Data Science Solutions With Python

Data Science
Solutions
with Python
Fast and Scalable Models Using
Keras, PySpark MLlib, H2O, XGBoost,
and Scikit-Learn
—
Tshepo Chris Nokeri
Data Science
Solutions with Python
Fast and Scalable Models Using
Keras, PySpark MLlib, H2O, XGBoost,
and Scikit-Learn
Tshepo Chris Nokeri

Data Science Solutions with Python: Fast and Scalable Models Using Keras, PySpark
MLlib, H2O, XGBoost, and Scikit-Learn
Tshepo Chris Nokeri

Pretoria, South Africa
ISBN-13 (pbk): 978-1-4842-7761-4 ISBN-13 (electronic): 978-1-4842-7762-1

https://doi.org/10.1007/978-1-4842-7762-1
Copyright © 2022 by Tshepo Chris Nokeri

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
Trademarked names, logos, and images may appear in this book. Rather than use a trademark symbol with
every occurrence of a trademarked name, logo, or image we use the names, logos, and images only in an
editorial fashion and to the benefit of the trademark owner, with no intention of infringement of the
trademark.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not
identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to
proprietary rights.
While the advice and information in this book are believed to be true and accurate at the date of publication,
neither the authors nor the editors nor the publisher can accept any legal responsibility for any errors or
omissions that may be made. The publisher makes no warranty, express or implied, with respect to the
material contained herein.
Managing Director, Apress Media LLC: Welmoed Spahr
Acquisitions Editor: Celestin Suresh John
Development Editor: Laura Berendson
Coordinating Editor: Aditee Mirashi
Cover designed by eStudioCalamar
Cover image designed by Freepik (www.freepik.com)
Distributed to the book trade worldwide by Springer Science+Business Media New York, 1 New York Plaza,
Suite 4600, New York, NY 10004-1562, USA. Phone 1-800-SPRINGER, fax (201) 348-4505, e-mail orders-ny@
springer-sbm.com, or visit www.springeronline.com. Apress Media, LLC is a California LLC and the sole
member (owner) is Springer Science + Business Media Finance Inc (SSBM Finance Inc). SSBM Finance Inc
is a Delaware corporation.
For information on translations, please e-mail booktranslations@springernature.com; for reprint,
paperback, or audio rights, please e-mail bookpermissions@springernature.com.
Apress titles may be purchased in bulk for academic, corporate, or promotional use. eBook versions and
licenses are also available for most titles. For more information, reference our Print and eBook Bulk Sales
web page at http://www.apress.com/bulk-sales.
Any source code or other supplementary material referenced by the author in this book is available to
readers on GitHub via the book’s product page, located at www.apress.com/9781484277614. For more
detailed information, please visit http://www.apress.com/source-code.
Printed on acid-free paper
I dedicate this book to my family and everyone who has merrily played
influential roles in my life.
Table of Contents
About the Author�� ix
About the Technical Reviewer�� xi

Acknowledgments�� xiii
Introduction��xv
Chapter 1: Exploring Machine Learning�� 1

Exploring Supervised Methods�� 1
Exploring Nonlinear Models�� 2
Exploring Ensemble Methods�� 3
Exploring Unsupervised Methods�� 3
Exploring Cluster Methods�� 3
Exploring Dimension Reduction�� 4
Exploring Deep Learning�� 4
Conclusion�� 5
Chapter 2: Big Data, Machine Learning, and Deep Learning Frameworks�� 7

Big Data�� 7
Big Data Features�� 8
Impact of Big Data on Business and People�� 8
Better Customer Relationships�� 8
Refined Product Development�� 9
Improved Decision-Making�� 9
Big Data Warehousing�� 9
Big Data ETL�� 9
Big Data Frameworks�� 10
Apache Spark�� 10
v
Table of Contents
ML Frameworks�� 13
Scikit-Learn�� 13
H2O�� 13
XGBoost�� 14
DL Frameworks�� 14
Keras�� 14
Chapter 3: Linear Modeling with Scikit-Learn, PySpark, and H2O�� 15

Exploring the Ordinary Least-Squares Method�� 15
Scikit-Learn in Action�� 17
PySpark in Action�� 20
H2O in Action�� 22
Conclusion�� 28
Chapter 4: Survival Analysis with PySpark and Lifelines�� 29

Exploring Survival Analysis�� 29
Exploring Cox Proportional Hazards Method�� 29
Lifeline in Action�� 30
Exploring the Accelerated Failure Time Method�� 34
Conclusion�� 37
Chapter 5: Nonlinear Modeling With Scikit-Learn, PySpark, and H2O�� 39

Exploring the Logistic Regression Method�� 39
Conclusion�� 57
vi
Table of Contents
Chapter 6: Tree Modeling and Gradient Boosting with Scikit-Learn, XGBoost,

PySpark, and H2O�� 59
Decision Trees�� 59
Preprocessing Features�� 60
Gradient Boosting�� 66
XGBoost in Action�� 66
Conclusion�� 74
Chapter 7: Neural Networks with Scikit-Learn, Keras, and H2O�� 75

Exploring Deep Learning�� 75
Multilayer Perceptron Neural Network�� 75
Keras in Action�� 82
Deep Belief Networks�� 87
Conclusion�� 88
Chapter 8: Cluster Analysis with Scikit-Learn, PySpark, and H2O�� 89

Exploring the K-Means Method�� 89
Conclusion�� 99
vii
Table of Contents
Chapter 9: Principal Component Analysis with Scikit-Learn,

