; it will
only identify those that are at risk. This is an important distinction, especially
since our model intends to prevent violent incidents before they occur.
It is important to note some potential limitations of our study. All survey
data rely on individual’s reports of abuse perpetrated, and their experiences
of violence may be subject to response bias. Despite extensive interviewee
training and efforts to ensure privacy, individuals might not have felt able to
disclose their experiences of perpetrating violence on their intimate partners.
Although our study provides important insight on what factors are the most
important in perpetration of violence, we might have limited details about the
degree of violence or the conditions in which the violence occurred, lack of
some of which might change our results drastically. Also, the data set we had
was too small to verify the correct applicability of machine learning and data
mining algorithms and to build a model based on that. In social sciences a data
set with more than 300 observations with rich behavioral data is applauded, but
in computer science it is not. This modeling was also only done within a sample
of homeless youth in Los Angeles. In the future, we can have more generalized
versions of the data set that span the demographics, living conditions, and age
of the individuals. Hopefully, we are able to predict and help the pepetration
even better in the coming future with using better data sets and better data
mining algorithms.

References
[1] Petering, R., Rhoades, H., Rice, E., and Yoshioka-Maxwell, A. (2015). Bidirec-
tional intimate partner violence and drug use among homeless youth. Journal of
Interpersonal Violence, 32(14):2209–2217.

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 Robin Petering et al. 209

[2] Meyer, S. (2011). Seeking help for intimate partner violence: Victims’ experi-
ences when approaching the criminal justice system for IPV-related support and
protection in an australian jurisdiction. Feminist Criminology, 6(4):268–290.
[3] Feder, L. and Wilson, D. B. (2005). A meta-analytic review of court-mandated
batterer intervention programs: Can courts affect abusers’ behavior? Journal of
Experimental Criminology, 1(2):239–262.
[4] McGinn, T., Taylor, B., McColgan, M., and Lagdon, S. (2016). Survivor per-
spectives on ipv perpetrator interventions: A systematic narrative review. Trauma,
Violence, & Abuse, 17(3):239–255.
[5] Murphy, C. M. and Baxter, V. A. (1997). Motivating batterers to change in the
treatment context. Journal of Interpersonal Violence, 12(4):607–619.
[6] Petering, R., Rice, E., and Rhoades, H. (2016). Violence in the social networks
of homeless youths: Implications for network-based prevention programming.
Journal of Adolescent Research, 31(5):582–605.
[7] Heerde, J. A., Hemphill, S. A., and Scholes-Balog, K. E. (2014). “Fighting”
for survival: A systematic review of physically violent behavior perpetrated
and experienced by homeless young people. Aggression and Violent Behavior,
19(1):50–66.
[8] Eaton, D. K., Kann, L., Kinchen, S., Shanklin, S., Flint, K. H., Hawkins, J.,
et al. (2012). Youth risk behavior surveillance-united states, 2011. Morbidity
and Mortality Weekly Report. Surveillance Summaries (Washington, DC: 2002),
61(4):1–162.
[9] Milburn, N. G., Rotheram-Borus, M. J., Rice, E., Mallet, S., and Rosenthal, D.
(2006). Cross-national variations in behavioral profiles among homeless youth.
American Journal of Community Psychology, 37(1–2):63.
[10] Robertson, M. J. and Toro, P. A. (1999). Homeless youth: Research, intervention,
and policy. In Practical lessons: The 1998 national symposium on homelessness
research. Washington, DC: US Department of Housing and Urban Development
and US Department of Health and Human Services.
[11] Whitbeck, L. B. and Hoyt, D. R. (1999). Nowhere to grow: Homeless and
runaway adolescents and their families. Piscataway, NJ: Transaction Publishers.
[12] Petering, R., Wenzel, S., and Winetrobe, H. (2014). Systematic review of current
intimate partner violence prevention programs and applicability to homeless
youth. Journal of the Society for Social Work and Research, 5(1):107–135.
[13] Wolfe, D. A. and Jaffe, P. G. (1999). Emerging strategies in the prevention of
domestic violence. The Future of Children, 9(3):133–144.
[14] Limbos, M. A., Chan, L. S., Warf, C., Schneir, A., Iverson, E., Shekelle, P., and
Kipke, M. D. (2007). Effectiveness of interventions to prevent youth violence: A
systematic review. American Journal of Preventive Medicine, 33(1):65–74.
[15] Bandura, A. (1979). The social learning perspective: Mechanisms of aggression.
In H. Toch (Ed.), Psychology of crime and criminal justice (pp. 198–236).
Prospect Heights, IL: Waveland Press.
[16] Johnson, W. L., Giordano, P. C., Manning, W. D., and Longmore, M. A. (2015).
The age–IPV curve: Changes in the perpetration of intimate partner violence
during adolescence and young adulthood. Journal of Youth and Adolescence,
44(3):708–726.

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 210 Predict Intimate Partner Violence Perpetration

[17] Delsol, C. and Margolin, G. (2004). The role of family-of-origin violence in men’s
marital violence perpetration. Clinical Psychology Review, 24(1):99–122.
[18] Eriksson, L. and Mazerolle, P. (2015). A cycle of violence? Examining family-of-
origin violence, attitudes, and intimate partner violence perpetration. Journal of
Interpersonal Violence, 30(6):945–964.
[19] Franklin, C. A. and Kercher, G. A. (2012). The intergenerational transmission
of intimate partner violence: Differentiating correlates in a random community
sample. Journal of Family Violence, 27(3):187–199.
[20] McKinney, C. M., Caetano, R., Ramisetty-Mikler, S., and Nelson, S. (2009).
Childhood family violence and perpetration and victimization of intimate partner
violence: Findings from a national population-based study of couples. Annals of
Epidemiology, 19(1):25–32.
[21] Widom, C. S., Czaja, S., and Dutton, M. A. (2014). Child abuse and neglect
and intimate partner violence victimization and perpetration: A prospective
investigation. Child Abuse & Neglect, 38(4):650–663.
[22] Widom, C. S. and Wilson, H. W. (2015). Intergenerational transmission of
violence. In Violence and mental health (pp. 27–45). New York: Springer.
[23] Widom, C. S. (1989). The cycle of violence. Science, 244(4901):160.
[24] Wasserstein, R. L. and Lazar, N. A. (2016). The ASA’s statement on p-values:
Context, process, and purpose. The American Statistician, 70(2):129–133.
[25] Tibshirani, R. (2011). Regression shrinkage and selection via the LASSO: A
retrospective. Journal of the Royal Statistical Society: Series B (Statistical
Methodology), 73(3):273–282.
[26] Press, W. H., Flannery, B. P., Teukolsky, S. A., Vetterling, W. T., et al. (1989).
Numerical Recipes, Vol. 3. Cambridge: Cambridge University Press.
[27] Silverman, J. G., Raj, A., and Clements, K. (2004). Dating violence and associated
sexual risk and pregnancy among adolescent girls in the united states. Pediatrics,
114(2):e220–e225.
[28] Russell, S. J. & Norvig, P. (1995). Artificial intelligence: A modern approach.
Englewood Cliffs, NJ: Prentice Hall.
[29] James, G., Witten, D., and Hastie, T. (2013). An introduction to statistical
learning: With applications in R. New York: Springer.
[30] Jose, C., Goyal, P., Aggrwal, P., and Varma, M. (2013). Local deep kernel learning
for efficient non-linear svm prediction. In International Conference on Machine
Learning, pp. 486–494.
[31] Flannery, D. J., Singer, M. I., and Wester, K. L. (2003). Violence, coping, and
mental health in a community sample of adolescents. Violence and Victims,
18(4):403–418.

https://doi.org/10.1017/9781108669016.013 Published online by Cambridge University Press

 12
SHIHbot
Sexual Health Information on HIV/AIDS, chatbot

Joshua Rusow, Jacqueline Brixey, Rens Hoegen, Lan Wei,

Karan Singla, and Xusen Yin

Motivating Social Problem

Human immunodeficiency virus (HIV) is an incurable virus that leads to
chronic illness – and the potentially fatal disease, acquired immune deficiency
syndrome (AIDS) [1]. HIV is currently incurable because the human immune
system has not adapted to completely rid the body of the virus after it
takes hold, and neither medicine is available that can completely eradicate
it, nor has a vaccine been developed. HIV uses the body’s immune system
as its mechanism of attack, infecting and reducing the number of CD4 cells
(also called T cells), leaving infected individuals at risk for other infections
that would otherwise be easily fought off [1]. These so-called opportunistic
infections are potentially fatal, and lead to the distinction between HIV and
AIDS. While no cure exists, there are several effective prevention techniques
available at reducing or eliminating the rates of future HIV infection [2].
However, for many prevention techniques to work, individuals need to know
if they are currently HIV-infected, to select the most appropriate prevention
strategy. The only way to know for certain if one is HIV-infected is to be tested
for HIV antibodies [1].
HIV/AIDS develops in the body of most people in a similar way, regardless
of gender, sexual orientation, or racial/ethnic identity. However, the prevalence
of HIV among the general US population is uneven. For example, while gay,
bisexual, and other men who have sex with men (hereafter referred to as
MSM) make up approximately 2% of the US population, in 2013 they made
up 55% of individuals living with HIV [2]. In terms of new diagnoses in 2014,
MSM accounted for 83% of new diagnoses among males over the age of 12
years, 67% of all new diagnoses, and 92% of diagnoses of male youth ages
13–24 years [2]. Stark racial/ethnic differences also exist, with black MSM
accounting for the highest number of new diagnoses in 2014, despite their

211

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 212 SHIHbot

relatively lower makeup of the general population, and Hispanic/Latino MSM

showing the greatest increase in new diagnoses when compared to rates in
2010 [2].
These figures show that the highest HIV burden in the United States falls on
gay, bisexual, and other MSM. Since this group is somewhat insular when it
comes to finding sexual partners, gay and bisexual men are also at an increased
chance of HIV exposure. Overall estimates place the prevalence of HIV
infection among gay and bisexual men in the United States at 15%, or around
1-in-7; however, only approximately 73% of those infected actually are aware
of their HIV status through diagnosis [2]. Without diagnosis information, HIV-
infected people are unable to access medication that can vastly improve their
health, while also reducing their risk of passing HIV on to their sexual partners.
HIV-negative individuals can also take a daily medication called pre-exposure
prophylaxis (PrEP) to prevent becoming HIV-infected, but again, medical
intervention is required. Since anal intercourse with an HIV-infected person
when condoms or proper medication are not present is the most likely way
to acquire HIV in this group, knowledge about transmission and prevention
methods is paramount. Condoms, in particular, can prevent other STDs like
chlamydia, gonorrhea, and syphilis, for which MSM are also at increased
risk [2]. However, stigma and/or discrimination, particularly in the form of
homophobia, may prevent MSM from accessing providers for support in
these areas.
As HIV is a chronic, lifelong illness, and the rates among gay, bisexual, and
other MSM are most pronounced in youth aged 13–24 years, reaching these
young men who have sex with men (YMSM) with prevention and treatment
options at a young age is vital. Despite efforts to prevent early sexual debut,
nearly two-thirds (64.1%) of American youth have had sex by the 12th grade
[3]. Additionally, of the approximately 20 million newly diagnosed sexually
transmitted infections each year, half occur among individuals between the
ages of 15 and 24 [4]. The behaviors that put youth at risk for these unintended
consequences are further pronounced among sexual- and gender-minority
(i.e., lesbian, gay, bisexual, transgender, questioning and queer [LGBTQ])
youth [5, 6].
Many schools implement some form of sexual health education, but when
asked, youth indicate that these programs do not meet their needs [7]. Students
report that sex education often provides technical but not practical information,
causing them to seek out additional resources to supplement the gap [7]. Many
students report being uncomfortable discussing sensitive topics with their
teachers in the classroom, who play an ongoing role outside of sexual health

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 Joshua Rusow et al. 213

education instruction, or who might inform their parents about their inquiries
[7]. Access to medically accurate information and/or confidential, individual
counseling has been shown to decrease negative consequences by increasing
condom and contraception use and delaying sexual debut while encouraging
abstinence [8, 9].
Students and other youth may respond better to culturally competent [10]
and confidential sexual health counseling if it were made available on demand
and throughout the year. Youth may not be currently sexually active [11], or
may not have developed questions about sexuality, during the prescribed times
that sexual health education classes are offered (if they are offered at all).
Perhaps some LGBTQ youth are concealing their sexual or gender identity
because of victimization [12] fears and thus may not be comfortable discussing
their unique sexual health needs in a group setting.
Disseminating sexual health information is key in curbing the spread of
HIV, but youth may be embarrassed to seek information from sources that they
would normally go to for medical or other advice. A plethora of information
exists online, but the quality can be very hard to assess, particularly for young
people who might not have much exposure to the topic. A chatbot that utilizes
artificial intelligence (AI) to train on only information vetted to be medically
accurate can provide information to users in an anonymous fashion, while
keeping anonymized logs that can be used to further improve the usability
of the virtual chat agent. This chapter goes on to explain the conceptualization
and implementation of the sexual health information on HIV/AIDS chatbot,
or SHIHbot. First, we discuss the architecture of SHIHbot, followed by the
strategy we used to collect medically accurate information, as well as some
example questions from actual users to train the system. We then describe
NPCEditor, which selects the best response to user questions, after which we
discuss our deployment and evaluation strategies.

Overview
In order to inform our target group of YMSM with reliable information on
HIV/AIDS, we propose a system that employs a chatbot to dispense this
information. The benefit of using a chatbot is that people often feel less
judged when interacting with a virtual agent than with a real person [13].
The system will also be readily available to users all the time, allowing
users to receive answers faster than seeking answers from a professional. We
hypothesize that a system using a virtual agent in the form of a chatbot will

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 214 SHIHbot

be an efficient way of informing YMSM on a subject that is often considered

taboo.
Although a lot of the information can be obtained by simply searching
the Web, some of this information might contain factual errors, such as
information coming from open forums. Therefore, a second benefit of using
a more closed environment like the chatbot is that this system can be
programmed to use only data from reliable sources.
In the following sections we provide explanations of the general architec-
ture of the system (Figure 12.1). Starting by describing our data-gathering
process, we then present how this data is preprocessed and formatted. The

Figure 12.1 High-level overview of the architecture. We first find Web sources
that have correct answers to generally asked questions about HIV/AIDS for
YMSM. Data crawler automatically extracts data from various sources and brings
them to the standard XML format. Data formatter and cleaner further cleans
and prunes the repetitive data according to question topics. Data are then fed to
NPCEditor, which learns to interpret queries given by the user and interacts with
the system and then to select an appropriate response. The end user only interacts
with the system interface to get answers to their questions related to HIV/AIDS.

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 Joshua Rusow et al. 215

preprocessing also involves merging semantically similar questions into one

topic. In the following section, we describe the use of NPCEditor and then
talk about the interface provided to the end user. NPCEditor also allows the
system to improve based on user questions and statements. Finally, we describe
deploying SHIHbot on Facebook Messenger, and our offline evaluation results.

Data Gathering
To build an AI model for answering questions about HIV/AIDS, a database
containing proper questions and answers should be available for reference.
The sources for both the questions and answers are critical. The qualities of
the questions will directly affect the understandability of this model for all
potential questions asked by our users. The answers provided through our
model should be verified for accuracy. It is preferable that the interaction
created by our model does not make our users feel uncomfortable.
The question sources provide coverage for potential questions coming from
the ultimate service users. The database of questions will focus on three
perspectives: professional knowledge, language habits, and synonyms.
First, users come to our platform in need of professional knowledge (e.g.,
“whether or not AIDS can be cured”; “what is the relationship between
hepatitis and HIV infections?”). The questions should cover a range of
profession keywords for the model to capture the most effective information,
such as “relationship,” “hepatitis,” “HIV/AIDS,” “cure,” etc.
Second, users of the platform have different habits when it comes to asking
questions, e.g., some users will not invert the subject and the verb when asking
questions. Some users tend to ask very subjective questions, so instead of
asking “Is AIDS curable?” they might ask, “Am i going to die because i am
infected with AIDS?” There are any number of ways to ask questions that
might lead to the same correct answer.
Third, users of the platform might be of different cultural backgrounds,
different ages, or have differing levels of knowledge about medication, health
care, or diseases. For example, people with medical research backgrounds will
refer to the white blood cells that send signals to activate the body’s immune
responses as CD4 cells, but other people might call them T cells, or T-helper
cells. Another example is that users may refer to semen as sperm or vice
versa, or some of them will not recognize semen but use cum. By increasing
the number of questions relating to an answer we can address the variety in
language habits of users.

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 216 SHIHbot

To achieve the professional knowledge coverage for HIV/AIDS informa-

tion, we use questions and answers organized by federal agencies and institutes
and departments of health from states. Since these organizations focus on
developing and applying HIV/AIDS control and prevention nationally or
within a whole state for a large range of different population over years,
we trust them to be a credible source of information. Two example question
sources we use are the HIV/AIDS question and answer section of Centers for
Disease Control and Prevention (CDC; https://www.cdc.gov/actagainstaids/
basics/whatishiv.html) and New York State Department of Health (https://
www.health.ny.gov/publications/0213.pdf).
To improve the source of different forms of questions considering the
different habits of language use, we also use sources from non-government
or non-federal institution. In smaller forum sites built by volunteers and
researchers, the questions asked are more personal and very likely to appear
when users are chatting with a chatbot. At the current phase, we collect these
personal questions from forums like i-base (http://i-base.info/qa).
To identify relevant synonyms and similar phrases for key terms, we uses
sources from CDC and other credible institutions’ records. These records
usually use parentheses to mark a word or a phrase with different usages,
such as, Polyurethane (plastic) or polyisoprene (synthetic rubber) condoms
are good options for people with latex allergies, but plastic ones break more
often than latex ones.
The sources of our answers all come from credible institutions like the CDC
and the New York State Department of Health. We have made this a limited
source to make sure the answers will not harm or distress the users. In future
iterations we may add information from other trustworthy sources.
To obtain the data from our resources, we used Web-scraping techniques.
Using a Web crawler to crawl the websites for the required data and then format
it to specifically formatted comma-separated values (CSV) files containing
question-answer pairs to be used by the chatbot. In addition to the professional
information websites of the CDC and the New York State Department of
Health, we obtained data from the i-base website. We obtained over 3,000
question-answer pairs from these websites, allowing us to address a wide
variety of topics on HIV and AIDS.

NPCEditor
To drive our chatbot responses we used NPCEditor, a response classifier
and dialogue management system [14, 15]. NPCEditor employs a statistical

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 Joshua Rusow et al. 217

Figure 12.2 Example of potential input statements and responses. The connecting
lines between the question panel and the answer panel indicate to which responses
in the answers an input in the questions list is connected.

classifier that is trained on linked questions and responses; these links are
visualized in Figure 12.2. For each new user utterance, the classifier ranks all
the available responses; the full details of the computations can be found in
Leuski and Traum [14]. We trained the classifier on our QA corpus, which is
augmented by questions and responses about the chatbot itself and utterances
that maintain dialogue flow such as greetings and closings (Table 12.1).
The dialogue manager functionality within NPCEditor chooses which
response generated by the classifier to give to the user. Typically it will choose
the response that was ranked highest by the classifier, but it may choose a
lower-ranked response in order to avoid repetition. In the event of multiple
responses scoring above the threshold, the returned response will be the one not
given most recently, and/or the response with the most relevant answer words
in the response. Examples of scores can be seen in Figure 12.3. If the score
of the top-ranked response is below a predefined threshold that is determined
during training, the dialogue manager will instead select an off-topic response
that indicates non-understanding (such as “please repeat that”).
The classifier also has special tokens to recognize when a user asks the
chatbot to repeat an answer or elaborate on a previous answer, and when such
a token is identified, the dialogue manager will repeat or elaborate, based
on the topic annotation of the responses. For example, should a user ask for
more information after receiving an answer tagged with “transmission,” the

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 218 SHIHbot

Table 12.1. Each response was tagged with a “type” and “topic” in
NPCEditor to help direct the dialogue manager

Type Purpose Example dialogue

Don’t know No relevant response in User: What is the weather in LA?

domain above threshold SHIHbot: I am not confident on an
answer for that question.
Personal Questions about the chatbot User: are you a robot?
SHIHbot: I am a computer, but I have
lots of human colleagues who helped
make me so smart.
Greeting Greetings and pleasantries User: hello
SHIHbot: Hi there! Do you have any
questions about HIV/AIDS?
Alternative Suggests a new User: will it rain today?
conversation topic after 3 SHIHBot: I’m not sure on that. I would
off-topic inputs from user be happy to answer any questions you
have about HIV/AIDS though.
Say more Gives a random response User: Should I get tested even if I do not
with the same topic as the have symptoms?
previous response SHIHBOT: If you think that you may
have been exposed to HIV, you should
get tested as soon as possible. Someone
can have HIV and have no symptoms for
many years.
User: tell me more about this topic.
SHIHBot: The CDC recommends that
everyone between the ages of 13 and 64
get tested for HIV at least once as part of
routine health care. About 1 in 8 people
in the United States who have HIV don’t
know they have it.
Repeat Repeats previous response
QR In-domain responses
annotated with topics

system would then give the next highest scoring answer with that same topic.
A counter keeps track of the number of consecutive times the chatbot has failed
to provide a direct answer, and on the third instance, an “alternative” response
is given to suggest returning to the HIV/AIDS domain. The counter restarts
after giving an “alternative” response (Figure 12.4).
Previous applications of NPCEditor have been used to drive interac-
tive characters in various domains such as interactive museum guides [16],
entertainment experiences [17], and interviews with Holocaust survivors [18].

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 Joshua Rusow et al. 219

Figure 12.3 Example of how potential responses while chatting are scored based
on input from the user. Panel 6 (the rightmost panel) shows the total score of
the response chosen (in this case above the threshold for an acceptable response
to return). Panel 5 (the middle panel) shows which words in the response led to
the overall score and gives insight into which words were most important for the
response being chosen.

Figure 12.4 Example of a classifier’s performance in NPCEditor

NPCEditor was applied to the HIV/AIDS domain in the development of a

virtual reality application designed for HIV-positive young men who have sex
with men (YMSM), to practice disclosing their status to intimate partners in
an immersive, nonjudgmental environment [19].

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 220 SHIHbot

Deployment
Deployment of the SHIHbot provides a real-world platform for users to interact
with it. Through it we can also collect more questions to expand the Q&A
database and further train NPCEditor. A version of SHIHbot was demonstrated
at Sigdial 2017 [20].
Many Web platforms provide programming interfaces for third-party devel-
opers to integrate their own applications. Among of them, several popular
ones include Facebook Messenger, Slack, Twitter, and PandoraBots. Facebook
Messenger was selected as our deployment platform as it currently has more
active and age-appropriate users and developed interfaces.
Our deployment involves three layers: the Messenger, the NPCEditor, and
a proxy Web service as an intermediate layer to isolate the user interface with
chatbot implementations (Figure 12.5).

Messenger as the User Interface

To deploy a chatbot on Facebook, we needed first to register a developer
account on Facebook, then add a new product, which we called SHIHbot, to
start our deployment.
To add a new role of a chatbot in the Messenger, we first create a new
Facebook page with the name “SHIHbot.” People who talk to the SHIHbot
are actually talking to the Facebook page. Then we add the page in our
SHIHbot product and apply for webhooks to the page with some of its events,
e.g., messages, message_deliveries, and message_reads. Facebook developer
platform will forward all applied events to the application.

Figure 12.5 SHIHbot deployment architecture. Users interact with the chatbot via
the Facebook page. Typed messages are passed via a message proxy to the HTTP
plugin within NPCEditor. Once the message has been received by NPCEditor,
the dialogue manager determines an appropriate response. The message is then
passed back through the proxy to show up as a response on Facebook messenger.

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 Joshua Rusow et al. 221

Intermediate Web Service as the Proxy

To receive the events mentioned above, we need to build a Web service with a
URL, which can accept RESTful requests as a callback. Then we register the
URL as a callback URL in the webhook product of the Facebook developer
platform. After this registration, the Facebook developer platform will forward
all events applied to the Web service. We use the Python Flask package to
build a test purpose Web service and hold it in the Heroku platform, which
provides a fast way to deploy a Web service. The intermediate layer can either
reply to simple questions from Messenger, like “what is your name” and reply
with “I’ll be back in a moment” if the backend chatbot is down, or forward
messages it received to any kind of chatbot backends.

NPCEditor as the Backend

The NPCEditor has many plugins to provide versatile interfaces for users. For
example, the e-mail plugin supports interacting with the NPCEditor via e-mail,
while http plugin provides RESTful interface. It also contains phone plugin,
Jabber plugin, and so on.
We use its http plugin to connect with the intermediate layer. First we create
a new user in the NPCEditor with a name “SHIHbot,” then we add a http
account for the user with a port number. After that, we can interact with the
SHIHbot user via the URL http://localhost:port/SHIHbot/1 to post questions
to it, e.g., http://localhost:port/SHIHbot/?q=”what is your name”&from=
”user-id”
Broken down, the above URL shows the user “user-id” asking the question
“what is your name” to the user “SHIHbot.” We use JSON to compose the
question and user id content in the intermediate layer, then forward the message
to the NPCEditor. After receiving the reply from NPCEditor, the intermediate
layer then forwards the response back to Messenger.

Evaluation Metrics
We measure the performance of SHIHbot to better serve users in three aspects:
the real-time interaction, the long-term engagement, and the learning and
improvement of SHIHbot over time. Accordingly, we use linguistic-driven,
online-driven, and social-driven metrics to measure our bot.

1
A public URL is needed for the NPCEditor for connecting it with the intermediate layer.

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 222 SHIHbot

Linguistic-Driven
Given the huge range of possible user input, chatbots often misinterpret or are
unable to understand what a user wants. The times when a bot responds with
a version of “I don’t understand” highlights the bot’s interpretation ability and
a potential lack of responding item in the database. Such incidences might
upset users.
However, there is no absolutely versatile chatbot, so to test the percentage of
“I don’t understand,” the question test set must be selective of reasonable ques-
tions. A robust chatbot serves reasonable “I don’t understand” to unreasonable
questions.
Conversation steps represent another metric for interpretation. A single
back-and-forth exchange between a user and a bot marks one conversation
step. For example, if a user asks one question, and the chatbot asks two
subquestions and then delivers the final answer, it means there are three
conversations steps.
The definitions are as follows,
Confusion Trigger Rate (CTR): The percentage of answers returned as “I
don’t understand,” which happens when the bot fails to find any appropriate
reply to a question.
Conversational Steps: Number of subquestions raised by bot following an
original question. This shows the bot’s robustness in engaging a user in a more
natural conversation and giving an appropriate answer.

Online-Driven
As with all online services, we expect to measure the retention: the ratio of
long-term users to registered users for our chatbot.
Also, we measure the engagement rate for each user. Each time a user
responds to the chatbot is considered an engaged session.

Social-Driven
A customer satisfaction questionnaire is planned for our chatbot when users
do not respond with the chatbot but leave the website open. A prompt will
ask the user a concluding question, such as, “Was I able to answer your
questions today?” Users will be prompted to send feedback through the chatbot
if they wish.
To research on correctness and social acceptability, we can ask human
annotators to annotate each answer for a question in terms of user acceptability.

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 Joshua Rusow et al. 223

Each annotator marks a reply for a question in terms of two metrics, Comfort
and Trust. Both will be measured on a Likert-like scale of 0 to 5, where
0 is “not at all” and 5 is “completely.” Comfort pertains to whether or not
the answer provided is sensitive to the way the question was asked; trust is
whether or not the answer provided was accurate. Meanwhile, further research
can determine if SHIHbot is effective in reducing HIV-exposure risk.

Alpha Test
The current version of evaluation is a test of the alpha version of SHIHbot,
which means it is only open to insiders offline. We expanded the Confusion
Trigger Rate (CTR) into the a scale from 0 to 5 (see Table 12.2). We use human
experts available to rate our question data set. The question data set we use is
from a government source; in this way, we make sure all these questions are
already supplied with a credible answer that our human experts can evaluate.

Data
We use two different data sets to evaluate the alpha version of SHIHbot. The
first data set, which we call “va38q,” consists of 38 questions extracted from the
Frequency Asked Questions section of the US Department of Veterans Affairs
(https://www.hiv.va.gov/HIV/patient/faqs/index.asp). The second data set con-
sists of 100 questions on the subject of HIV/AIDS, published by the New York
State Department of Health (https://www.health.ny.gov/publications/0213.pdf),
which we call “ny100q.” Feeding these questions into the SHIHbot, we get
138 answers related to these questions. Then we evaluate our Q&A pairs using
the rubric provided in Table 12.2.

Observations
Figures 12.6 and 12.7 show three consistent cumulative distribution results
from our four human experts. For both data sets, the distribution of scores
does not differ greatly between the human experts, implying agreement in
interpretation of the overall performance of SHIHbot. We see in Figure 12.6
that human expert 2 has a slightly different scoring distribution than the other
three, which can be attributed to the smaller sample size of questions in that
dataset. Figure 12.7, which has nearly three times the sample size and the same
human experts, has much more consistent scoring.

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 Table 12.2. Explanations and examples of rating

Rating Explanation Case (Q&A pairs from ny100q)

0 returns “I don’t understand,” question 17: Does sexual contact with many partners increase my risk of getting HIV?
“I’m still learning,” or similar answer 17: I’m not sure! I think Google might have a better answer.
answers
1 returns an answer other than “I question 21: Can a person with HIV who has an undetectable viral load pass HIV to someone
don’t understand,” but unrelated else?
to the topic of the question answer 21: It is okay for HIV-positive people to date anyone positive or negative. There are
no additional risks if you are both positive. Sometimes it means you have a shared
understanding about HIV, and this can be good.
2 returns a misleading answer question 7: Where can I find updated statistics on HIV and AIDS?

224
answer 7: There are a number of resources you can use to find this information. The first
resource is our guide to ARVs. The guide lists ARVs currently available. It also provides a
one-page summary on each drug. This includes information such as generic and brand drug
names, and brief characteristics and potential side effects of each drug. If you would like

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

information on managing side effects of the different drugs, please follow this link to our
guide on HIV and your quality of life. Other resources you can use are:
The European Medicines Agency website. The site provides information in two formats. The
first is patient information in nontechnical language and the second is more technical drug
product information.
The European Medicines Compendium Website. This site provides a concentrated technical
summary on each drug. Also know as the SPC [Summary of Product Characteristics], it
includes information on how the drug works; how it was studied; side effects;
contraindications (when it should not be used); and why it was approved.
 The NHS (National Health Service) website.
The American National Institute for Health website.
The third and most personal option is to talk to your HIV clinician or pharmacist. This option
is a good opportunity to ask direct questions on anything you’re not too clear about.
3 returns an answer that only question 1: Who is at risk for getting HIV?
partially answers the question answer 1: Case reports of female-to-female transmission of HIV are rare. The
well-documented risk of female-to-male transmission shows that vaginal fluids and
menstrual blood may contain the virus and that exposure to these fluids through mucous
membranes (in the vagina or mouth) could potentially lead to HIV infection.
4 returns an answer covering most question 9: Is it easy to get HIV?
of the question answer 9: It sounds as if you would benefit from some professional counseling about why

225
you are so concerned about getting HIV. With today’s modern drugs, someone with HIV can
lead a normal life with a near-normal life expectancy so there is really nothing to worry
about. HIV is not that easy to catch, especially in the UK where the prevalence is very low
and where HIV-positive people have access to treatment making them noninfectious. You

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

cannot get HIV from handling a parcel and you cannot get it from a needle that has been
lying around for 4 hours.
5 returns an exactly correct and question 4: Is there a cure for HIV or AIDS?
related answer answer 4: No effective cure currently exists for HIV or AIDS. But with proper medical care,
HIV can be controlled.
 226 SHIHbot

Figure 12.6 CDF of ratings of the va38q

Based on rating standards in Table 12.2, the start points of lines in Figures
12.6 and 12.7 mean the experts giving the 0 rating, which shows there is a
lack of knowledge for SHIHbot to return the good answers. The current results
show that there are about 20–50% answers for which SHIHbot has no related
records. This lack of information will be addressed after SHIHbot is activated
online, and information can be added according to collections of unsolved
questions.
If we exclude the starting points from Figures 12.6 and 12.7, both graphs
show a greater slope from rating 4–5 than from rating 2–3, which indicates that
when SHIHbot has information about the questions that users are asking in the
referral database, it uses it to create a satisfactory answer.

Future Work and Directions

Currently, SHIHbot is only in its first iteration. Although SHIHbot is a
fully functional chatbot capable of answering questions on HIV and AIDS,

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 Joshua Rusow et al. 227

Figure 12.7 CDF of ratings of the ny100q

we envision the following improvements that will improve its capability of

reaching its target audience and providing information.

Support for Additional Platforms

SHIHbot currently supports the Facebook Messenger platform. Although
Facebook is widely used, having the chatbot deployed on only one platform
will limit the number of people we can reach. Additionally some people might
not want to use the chatbot on Facebook, as the platform technically allows
us to gather personally identifiable information. Although we do not track or
store any of these data, it is hard to convince skeptical users of this.
One additional platform to deploy SHIHbot on is Twitter. Users could
privately send direct messages to the chatbot. Another option is to build our
own website that allows its users to interact with the chatbot. That would
enable users without a Facebook or Twitter account to interact with the bot.
Similar to a separate website would be to build an app for the iOS and Android

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 228 SHIHbot

operating systems giving users a dedicated app for obtaining information on

HIV/AIDS on their smartphones.

Obtain Additional Information

SHIHbot currently provides information on HIV/AIDS using information
obtained from the Centers for Disease Control and Prevention (CDC), the New
York State Department of Health, and the Q&A section of the i-base website.
We can improve the bot by obtaining new information from additional websites
(e.g., aids.gov, Scarleteen, Go Ask Alice).
We also can broaden our search for information on sexual health in general.
The current version of SHIHbot only answers questions directly relating to
HIV and AIDS. However, it would make sense to create additional question
categories on topics such as other STDs, sexual orientation, and gender
identity.

Add Additional Functionality

SHIHbot is currently only capable of answering a single question. This means
that it does not keep track of the entire conversation it has been having with the
user. We can improve this behavior by tracking the history of a conversation.
History would allow SHIHbot to give more appropriate answers and also to
allow its users to ask follow-up questions on some topics. By doing this,
chatting with SHIHbot will feel more natural, perhaps enticing the user to also
spend more time and ask more questions.

Analyze User Data to Improve SHIHbot

SHIHbot can be continuously improved by analyzing the anonymized user
data. A review of the questions being posed to the system and the answers
provided offer a metric of how well the system is performing and also indicate
areas that need to be strengthened. By gaining a better understanding on what
users are interested in, it will be possible to better address their questions and
concerns. In addition to the anonymous user data within the system, users will
also have the ability to provide feedback directly to the system, or through
usability questionnaires and feedback forms available on the platform they
are using.
Overall, the system can be iteratively developed and improved to provide
a satisfying user experience. While the topics tackled by the system are

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 Joshua Rusow et al. 229

potentially sensitive, they also have great public health and social justice
implications. It is our hope that SHIHbot will move online and provide the
most vulnerable – or the otherwise curious – the best information to reduce
HIV transmission.

References
[1] US DHHS (Department of Health and Human Services). (2016). HIV/AIDS 101:
What is HIV/AIDS? Retrieved from www.aids.gov/hiv-aids-basics/hiv-aids-101/
what-is-hiv-aids/
[2] CDC (Centers for Disease Control and Prevention). (2017). HIV among gay and
bisexual men. Retrieved from www.cdc.gov/hiv/group/msm/index.html
[3] Kann, L., et al., Youth risk behavior surveillance – United States, 2013. MMWR
Surveill Summ, 2014. 63(Suppl 4): 1–168.
[4] Centers for Disease Control and Prevention. Incidence, prevalence, and cost
of sexually transmitted infections in the United States. 2013; Available from:
www.cdc.gov/std/stats/.
[5] Robinson, J. P. and D. L. Espelage, Peer victimization and sexual risk differences
between lesbian, gay, bisexual, transgender, or questioning and nontransgender
heterosexual youths in grades 7–12. American Journal of Public Health, 2013.
103(10): 1810–1819.
[6] Mojola, S. A. and B. Everett, STD and HIV risk factors among US young
adults: Variations by gender, race, ethnicity and sexual orientation. Perspectives
on Sexual and Reproductive Health, 2012. 44(2): 125–133.
[7] DiCenso, A., V. W. Borthwick, and C. Creatura, Completing the picture: Adoles-
cents talk about what’s missing in sexual health services. Canadian Journal of
Public Health, 2001. 92(1): 35.
[8] Kirby, D., et al., School-based programs to reduce sexual risk behaviors: A review
of effectiveness. Public Health Reports, 1994. 109(3): 339.
[9] Underhill, K., D. Operario, and P. Montgomery, Systematic review of abstinence-
plus HIV prevention programs in high-income countries. PLOS Medicine, 2007.
4(9): e275.
[10] Hoban, M. T. and R. L. Ward, Building culturally competent college health
programs. Journal of American College Health, 2003. 52(3): 137–141.
[11] Cavazos-Rehg, P. A., et al., Age of sexual debut among US adolescents. Contra-
ception, 2009. 80(2): 158–162.
[12] D’Augelli, A. R., N. W. Pilkington, and S. L. Hershberger, Incidence and mental
health impact of sexual orientation victimization of lesbian, gay, and bisexual
youths in high school. School Psychology Quarterly, 2002. 17(2): 148.
[13] Gratch, J., Lucas, G. M., King, A. A., and Morency, L. P. (2014, May). It’s
only a computer: the impact of human-agent interaction in clinical interviews.
In Proceedings of the 2014 International Conference on Autonomous Agents
and Multi-Agent Systems (pp. 85–92). International Foundation for Autonomous
Agents and Multiagent Systems.

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 230 SHIHbot

[14] Leuski, A., and D. Traum, NPCEditor: Creating virtual human dialogue using
information retrieval techniques. Ai Magazine, 2011. 32(2): 42–56.
[15] Leuski, A., and D. R. Traum, NPCEditor: A tool for building question-answering
characters. Proceedings of the Seventh conference on International Language
Resources and Evaluation. L’REC 2010. pp. 2463–2470.
[16] Swartout, W., et al. Ada and Grace: Toward realistic and engaging virtual museum
guides. International Conference on Intelligent Virtual Agents. Springer Berlin
Heidelberg, 2010.
[17] Hartholt, A., J. Gratch, and L. Weiss, At the virtual frontier: Introducing Gun-
slinger, a multi-character, mixed-reality, story-driven experience. International
Workshop on Intelligent Virtual Agents. Springer Berlin Heidelberg, 2009.
[18] Traum, D., et al. New dimensions in testimony: Digitally preserving a Holocaust
survivor’s interactive storytelling. International Conference on Interactive Digital
Storytelling. Springer International Publishing, 2015.
[19] Knudtson, K., K. Soni, K. Muessig, M. Adams-Larsen, R. Artstein, et al. Tough
talks: Developing a virtual reality application to support HIV status disclosure
among young MSM. 21st International AIDS Conference (AIDS 2016) Abstract
Book, 2016.
[20] Brixey, J., R. Hoegen, W. Lan, J. Rusow, K. Singla, X. Yin, R. Artstein and
A. Leuski. SHIHbot: A Facebook chatbot for Sexual Health Information on
HIV/AIDS. Proceedings of the 18th Annual SIGdial Meeting on Discourse and
Dialogue, 2017.

https://doi.org/10.1017/9781108669016.014 Published online by Cambridge University Press

 13
Ethics and Artificial Intelligence in Public
Health Social Work
David Gray Grant

Introduction
As demonstrated by a number of chapters in this volume, autonomous software
agents based on artificial intelligence (AI) have many potential applications
at the intersection of public health and social work, known as public health
social work. These applications are enormously promising, but often pose
novel ethical problems for researchers that can be difficult to assess, much
less to resolve. The aim of this chapter is to show how analytical tools from
moral philosophy and theoretical computer science can be combined to better
understand these problems and to develop strategies for addressing them in
practice.
I will focus on a specific set of problems that arise in the development of
public health social work interventions based on AI. Specifically, I will focus
on a class of problems that I call beneficence problems (the rationale behind
the name will become clear further on). Beneficence problems occur in the
context of public health social work interventions that are partially planned, in
the field, with the help of an artificially intelligent autonomous software agent
(call these AI planning interventions).
Consider cases such as the following. Suppose that researchers are devel-
oping a public health social work intervention to be conducted by some
intervention team with some target population. The goal of the proposed
intervention is to provide some specific set of benefits to the target population,
such as (for example) reducing the incidence of HIV in the population. In
the field, a software agent will assist the intervention team by recommending
an intervention plan, consisting of some sequence of actions to be performed
by the intervention team. The agent’s goal is to identify and recommend the
intervention plan that, if carried out, would benefit the target population in the
intended way to the greatest extent possible. The intervention team believes

231

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 232 Ethics and Artificial Intelligence in Public Health Social Work

that the intervention plans that will be recommended by the agent are likely
to benefit the target population, considered as a whole, to a greater extent than
would alternative interventions they could perform. However, they also believe
that the software agent may, under some foreseeable conditions, recommend
a plan that, while well-suited to maximizing benefits to the target population,
will also pose a significant risk of harm to some of that population’s members.
Moreover, they have the capability to predict the individual- and population-
level expected benefits and harms of particular intervention plans the agent
considers. What should the researchers do?
The answer is far from obvious. Cases such as this one pose a dilemma
between two moral duties whose importance is widely recognized in the
fields of both public health and social work. The first is the duty not to
conduct interventions that are expected to harm others, otherwise known as
the duty of non-maleficence. The second is the duty intervention teams have
to conduct interventions that they expect will provide the greatest possible
benefits to the populations they work with, otherwise known as the duty of
benefit maximization. The challenge posed for the design of the autonomous
agents involved in such interventions is how to ensure that the intervention
plan that is eventually conducted by the intervention team will demonstrate
appropriate respect for both duties. Call the problem of meeting this challenge
in a particular AI planning intervention a beneficence problem, and call an AI
planning intervention that poses a beneficence problem an AI intervention*.
(The asterisk next to this term throughout the chapter indicates that it is
different from a standard AI planning intervention.)
As we shall see, it can be difficult in practice to modify the software
agent used in interventions of this kind so that there are reasonable guarantees
that both duties just mentioned will be respected. That is, it can be difficult
to ensure that the intervention plans the agent recommends will both (1)
minimize the expected risks of the intervention to a degree compatible with
the duty of non-maleficence and (2) maintain the expected benefits of the
intervention at a level consistent with the duty of benefit maximization. On
the one hand, if the balance researchers strike between minimizing expected
harms and maximizing expected benefits weights expected benefits too heavily,
favoring the duty of benefit maximization, then the risks posed to particular
individuals may be too significant to be morally justifiable. On the other hand,
if the balance struck weights expected harms too heavily, favoring the duty
of non-maleficence, then the intervention may be rendered ineffective enough
to obligate researchers to abandon the proposed intervention in favor of a
more beneficial alternative. Either kind of failure would render the proposed

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 David Gray Grant 233

intervention morally impermissible to conduct, in the absence of further

modifications.
My ultimate objective in what follows is to provide practical guidance about
how to approach designing the software agents used in AI interventions* so
that both extremes are avoided. The plan for the chapter is as follows. To make
the discussion more concrete, I begin in the next section by setting out a real-
world beneficence problem currently being faced by researchers at the Center
for Artificial Intelligence in Society at the University of Southern California.
In the third section, I provide a more formal definition of beneficence prob-
lems and set out some necessary technical background. Since implementing
appropriate safety constraints requires a working understanding of the moral
duties involved, the fourth discusses the duties of non-maleficence and benefit
maximization in more detail. The penultimate section proposes a framework,
offered in a tentative spirit, for addressing conflicts between the two duties
in public health social work interventions. I conclude in the final section by
sketching some potential strategies for operationalizing this framework in the
context of AI planning interventions.

Case Study: Adapting TND Network for Homeless Youth

Researchers at the Center for Artificial Intelligence in Society (CAIS) are in
the process of adapting a drug abuse prevention program called TND (Towards
No Drug Abuse) Network for use in residential shelters for homeless youth.
The program divides participants into small groups of approximately five
members. The goal of the program, which relies heavily on peer interaction,
is to use the power of peer influence to help participants develop the attitudes,
skills, and confidence required to resist drug abuse. This program has been
demonstrated to be effective in schools with at-risk youth – an important
result, given how difficult drug abuse has proven to address – but current
versions have a significant downside. In a recent large-scale trial of TND
Network, Valente et al. (2007) found that particularly at-risk participants failed
to benefit from the program. In fact, the program seemed to place them at
even higher risk of future drug abuse. Valente and colleagues hypothesized
that this increased risk was the result of deviancy training, a well-documented
phenomenon that occurs when individuals are encouraged by their peers to
adopt harmful or antisocial behaviors.1 When such high-risk youth are placed
together in an intervention group, Valente and colleagues concluded, they are
likely to influence each other to use substances more, not less.

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 234 Ethics and Artificial Intelligence in Public Health Social Work

Researchers at CAIS are currently attempting to modify the program in

two ways.2 The first is to make it more suitable for homeless youth at
high risk of serious drug abuse living temporarily in residential homeless
shelters. The second is to address its deviancy training effects. To achieve the
second goal, CAIS developed an autonomous software agent to help improve
how the program divides intervention participants (the current residents of a
medium-term residential shelter for homeless youth) into groups. The agent’s
goal is to maximize the intervention’s positive effects of peer influence on
participants’ future drug use, and to minimize its negative (deviancy training)
effects.
To do this, the agent uses a peer influence model (and more specifically, a
linear threshold model; see Kempe et al., 2003) to predict how attitudes about
drug abuse will spread from person to person as a result of the intervention.
(The model first predicts changes in participants’ social networks that will
occur as a result of participation, based on who is assigned to their participation
group.) The intervention team conducts interviews with each participant in
order to acquire information about their past drug use, who is in their personal
(or “egocentric”) social network, how strongly they are related to those
individuals, and to what extent those individuals are known by the participant
to have used drugs, currently or in the past. The agent’s influence model is then
updated based on this information, allowing it to make predictions about how
various choices of intervention groups will affect participants’ future drug use
behavior. The agent then attempts to identify and recommend the “optimal” set
of intervention groups – the set of intervention groups predicted by its model
to result in the greatest aggregate (or population-level) reduction in future drug
abuse by all participants.
Researchers hope that this modification to TND Network, which we can
call TND Network-SA (“SA” for “Software Agent”), will render the program
even more effective at reducing drug abuse by minimizing potential deviancy
training effects. This is particularly important given that their target population,
homeless youth, has an unusually high proportion of individuals who currently
abuse or have abused drugs in the past, which increases the risk that the
intervention will lead to deviancy training by strengthening participants’ social
ties to other youth who view drug abuse favorably.
On early tests on sample data, however, researchers discovered a potential
beneficence problem. Researchers fed previously collected data about the
social networks of members of a particular population of homeless youth in
Los Angeles, California (where the center is based) into the agent’s influence
model, and then instructed the agent to recommend intervention groups for a
hypothetical set of participants in that population. The agent made a surprising

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 David Gray Grant 235

recommendation: it recommended that the intervention team put the youth

currently at highest risk of drug abuse into one group, and divide the youth
at lower risk among the remaining groups. The agent’s model, it turned out,
had predicted that grouping the hypothetical participants in this way would
result in a substantial reduction in aggregate future drug abuse for the lower-
risk youth – at the cost of a substantial predicted increase in future drug abuse
for the higher-risk youth. The expected decrease in risk for the lower-risk youth
was great enough that, despite the increased risk for the higher-risk youth, the
resulting grouping was optimal from the point of view of minimizing risk at
the population level.
CAIS researchers judged that the model’s predictions here were reliable
enough for the risk of significant harm here to be genuine if the hypothetical
intervention were actually conducted. As a result, the intervention poses a
beneficence problem, as defined earlier.

Beneficence Problems
In this section, I provide a more formal definition of beneficence problems
in general. In the field of artificial intelligence, a planning problem is the
problem of identifying the sequence of actions that is best suited for achieving
some goal. A particular planning problem can be defined by specifying the
following:
1. An agent of some kind (e.g., a human, a software agent, a robot, or a team
consisting of some combination thereof);
2. A goal to be achieved by that agent;
3. A set of possible states of the world – particular ways for the world
to be;
4. A set of possible actions that the agent could perform in order to achieve
that goal;
5. The agent’s beliefs about the starting state of the world, e.g., in the form
of a probability distribution over the states referenced above;
6. An optional set of constraints on how the goal can be achieved, used to
rule out as possible solutions to the problem actions or sequences of
actions known to be infeasible or otherwise undesirable for the agent to
perform.3
In a beneficence problem, the agent in question includes both (a) a human team
that is planning a public health social work intervention and (b) a software
agent that will assist that team in planning the intervention. The goal to be

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 236 Ethics and Artificial Intelligence in Public Health Social Work

achieved is the goal of maximizing benefits of a particular kind for some target
population. The states are various possible features of the intervention context
that are relevant to predicting the intervention’s effects. Both the intervention
team and the agent have relevant beliefs about the starting state of the world
prior to the intervention. The intervention team has information about the
starting state and supplies that information to the software agent in the form
of input. This input is used to generate a formal representation of the starting
state in terms that the agent is capable of understanding. There may or may
not be any preexisting constraints on how the team achieves the relevant goal.
The software agent and the intervention team will plan different aspects of
the intervention, depending on what planning tasks are being offloaded to the
software agent.
We need to specify two further features of the planning problem to
complete our definition of beneficence problems. The first is that it has an
explicitly moral constraint on how the goal of maximizing benefits for the
target population can be achieved: that it be reasonable to expect that the
intervention plan that the intervention team conducts based on the agent’s
recommendations will appropriately respect the duties of benefit maximization
and non-maleficence. Call this the beneficence constraint. The second is
that the agent is capable, or could be made capable, of providing useful
predictions about the expected benefits and harms of the intervention for
particular members of the target population.
This last feature is what distinguishes AI planning interventions that pose
beneficence problems (AI planning interventions*) from other public health
social work interventions that generate similar conflicts between the duties
of benefit maximization and non-maleficence. With the greater ability to
assess individual-level effects that AI interventions* offer comes a greater
responsibility to minimize potentially harmful effects. This leads to difficult
questions about how to resolve the associated dilemma between beneficence
and non-maleficence.

Moral Duties and Beneficence Problems

A moral duty, as I use the term here, is a relatively general moral obligation that
applies across a range of possible contexts. As I have mentioned, beneficence
problems pose a dilemma between two different and sometimes conflicting
moral duties that public health social work professional have: maximizing the
benefits of interventions they conduct and avoiding causing harm to particular
individuals. These two duties are combined in a foundational principle of

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 David Gray Grant 237

human subjects research ethics, the principle of beneficence. The Belmont

Report states the principle as follows:

Persons are treated in an ethical manner not only by respecting their decisions
and protecting them from harm, but also by making efforts to secure their
well-being. Such treatment falls under the principle of beneficence. The term
“beneficence” is often understood to cover acts of kindness or charity that go
beyond strict obligation. In this document, beneficence is understood in a
stronger sense, as an obligation. Two general rules have been formulated as
complementary expressions of beneficent actions in this sense: (1) do not harm
and (2) maximize possible benefits and minimize possible harms.4

As the last sentence brings out, there are really two principles here. The
first, “do not harm,” is ancient in its origins5 and is commonly referred to
as the principle of non-maleficence. The second, “maximize possible benefits
and minimize possible harms,” is a version of the utilitarian injunction to bring
about “the greatest good for the greatest number.” I refer to it as the principle of
benefit maximization. The duties of non-maleficence and benefit maximization
are widely accepted in public health and social work, whether or not human
subjects research is involved, though there is significant controversy over how
to address conflicts between the two duties.6
It is worth emphasizing that the benefits the duty of benefit maximization
refers to are expected benefits, net benefits, and aggregate benefits. They
are expected benefits because there is uncertainty involved. The intervention
could have various possible effects, and researchers are really only in a
position to estimate how beneficial or harmful those effects might be under
various possible conditions, and assess the expected value or disvalue of the
intervention across the full range of possible ways things could turn out.
They are net benefits because benefits are allowed to compensate for harms
in determining the total benefit provided, at least for some types of harms and
benefits. (The harms might be greater than the benefits, in which case the net
benefit would be negative, a net harm.) And they are aggregate benefits because
they are calculated by summing the expected benefits for each individual
to produce an aggregate score representing the total expected net aggregate
benefit for the population as a whole. We can also call them population-level
benefits for this reason, as opposed to individual-level benefits.
The duty of non-maleficence refers to expected harms in the same sense
that the duty of benefit maximization refers to expected benefits. Since an
intervention’s effects cannot be predicted with certainty, public health social
work professionals are enjoined to avoid risks that they are in a position to
foresee, and weigh those risks both by their severity and by how likely they are

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 238 Ethics and Artificial Intelligence in Public Health Social Work

to occur. The level of expected risk an intervention poses is therefore a function

of both the likelihood and severity of the relevant risks. However, the relevant
harms are neither net nor aggregate: the duty directs public health social work
professionals to avoid expected harms to particular individuals, regardless of
whether they are compensated for by benefits to those same individuals or to
others.
It will simplify the discussion in what follows to confine our attention to
cases where the expected harms and benefits of an intervention are differ-
ent sides of the same coin: the benefits alleviate some potentially harmful
condition a person can be in (e.g., being disposed to abuse drugs), whereas
the harms make that condition worse. This is true in TND Network-SA:
the relevant type of harm is being influenced to increase future drug abuse
behaviors; the relevant type of benefit is to decrease future drug abuse
behaviors. Restricting our attention in this way will let us set aside issues of
intrapersonal compensation, such as determining whether a person is harmed
on balance when they are made more likely to abuse drugs but less likely to
contract HIV.
Some further distinctions between the two duties will also be helpful for our
purposes. One such distinction that is worth making explicit (even though it is
obvious) is that the duty of non-maleficence requires that specific individuals
be treated in certain ways, whereas the duty of benefit maximization requires
that larger populations be treated in certain ways. Call moral duties of
the former kind individual-oriented duties and moral duties of the second
kind population-oriented duties. One thing to notice right away is that the
agent used in an AI intervention* will need the capability to predict both
individual- and population-level effects in order to help the intervention team
make determinations about whether the duties of non-maleficence and benefit
maximization are satisfied.
Another crucial distinction between the two duties for our purposes is that
the duty of non-maleficence is what is sometimes known in moral theory
as a perfect duty, whereas the duty of aggregate benefit maximization is an
imperfect duty. At least one way of understanding the distinction is that perfect
duties specify types of actions that are morally required or morally prohibited.
The duty of non-maleficence, then, is a perfect duty because it prohibits us
from performing any action that would harm another person. Imperfect duties,
by contrast, require us to pursue certain goals. The duty to maximize benefits is
most naturally understood as an imperfect duty, as it requires those designing
and conducting public health social work interventions to make reasonable
efforts to maximize the expected aggregate benefits of those interventions.
However, the duty does not specify what kinds of means should be used

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 David Gray Grant 239

to achieve this goal: it does not require or prohibit actions of any specific
kind, except to say that the interventions in question should be well-suited
to promoting the goal of maximizing population-level benefits.
There are also imperfect/individual-oriented duties and perfect/population-
oriented duties, some of which are highly relevant for thinking about
researchers’ obligations with respect to beneficence problems. Space
constraints prohibit more than a cursory discussion of these and other duties,
but it is worth mentioning a few in particular. First, social workers are generally
understood, as acknowledged by the National Association of Social Work’s
code of ethics, to have a general duty to promote the interests of their clients,
at least in certain respects.7 This includes ensuring that services provided to
clients – such as enrolling them in a public health intervention – are reasonably
expected to be of benefit to them, except under special circumstances. This
duty is both imperfect and individual-oriented.
Second, public health social work professionals are generally understood to
have a moral duty to ensure that the burdens and benefits of the interventions
they conduct are fairly distributed among various subpopulations.8 This is a
perfect and population-oriented duty. Researchers should take special care,
then, when designing AI interventions* that may impose unfair burdens on
some subpopulation. In AI interventions* as I have described them, the
information about social status necessary to determine whether this duty is
satisfied by particular intervention plans may well not be represented in the
formal model the agent uses to plan the intervention. Building this information
into the model can help the researchers developing the intervention determine
whether the agent is like to recommend plans that may distribute benefits
unfairly, and could potentially be used by the agent itself as it assists attempts
to identify and recommend potentially desirable plans.
I will set aside these further duties in what follows, and confine my attention
to the duties of benefit maximization and non-maleficence. However, these
other moral obligations are worth keeping in mind, and it is important to note
that public health social workers have other moral duties that will eventually
need to be factored in as well when beneficence problems are addressed.
One reason the distinction between imperfect and perfect duties is useful
for our purposes is that it maps roughly onto the distinction between goals
and constraints in the literature on AI planning. Imperfect duties specify moral
goals that human agents should pursue; perfect duties specify moral constraints
that human agents need to observe as they pursue these goals and others.
The similarity here is not just superficial: it mirrors the way that perfect
and imperfect duties typically interact with one another. Insofar as the two
conflict, perfect duties are generally understood to have priority over imperfect

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 240 Ethics and Artificial Intelligence in Public Health Social Work

duties. The constraints on action imposed by perfect duties, that is, restrict the
range of actions an agent is morally permitted to perform as she pursues the
goals imposed on her by imperfect duties. Ordinarily at least, an agent is not
permitted to violate a perfect duty for the sake of better promoting the goal
an imperfect duty obligates her to pursue. We will consider some potential
exceptions to this general rule further on.
This suggests a natural strategy for accommodating perfect and imperfect
duties in the design of AI interventions*: accommodate imperfect duties
through planning goals; accommodate perfect duties through planning con-
straints. In an AI intervention*, the software agent assists the intervention team
by planning certain aspects of the intervention on their behalf. In order to help
the intervention team avoid violating imperfect duties, appropriate changes can
be made to the goals of the planning problem the agent seeks to solve on their
behalf. In order to help them avoid violating perfect duties, changes can be
made to the planning problem’s constraints.
I’ll talk more about how this might work in the “Operationalizing the
Framework” section, but first we should consider what obligations public
health social work professionals have when the duties of non-maleficence and
benefit maximization conflict.

A Framework for Resolving Conflicts

Ethicists working in a variety of fields have long recognized that the duties of
beneficence and non-maleficence can come into conflict in ways that generate
difficult ethical dilemmas. The Belmont Report acknowledges this possibility
early on:

[The] role of the principle of beneficence is not always so unambiguous. A

difficult ethical problem remains, for example, about research that presents
more than minimal risk without immediate prospect of direct benefit to the
children involved. Some have argued that such research is inadmissible, while
others have pointed out that this limit would rule out much research promising
great benefit to children in the future. Here again, as with all hard cases, the
different claims covered by the principle of beneficence may come into conflict
and force difficult choices.9

A beneficence problem presents researchers with a dilemma with exactly this

structure: in some foreseeable intervention contexts, the software agent may
recommend an intervention plan that maximizes expected aggregate benefits
for the target population by generating expected benefits for some of its
members and expected harms for others. The duties of non-maleficence and

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 David Gray Grant 241

benefit maximization appear to conflict in these cases, posing the question of

how to resolve the conflict.
I mentioned earlier in the chapter that perfect duties typically place stringent
constraints on how imperfect duties may be pursued. Given this, a natural
first thought is that these conflicts are easy to resolve: researchers are simply
prohibited from imposing expected harms on particular individuals for the sake
of producing greater aggregate benefits for the larger population. However,
there are good reasons to think that there will be many exceptions to this
general rule in the context of public health in general and social work in
particular.
There are various ways to defend the claim that public health social work
interventions may sometimes violate the duty of non-maleficence, consistent
with being morally justifiable on balance, but one standard (if not uncontro-
versial) argument is as follows.10 The moral justification for particular public
health interventions is often understood to derive from the justification of the
public health system as a whole. What justifies the sum total of interventions
conducted by a public health system, in spite of the fact that particular
interventions impose net costs (including harms and other kinds of costs) on
some individuals, is that each individual has a strong reason to prefer living
under a public health regime that allows individual interventions to trade off
costs for some individuals to achieve comparatively greater benefits for others.
That reason is that a system prohibited from intervening to improve the health
of a population at a cost to some of its members would be far less effective
overall. Provided that the system is designed to ensure that the burdens an
individual is subjected to from any given intervention are compensated for
by benefits from other interventions, the net result should be that everyone
experiences greater benefits on balance from systems that allow such trade-
offs to be made than those that do not.
Excise taxes on cigarettes, for example, improve aggregate well-being in a
population by reducing the incidence of various diseases, but accomplish this at
a cost to those smokers who are not persuaded to reduce their intake. Similarly,
access restrictions on performance-enhancing drugs such as Ritalin benefit
populations by reducing their abuse, but leave some individuals who would
benefit from them worse off. What makes these trade-offs morally acceptable?
According to the foregoing argument, these trade-offs are justifiable because
those individuals who are negatively affected nonetheless have good reason
to prefer living under a public health regime that allows them to be made –
provided that there are restrictions on how significant the expected burdens
can be in particular cases. Call this argument the higher-order justification
argument.

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 242 Ethics and Artificial Intelligence in Public Health Social Work

This argument is at least initially plausible as it applies to public health

social work interventions in particular. The individual members of the popu-
lations served by public health social workers have a reason to prefer being
served a system that allows public health social work interventions to make
the kinds of trade-offs just described, provided that safeguards are in place to
ensure that the end result is that each individual expects to benefit more if those
trade-offs are allowed in some cases than if they are not. This justifies relaxing
the requirements of the duty of non-maleficence as it applies to individual
interventions to at least some degree, allowing for net expected harms to
individuals in some cases.
Suppose, then, that this argument, or another one with the same conclusion,
is correct. If so, then some exceptions to the principle of non-maleficence
as it applies to public health social work interventions can be justified. It
remains to be seen how public health social work professionals should go
about determining whether the expected harms imposed by a specific public
health intervention are morally justifiable. This is of course a complex question
requiring extensive treatment, but a few preliminary points can be made. I will
briefly sketch a framework, based on existing work in public health ethics,11
that offers a useful and at least initially plausible proposal about how public
health social work professionals should go about negotiating conflicts between
the duties of non-maleficence and benefit maximization.
The framework posits three moral duties that it takes to apply to public
health social work interventions where the duties of non-maleficence and
benefit maximization conflict. The first two duties apply to interventions
considered in isolation; the third applies to interventions considered in light
of available alternative interventions.
First, the duty of necessity. Suppose an intervention is believed to offer
certain magnitude of expected benefit to the target population and a certain
magnitude of expected harm to some of its individual members. The duty of
necessity requires that the expected harms of the intervention be necessary to
achieve its expected benefits. If the intervention can be modified in ways that
reduce those expected harms while maintaining a comparable level of expected
benefit, then the intervention must be modified accordingly. Doing so reduces
the severity of the intervention’s violation of the duty of non-maleficence.
It also increases the degree to which the duty of benefit maximization is
satisfied by the intervention, since expected harms for individual members
of the target population entail reduced aggregate expected benefits for the
population considered as a whole.
Second, the duty of proportionality requires that the expected benefits of an
intervention be great enough to justify the expected harms involved—that the
two be suitably “proportional” to one another. The idea here is that sufficiently

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 David Gray Grant 243

great benefits can outweigh comparatively minor harms. We can see this logic
at work in an argument mentioned in the second quote from the Belmont report
cited earlier (my emphasis added):

A difficult ethical problem remains, for example, about research that presents
more than minimal risk without immediate prospect of direct benefit to the
children involved. Some have argued that such research is inadmissible, while
others have pointed out that this limit would rule out much research promising
great benefit to children in the future.

The implicit argument here is that the expected benefits of research of this
kind for future children is so great that it justifies the uncompensated risks
posed to the children that will be enrolled as research subjects, because the
relevant risks and benefits are “proportional” given the relative importance of
those two moral considerations.
Why should we accept the duty of proportionality as a necessary condition
on permissible interventions? Here is one argument. That an intervention is
expected to generate substantial benefits for the target population counts in its
favor, from a moral point of view, relative to alternative options. Other things
being equal, it is better to perform interventions that produce greater aggregate
benefits, as recognized by the duty of benefit maximization. Further, that an
intervention poses a significant expected risk of harm to some of the target
population’s members counts against it, again from a moral perspective. But
the two kinds of moral considerations just mentioned are not equally important.
Expected risks count more against an intervention than expected benefits of
a comparable magnitude. This way of thinking about the comparative moral
importance of expected risks and expected benefits recognizes the fact that the
duty of non-maleficence imposes a meaningful constraint on the way expected
benefits can be pursued by public health professionals. If the two were not
weighted differently, in this way, then the duty of non-maleficence would not
impose a meaningful constraint.
Third, the duty of trade-off optimization. Unlike the first two duties, this
duty requires that the expected benefits and harms of an intervention be
compared to those of viable alternative interventions. Suppose an intervention
team is faced with a choice between possible interventions that all satisfy
the duties of necessity and proportionality – various versions of their own
intervention, perhaps – and other alternative interventions they could perform.
How should they choose among these possibilities? According to the duty
of tradeoff optimization, researchers are obligated to choose the intervention
from the set that strikes the best balance between expected benefits and
expected harms – the expected benefit/expected harm profile that best satisfies
the duty of proportionality.12 The thought here is that the elements of a set

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 244 Ethics and Artificial Intelligence in Public Health Social Work

of interventions can each satisfy the duty of proportionality, consistent with

it being true that the expected benefits and harms are more “proportional”
for some of the interventions in the set than for others. More proportional
interventions strike a more morally desirable balance between the conflicting
demands of the duty of non-maleficence and benefit maximization, and so
should be preferred.
The tentative proposal I am offering here is that AI interventions* that
satisfy all three of these duties – of necessity, of proportionality, and of
trade-off optimization – are morally justifiable insofar as the duties of non-
maleficence and benefit maximization are concerned. This is just a first
step and requires much further refinement and assessment in light of other
ways of approaching conflicts between the two duties. Moreover, other moral
duties, such as those mentioned earlier, are not yet taken into account by the
framework. However, I think the framework is at least initially plausible as
far as it goes, and – despite its partial nature and the complexity of the issues
involved – we can use it to draw a number of tentative conclusions about how
AI interventions* should be designed.
We need to introduce one final element into the mix before we get down to
brass tacks and consider how to operationalize the framework just described.
As I mention earlier, autonomous agents such as those considered here
make it possible to make more accurate individual-level predictions about an
intervention’s benefits and risks than was previously possible. One question
researchers need to consider as they attempt to satisfy the foregoing three moral
duties is to what extent they are obligated to use these tools to improve their
estimates of those individual-level risks. This question is important because the
goal of achieving a more accurate and comprehensive estimate of the expected
individual-level harms of an AI intervention* is to some extent in tension with
the goal of maximizing expected benefits for the population as a whole. One
reason for this is that there is only so much research time that can be spent
improving an intervention in the variegated respects that are morally desirable.
Another reason is that many ways of improving a software agent’s estimates
of expected individual-level effects for particular intervention plans introduce
additional computational complexity, which can in itself reduce the agent’s
ability to optimize population-level benefits.

Operationalizing the Framework

We can now return to the question of how researchers should proceed when
they believe an AI planning intervention they are developing has a beneficence

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 David Gray Grant 245

problem. Assume researchers are designing the software agent associated with
such an intervention.
Let’s begin by revising our original definition of beneficence in light of
the framework I have proposed for resolving conflicts between the duties
of benefit maximization and non-maleficence. Specifically, we can remove
the beneficence constraint and replace it with a new constraint: that it be
reasonable to expect that the intervention plan the intervention team actually
conducts, based on recommendations from the agent, will satisfy all three
of the moral duties just described: the duty of necessity, the duty of pro-
portionality, and the duty of optimal trade-offs. This in itself is progress, as
the redefined problem gives us a better idea of what is required to respect
the duties of benefit maximization and non-maleficence when they come into
conflict.
As is probably clear by now, it is extremely unlikely that this constraint
can be fully operationalized in the design of the agent so that it never recom-
mends plans inconsistent with it. Determining whether the three moral duties
involved are satisfied by a particular intervention requires time, experience,
consideration of the merits of possible alternative interventions, and open
discussion with other professionals and stakeholders (including members of
the target population) with a diverse enough set of perspectives to minimize the
chances that something important will be overlooked.13 Formulating general
and operationalizable rules that will provide reasonable assurances that they
are never violated by the intervention plans that the software agent in an AI
intervention* recommends is almost certainly impossible.
What this means is that human judgment will always be required – the
intervention team will need to assess the intervention plans recommended by
the agent to determine whether it satisfies the duties of necessity, proportional-
ity, and optimal trade-offs. However, the software agent might nonetheless be
designed in ways that make it less likely than it would otherwise be that the
duties will be violated in the field. I will sketch out three possible strategies.
First, the agent could be designed so that when it recommends a particular
intervention plan, it provides the intervention team with supplemental infor-
mation relevant to determining whether the recommended plan satisfies the
constraints of necessity, proportionality, and optimal trade-offs. Call this the
interpretability approach.
How might the interpretability approach look in practice? Researchers at
CAIS are currently redesigning the agent to be used in TND Network-SA
to supply the intervention team with more information about the expected
individual- and population-level harms and benefits associated with the plans
it recommends. With this information in hand, researchers will be in a

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 246 Ethics and Artificial Intelligence in Public Health Social Work

much better position to assess whether the moral duties associated with
beneficence problems are satisfied. An obvious first step here is for the agent
to be programmed to calculate the change in expected value predicted to be
achieved by conducting a particular intervention plan, relative to conducting
no intervention, for each individual member of the target population, as well
as for the target population considered as a whole. This would require, inter
alia, that the agent’s model include what we can call an individual-level value
function. Let an individual-level value function be a function from pairs of
states and individual members of a target population to real numbers on the
interval between 0 and 1. The value assigned to a pair is the value for the
individual of being in that state.
If the change in expected value for an individual (or population) associated
with a particular intervention plan is positive, then the intervention is expected
to benefit the individual (population) relative to doing nothing; if it is negative,
then it is expected to harm the individual (population). Providing this infor-
mation to the researchers lets them know how conducting the recommended
intervention plan is expected to positively and negatively affect both particular
individuals and the population considered as a whole. This would at least put
the researchers in a better position to assess whether the proportionality and
necessity constraints are satisfied, in conjunction with the information already
available to them.
Second, the agent could be redesigned so that it automatically rules out
plans with features that can be determined in advance to violate one of the
three duties in the framework. This could be accomplished using what AI
researchers call safety constraints. Call this the safety constraint approach.
How might this work? One possibility is to model the planning problem
solved by the agent as a constrained partially observable Markov decision
process (constrained POMDP). A partially observable Markov decision pro-
cess (POMDP) is a mathematical tool for modeling planning problems where
there is uncertainty about the initial state and how it will evolve over time
based on actions performed by the agent.14 This makes POMDPs useful for
modeling beneficence problems, where both kinds of uncertainty are present.
In a constrained POMDP, the agent’s goal is to maximize the value of one
function while simultaneously bounding the values of other functions within
a specified range.15 Researchers could determine a level of expected harm
that they judge would be impermissible to impose on an individual in the
preponderance of foreseeable starting intervention states, and direct the agent
to bound the expected change in individual-level value functions associated
with each member of the target population accordingly. An advantage of

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 David Gray Grant 247

this strategy is that the intervention team would have to reject recommended
strategies less often, which could dramatically reduce the amount of time it
takes to identify a morally acceptable intervention plan.
Another variant on the safety constraint approach would be to generate
recommended intervention plans using a two-step procedure. In the first step,
the agent would eliminate all possible end states that exceed a level of expected
harm for individuals that they have decided is unlikely to be acceptable.
It would then work backward by considering the states that are possible
in the preceding step and eliminating any that have a high probability of
resulting in one of those end states. This could be repeated until the step
immediately following the initial state is reached. Pynadath and Tambe (2001)
show how this can be done in planning problems that can be modeled as
vanilla Markov decision processes. Whether the strategy could be extended in
a computationally tractable way to beneficence POMDPs is beyond the scope
of this paper, but is worth exploring.
Third, researchers could implement what Pynadath and Tambe (2001)
call safety conditions. Safety conditions specify conditions under which the
software agent assisting a human team should transfer control to its human
operators in order to request more information or a decision of some kind. The
general idea behind safety conditions is that the human operators of a software
agent designed to solve a planning problem often have important information
about the problem that the agent does not. In many cases, this information
cannot be reliably captured or processed by the agent as it decides how to act,
regardless of how well it is designed.16 In principle, such conditions could be
used to direct the agent in an AI intervention* to present the intervention team
with a selection of possible intervention plans, each with a different risk/benefit
profile. If feasible, this would put the intervention team in an even better
position to determine how to satisfy our three moral constraints, since they
will have more information about the effects of different possible intervention
plans. Note that, for this to work, the agent will need to choose alternative plans
in an intelligence way, by recommending plans that make the trade-off between
expected individual-level harms and population-level benefits in meaningfully
different ways. Call this the safety condition approach.
These three approaches could of course be combined, if computational
limitations allow. Various other strategies are also of course possible. In any
case, strategies like these have considerable potential to help the intervention
teams in AI interventions* select an intervention plan that promises morally
desirable expected population-level benefits without imposing morally unjus-
tifiable expected individual-level harms.17

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 248 Ethics and Artificial Intelligence in Public Health Social Work

Notes
1 Dishion et al., 1999.
2 See Center for Artificial Intelligence in Society (2017) for more information.
3 Russell and Norvig, 2010.
4 National Institutes of Health, 1979.
5 The Hippocratic Oath, for instance, was written as early as the fifth century BC,
and requires physicians to swear as follows: “I will utterly reject harm and
mischief” (Wikipedia, 2017). The more familiar phrasing, “first, do no harm,”
appears to have been coined by the English surgeon Thomas Inman in the 19th
century (Sokol, 2013).
6 On the public health side, see, for instance, Mann (1997) and Childress
et al. (2002). On the social work side, the NASW’s Code of Ethics explicitly
endorses a version of benefit maximization (“The primary mission of the social
work profession is to enhance human well-being and help meet the basic human
needs of all people, with particular attention to the needs and empowerment of
people who are vulnerable, oppressed, and living in poverty”) and repeatedly
enjoins social workers to avoid activities that would harm clients. See National
Association of Social Workers (2017).
7 National Association of Social Workers, 2017.
8 See Faden and Shebaya (2016) regarding the field of public health. The NASW
code of ethics makes clear that social workers are obligated to promote social
justice, including the fair distribution of resources of various kinds, in their work
(National Association of Social Workers, 2017).
9 National Institutes of Health, 1979.
10 My presentation of the argument presented later in the chapter owes much to
Faden and Shebaya (2016), section 2.1.
11 The framework I develop is based in large part on the framework proposed in
Childress et al. (2002), though it differs in important respects.
12 How to understand “best” in the preceding sentence is a further question that I
will set aside here, but it will almost certainly depend on the nature of the
argument that is given for allowing exceptions to the duty of non-maleficence.
13 There are also independent reasons to consult stakeholders in particular. See
Childress et al.’s (2002) discussion of a principle they call “public justification.”
14 See Russell and Norvig (2010), chapter 17, section 4.
15 See Isom et al. (2008).
16 Pynadath and Tambe, 2001.
17 The research for this paper was supported by the Center for Artificial Intelligence
in Society (CAIS) at the University of Southern California (USC) over the course
of a summer research fellowship in 2017. Thanks to Eric Rice, Milind Tambe, and
Hailey Winetrobe at CAIS, Robin Petering and Mee-Young Um at the USC
Suzanne Dworak-Peck School of Social Work, Aida Rahmattalabi and Bryan
Wilder at the USC Viterbi School of Engineering, and Hau Chan at Harvard
University for their insight and helpful discussion. Thanks also to the other 2017
summer fellows at CAIS and the members of the Teamcore research group at
USC for many useful conversations that contributed to my understanding of the
issues discussed in this essay. Special thanks to Lyndal Grant at the Massachusetts
Institute of Technology for detailed comments on previous drafts.

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 David Gray Grant 249

References
Center for Artificial Intelligence in Society. 2017. Social Network-Based Substance
Abuse Prevention for Homeless Youth. Available at: https://www.cais.usc.edu/
projects/social-network-based-substance-abuse-prevention-for-homeless-youth/.
Accessed August 15, 2017.
Childress, James F., Faden, Ruth R., Gaare, Ruth D., Gostin, Lawrence O., Kahn,
Jeffrey, Bonnie, Richard J., Kass, Nancy E., Mastroianni, Anna C., Moreno,
Jonathan D., and Nieburg, Phillip. 2002. Public health ethics: Mapping the terrain.
The Journal of Law, Medicine & Ethics, 30(2), 170–178.
Dishion, Thomas J., McCord, Joan, and Poulin, Francois. 1999. When interventions
harm: Peer groups and problem behavior. American Psychologist, 54(9), 755–764.
Faden, Ruth, and Shebaya, Sirine. 2016. Public health ethics. In: Zalta, Edward N. (ed.),
The Stanford Encyclopedia of Philosophy, winter edn. Stanford, CA: Metaphysics
Research Lab, Stanford University.
Isom, Joshua D., Meyn, Sean P., and Braatz, Richard D. 2008. Piecewise linear dynamic
programming for constrained POMDPs. AAAI-2008, 1, 291–296.
Kempe, David, Kleinberg, Jon, and Tardos, Éva. 2003. Maximizing the spread of
influence through a social network. KDD, 137.
Mann, Jonathan M. 1997. Medicine and public health, ethics and human rights.
Hastings Center Report, 27(3), 6–13.
National Association of Social Workers. 2017. NASW Code of Ethics (Guide to the
Everyday Professional Conduct of Social Workers). Available at: https://www
.socialworkers.org/LinkClick.aspx?fileticket=ms_ArtLqzeI%3d&portalid=0
National Institutes of Health. 1979. The Belmont Report: Ethical Principles and
Guidelines for the Protection of Human Subjects of Research. Available
at: https://www.hhs.gov/ohrp/regulations-and-policy/belmont-report/read-the-
belmont-report/index.html
Pynadath, David V., and Tambe, Milind. 2001. Revisiting Asimov’s First Law: A
response to the call to arms. 8th International Workshop on Intelligent Agents VIII,
307–320.
Russell, Stuart, and Norvig, Peter. 2010. Artificial Intelligence: A Modern Approach.
Upper Saddle River, NJ: Pearson Education.
Sokol, Daniel K. 2013. “First Do No Harm” Revisited. BMJ, 347(October), f6426.
Valente, Thomas W., Ritt-Olson, Anamara, Stacy, Alan, Unger, Jennifer B., Okamoto,
Janet, and Sussman, Steve. 2007. Peer acceleration: Effects of a social network
tailored substance abuse prevention program among high-risk adolescents. Addic-
tion, 102(11), 1804–1815.
Wikipedia. 2017. Hippocratic Oath. Available at: https://en.wikipedia.org/wiki/
Hippocratic_Oath. Accessed: August 15, 2017.

https://doi.org/10.1017/9781108669016.015 Published online by Cambridge University Press

 Glossary
Elizabeth Bondi and Mee-Young Um

Alters: An individual described by the ego as a person with whom the ego has
some type of relationship; the alters surround the ego in an egocentric analysis
and network map.
Binary: In the context of modeling (as opposed to binary numbers), a variable
can take one of two possible values. For example, information could be
propagated along an edge or not.
Categorical data: Also known as nominal data, can be divided into groups
based on a particular qualitative property. These data cannot be ordered. Some
examples of categorical data include race, gender, and education level.
Continuous data: Data that have an infinite number of possible values. Some
examples of continuous data include height and weight.
Dichotomize: To make a variable into a binary variable.
Dyads: Pairs of egos and their alters.
Edges: The connections between vertices in graph theory, which can be
directed or undirected. In sociometric network analysis, equivalent to ties.
Egos: The individual under study in an egocentric analysis (occupies the core
position in an egocentric network map); data are collected from the ego.
False negatives: In binary classification, instances that are incorrectly classi-
fied as the negative class (i.e., they should have been classified as positive). In
the case of multiple classes, this would be measured for each class (e.g., dog
or not, cat or not).
False positives: In binary classification, instances that are incorrectly classified
as the positive class (i.e., they should have been classified as negative). In the

250

https://doi.org/10.1017/9781108669016.016 Published online by Cambridge University Press

 Glossary 251

case of multiple classes, this would be measured for each class (e.g., dog or
not, cat or not).
Feature: Any characteristic that can vary in values (quantity or quality) among
individuals in a population or sample. (Same as Variable)
Graph: A mathematical structure used to model relations between vertices.
Relationships can be directed or undirected, meaning the order of the edges
may or may not matter. For example, friendship may be directed, since one
person (vertex) may nominate another person as his/her friend, but the other
person may not. (Same as Network)
Health Insurance Portability and Accountability Act (HIPAA): A federal
law passed by US Congress in 1996. HIPAA requires medical records and
other health information of patients to be kept confidential, protects continuity
of health insurance coverage for individuals who change or lose their jobs,
reduces fraud and abuse of health care, and standardizes electronic transactions
that contain health care information.
Homophily: Tendency for people to bond and associate with others who are
like themselves.
Institutional Review Board (IRB): A committee or administrative body who
reviews, approves, and monitors research activities involving human subjects
to ensure that they are conducted in accordance with all federal, institutional,
and ethical guidelines.
Longitudinal data: Data resulting from a study design involving data collec-
tion at different points in time. Often used interchangeably with panel data,
but generally has a larger number of participants.
Machine learning: An area of artificial intelligence research in which algo-
rithms are designed to learn from data (e.g., analyze patterns in the data
automatically) and use the learned models to make predictions given new data.
Model: A mathematical description of a phenomenon, such as independent
cascade to describe how information propagates through a social network, or
logistic regression to represent the relationship between two variables.
Network: A mathematical structure used to model relations between vertices.
Relationships can be directed or undirected, meaning the order of the edges
may or may not matter. For example, friendship may be directed, since one
person (vertex) may nominate another person as his/her friend, but the other
person may not. (Same as Graph)

https://doi.org/10.1017/9781108669016.016 Published online by Cambridge University Press

 252 Glossary

Observations: Experiences in the real world that help build a new social theory
or test an existing theory.
Ordinal data: Data with ordered categories that do not have a numeric
distance between each category. The Likert scale is a typical example of ordinal
data, where ordered categories can be strongly agree, agree, neutral, disagree,
and strongly disagree.
Panel of data: Data can be collected at different points in time (wave, panel,
or snapshot) to repeatedly examine the same variables and their changes over
time for the same participants. Variables can be added or removed from one
wave/panel to another. (Same as Waves of data and Snapshot)
Planning: An area of artificial intelligence research that involves reasoning
about a sequence of actions to achieve a goal, often involving uncertainty.
Policy: A (mathematical) description of what should be done in certain
situations, often associated with planning.
Precision: A number between 0 and 1 that can be calculated as the number of
true positives divided by the total number of instances classified as positive,
or the sum of true positives and false positives. In other words, it measures
whether those instances that were classified as positive are correctly classified
or not, or said another way, if there were mistakes while classifying instances
as positive. The closer to 1, the closer to correctly classifying all instances
predicted to be positive, the fewer mistakes made while classifying instances as
positive. Note that some false negatives would not affect this measure because
it does not take into account how many positive instances were missed.
Qualitative data: Information that cannot be expressed in numerical forms,
such as anecdotes or interviews.
Quantitative data: Information that can be expressed in numerical forms or
quantified, which is generally collected through surveys.
Recall: A number between 0 and 1 that can be calculated as the number of
true positives divided by all actual positives, or the sum of true positives and
false negatives. In other words, it measures the degree to which the classifier
is correctly classifying all positive instances, or said another way, whether
any positive instances are missed. The closer to 1, the closer to classifying
all positive instances as positive, the fewer positive instances missed. Note
that some false positives would not affect this measure because all positive
instances could still be correctly classified even if there were extra positive
instances. (Same as Sensitivity)

https://doi.org/10.1017/9781108669016.016 Published online by Cambridge University Press

 Glossary 253

ROC curve: A plot that can be used to evaluate the performance of a classifier.
The recall is plotted along the y-axis, and the false positive rate (the number
of false positives divided by the sum of false positives and true negatives, or
1 − specificity) is plotted along the x-axis. To generate a plot, and not simply
two numbers, the confidence threshold of the classifier is varied. For example,
the classifier may classify an instance as positive with 80% confidence. This
instance would be counted as correct when using a confidence threshold of
40%, but not for a confidence threshold of 90%, which would change the
measures used to generate the curve. A well-performing classifier would have
a curve close to lines at y = 1 for all x 0 to 1, and x = 1 for all y 0 to 1, meaning
there is perfect recall for all false positive rates and vice versa. The ROC curve
can also be used to select the optimal confidence threshold for the classifier.
ROC curve Stands for Receiver Operating Characteristic curve.
Sample: A subset of a population used to represent the entire population of
interest.
Sensitivity: A number between 0 and 1 that can be calculated as the number of
true positives divided by all actual positives, or the sum of true positives and
false negatives. In other words, it measures the degree to which the classifier
is correctly classifying all positive instances, or said another way, whether
any positive instances are missed. The closer to 1, the closer to classifying
all positive instances as positive, the fewer positive instances missed. Note
that some false positives would not affect this measure because all positive
instances could still be correctly classified even if there were extra positive
instances. (Same as Recall)
Snapshot: Data can be collected at different points in time (wave, panel, or
snapshot) to repeatedly examine the same variables and their changes over
time for the same participants. Variables can be added or removed from one
wave/panel to another. (Same as Panel of data and Waves of data)
Specificity: A number between 0 and 1 that can be calculated as the number
of true negatives divided by all actual negatives, or the sum of true negatives
and false positives. This can be thought of as the “opposite” of recall, in that it
measures the degree to which the classifier is correctly classifying all negative
instances, or said another way, whether any negative instances are missed. The
closer to 1, the closer to classifying all negative instances as negative, the fewer
negative instances missed. Note that some false negatives would not affect this
measure, because all negative instances could still be correctly classified even
if there were extra negative instances.

https://doi.org/10.1017/9781108669016.016 Published online by Cambridge University Press

 254 Glossary

Ties: The relationship between nodes in a sociometric network (also known

as an edge); ties can be derived from a variety of social relationships, such as
friendships, sexual contacts, persons with whom an individual uses drugs, or
collaborative networks among social service agencies.
True negatives: In binary classification, instances that are correctly classified
as the negative class. In the case of multiple classes, this would be measured
for each class (e.g., dog or not, cat or not).
True positives: In binary classification, instances that are correctly classified
as the positive class. In the case of multiple classes, this would be measured
for each class (e.g., dog or not, cat or not).
Unstructured data: Textual data, such as anecdotes or interviews.
Variable: Any characteristic that can vary in values (quantity or quality)
among individuals in a population or sample. (Same as Feature)
Vertices: Also known as nodes in graph theory, a set of fundamental points
connected by edges. These points are called egos in sociometric network
analysis.
Waves of data: Data can be collected at different points in time (wave, panel,
or snapshot) to repeatedly examine the same variables and their changes over
time for the same participants. Variables can be added or removed from one
wave/panel to another. (Same as Panel of data and Snapshot)

https://doi.org/10.1017/9781108669016.016 Published online by Cambridge University Press

 Index

AI: Artificial Intelligence DIME: Dynamic Influence Maximization

Definition, 8 Under Uncertainty
Ethics and Artificial Intelligence in Public Defined, 39
Health Social Work, 231 Problem statement, 43
History, 8 Solver, 42
Improving access to sexual health DOSIM
necessities for youth homelessness, 136 Algorithm for DIME problem, 50
Predicting intimate partner violence (IPV) Have You Heard? algorithm, 16, 16
perpetration, 195
Why we need AI, 28 Ethics and Artificial Intelligence in Public
ARISEN: Approximating with Random Health Social Work, 231
walks to Influence a Socially Explored Case Study: TND Network for Homeless
Network Youth, 233
Algorithm, 80 Framework for resolving conflicts, 240
Introduction, 231
CAIS: University of Southern California Moral duties and beneficence program, 236
(USC) Center for Artificial Intelligence Operationalizing the framework, 244
in Society TND: Towards No Drug Abuse, 233
Motivations, 4
Foster youth
Diabetes Patients Predictive Modeling Multidisciplinary study on the relationship
of Stress between foster care attributes and
Aims of the project, 157 Posttraumatic Stress Disorder Symptoms
Conclusion, 166 (PTSD) on foster youth, 169
Data description, 154 PTSD symptoms among youth with
Experimental results and discussion, 162 homelessness and foster care histories,
Experimental setup, 158 170
Introduction and statistics, 153
Neural networks, 159 GOFAI: Good Old-Fashioned AI
Predictive modeling of stress, 153 Defined, 9
Type 2 diabetes (T2D), 153 Grand Challenges of Social Work, 5

255

https://doi.org/10.1017/9781108669016.017 Published online by Cambridge University Press

 256 Index

Have You Heard? Local information, 78

Behavioral intervention, 16 SBM: Stochastic Block Model, 79
HEALER: Hierarchical Ensembling Based Intimate partner violence (IPV) perpetration
Agent predictions using artificial intelligence,
Construction, 43 195
Design, 41 Conclusion, 206
Have You Heard? algorithm, 16 Conflict Tactics Scale (CTS2), 197
Influence model, 45 Data set description, 197
Solution to DIME problem, 39 Introduction and statistics to IPV, 195
Uncertain network, 43 Methodology, 201
HIV: Human Immunodeficiency Virus Problem definition, 196
Prevention for homeless youth, 26 Results, 203
Prevention via social networks, 27 Support Vector Machines (SVM), 200
Raise awareness via social networks, 38
Risk in homeless youth, 23 LGBT: Lesbian, Gay, Bisexual, or
Young men who have sex with men Transgender
(YMSM), 24 Homeless youth, 20
Homelessness LGBTQ: Lesbian, Gay, Bisexual, Transgender,
Causes of youth homelessness, 19 Questioning and Queer
Couch surfing, 17 HIV, 212
Extent of youth homelessness, 17
LGBT youth, 20 Mental health
Travelers, 20 Homeless youth, 21
Violence, 21 Minimizing violence in homeless youth, 119
Data collection, 121
Improving access to sexual health necessities Difficulty in Emotion Regulation (DERS),
for youth experiencing homelessness 131
using AI Discussion, 129
Discussion and conclusions, 149 Experiments, 126
Greedy approach approximation guarantees, Greedy minimization, 124
144 Implications from social work perspective,
K-Center problem, 137 131
Population estimation, 145 Introduction, 119
Proposed approach, 142 Multiple waves, 130
Results, 148 Uncertain Voter Model (UVM), 123
Smart kiosks, 136 Voter Model, 123
Influence maximization in the field, 57 Multidisciplinary study on the relationship
Conclusion, limitations, and lessons between foster care attributes and
learned, 74 Posttraumatic Stress Disorder Symptoms
Introduction, 57 (PTSD) on foster youth, 169
Pilot study pipeline, 59 Data preprocessing for models, 175
Pilot study process, 60 Data samples and procedures, 173
Results, 63 Discussion and conclusions, 190
Influence maximization with unknown Introduction and statistics, 169
network structure, 77 Machine learning within the social sciences,
ARISEN: Approximating with Random 171
walks to Influence a Socially Explored Modeling techniques and evaluations, 178
Network Algorithm, 80 PTSD symptoms among youth with
Exploratory influence maximization, 78 homelessness and foster care histories,
Introduction, 77 170

https://doi.org/10.1017/9781108669016.017 Published online by Cambridge University Press

 Index 257

POMDP: Partially Observable Markov Minimizing violence in homeless youth,

Decision Process 119
Model, 48
YBW: Young Black women
RAAM Risk Amplification and Abatement Algorithms for solving the maximum
Model influence problem, 102
Definition, 25 Conclusions and future Work, 113
Findings, 25 Introduction and HIV statistics, 93
RAM Risk Amplification Model Mathematically modeling information
Definition, 25 diffusion, 98
Findings, 25 Maximizing the spread of sexual health
information in a multimodal
SBM: Stochastic Block Model communication network, 93
Influence maximization, 79 Quantitative network survey, 97
Sexual health information on HIV/AIDS, 211 Results and discussion, 104
Data gathering, 215 YMSM: Young men who have sex with men
Evaluation metrics, 221 HIV risk, 24
Future work and directions, 226 Youth homelessness
Motivating social problem, 211 AI for improving access to sexual health
NPCEditor, 216 necessities, 136
SHIELD Self-Help in Eliminating Causes, 19
Life-Threatening Diseases Census, 17
Defined, 27 Couch surfing, 17
Social networks Depression, 21
Raise HIV awareness, 38 Experiences of violence (overview),
Role in HIV prevention, 27 21
The importance of, 25 Extent of, 17
Social work Greedy minimization, 124
Definition, 7 Have You Heard?, 16
Ethics and Artificial Intelligence in Public HIV prevention, 26
Health Social Work, 231 HIV risk, 23
Goals, 8 Importance of social networks, 25
Grand Challenges of, 5 Minimizing violence in homeless youth,
Substance use 119
Homeless youth, 22 Pregnancy, 23
Travelers PTSD: Posttraumatic Stress Disorder in
Social network, 21 homeless youth, 21
Subpopulation of homelessness, 20 Sexual- and gender-minority, 20
Sexual violence, 21
Violence Substance use/abuse, 22
Artificial intelligence to predict intimate YMSM: Young men who have sex with
partner violence perpetration, 195 men, 24

https://doi.org/10.1017/9781108669016.017 Published online by Cambridge University Press

https://doi.org/10.1017/9781108669016.017 Published online by Cambridge University Press