PySpark, and H2O�� 101
Exploring the Principal Component Method�� 101
Conclusion�� 110
Chapter 10: Automating the Machine Learning Process with H2O�� 111
Exploring Automated Machine Learning�� 111
H2O AutoML in Action�� 112
Conclusion�� 116
Index�� 117
viii
About the Author
Tshepo Chris Nokeri harnesses advanced analytics and
artificial intelligence to foster innovation and optimize
business performance. In his work, he delivered complex
solutions to companies in the mining, petroleum, and
manufacturing industries. He earned a Bachelor’s degree
in Information Management and then graduated with an
honour’s degree in Business Science from the University of
the Witwatersrand, on a TATA Prestigious Scholarship and
a Wits Postgraduate Merit Award. He was also unanimously
awarded the Oxford University Press Prize. He is the author of Data Science Revealed,
Implementing Machine Learning in Finance, and Econometrics and Data Science, all
published by Apress.
ix
About the Technical Reviewer
Joos Korstanje is a data scientist with over five years of industry experience in
developing machine learning tools, a large part of which has been forecasting models.
He currently works at Disneyland Paris, where he develops machine learning for a
variety of tools. His experience in writing and teaching have motivated him to contribute
to this book on advanced forecasting with Python.
xi
Acknowledgments
Writing a single-authored book is demanding, but I received firm support and active
encouragement from my family and dear friends. Many heartfelt thanks to the Apress
Publishing team for all their support throughout the writing and editing processes.
Lastly, my humble thanks to all of you for reading this; I earnestly hope you find it
helpful.
xiii
Introduction
This book covers the in-memory, distributed cluster computing framework called
PySpark, the machine learning framework platforms called Scikit-Learn, PySpark MLlib,
H2O, and XGBoost, and the deep learning framework known as Keras. After reading this
book, you will be able to apply supervised and unsupervised learning to solve practical
and real-world data problems. In this book, you will learn how to engineer features,
optimize hyperparameters, train and test models, develop pipelines, and automate the
machine learning process.
To begin, the book carefully presents supervised and unsupervised ML and DL
models and examines big data frameworks and machine learning and deep learning
frameworks. It also discusses the parametric model called Generalized Linear Model
and a survival regression model known as the Cox Proportional Hazards model and
Accelerated Failure Time (AFT). It presents a binary classification model called Logistic
Regression and an ensemble model called Gradient Boost Trees. It also introduces DL
and an artificial neural network, the Multilayer Perceptron (MLP) classifier. It describes
a way of performing cluster analysis using the k-means model. It explores dimension
reduction techniques like Principal Components Analysis and Linear Discriminant
Analysis and concludes by unpacking automated machine learning.
The book targets intermediate data scientists and machine learning engineers who
want to learn how to apply key big data frameworks, as well as ML and DL frameworks.
Before exploring the contents of this book, be sure that you understand basic statistics,
Python programming, probability theories, and predictive analytics.
The books uses Anaconda (an open source distribution of Python programming) for
the examples. The following list highlights some of the Python libraries that this book
covers.
• Pandas for data structures and tools.
• PySpark for in-memory, cluster computing.

• XGBoost for gradient boosting and survival regression analysis.
• Auto-Sklearn, Tree-based Pipeline Optimization Tool (TPOT),

Hyperopt-Sklearn, and H2O for AutoML.
xv
Introduction
• Scikit-Learn for building and validating key machine learning

algorithms.
• Keras for high-level frameworks for deep learning.
• H2O for driverless machine learning.
• Lifelines for survival analysis.
• NumPy for arrays and matrices.
• SciPy for integrals, solving differential equations, and optimization.
• Matplotlib and Seaborn for recognized plots and graphs.
xvi
CHAPTER 1
Exploring Machine
Learning
This chapter introduces the best machine learning methods and specifies the main
differences between supervised and unsupervised machine learning. It also discusses
various applications of both.
Machine learning has been around for a long time; however, it has recently gained
widespread recognition. This is because of the increased computational power of
modern computer systems and the ease of access to open source platforms and
frameworks. Machine learning involves inducing computer systems with intelligence by
implementing various programming and statistical techniques. It draws from fields such
as statistics, computational linguistics, and neuroscience, among others. It also applies
modern statistics and basic programming. It enables developers to develop and deploy
intelligent computer systems and create practical and reliable applications.
Exploring Supervised Methods

Supervised learning involves feeding a machine learning algorithm with a large data
set, through which it learns the desired output. The main machine learning methods
are linear, nonlinear, and ensemble. The following section familiarizes you with the
parametric method, also known as the linear regression method. See Table 1-1.
Linear regression methods expect data to follow a Gaussian distribution. The
ordinary least-squares method is the standard linear method; it minimizes the error in
terms and determines the slope and intercept to fit the data to a straight line (where the
dependent feature is continuous).
Furthermore, the ordinary least-squares method determines the extent of the
association between the features. It also assumes linearity, no autocorrelation in
residuals, and no multicollinearity. In the real world, data hardly ever comes from a
1
© Tshepo Chris Nokeri 2022
T. C. Nokeri, Data Science Solutions with Python, https://doi.org/10.1007/978-1-4842-7762-1_1
Chapter 1 Exploring Machine Learning
normal distribution; the method struggles to explain the variability, resulting in an

under-fitted or over-fitted model. To address this problem, we introduce a penalty term
to the equation to alter the model’s performance.
Table 1-1. Types of Parametric Methods

Model Description
Linear regression method Applied when there is one dependent feature (continuous feature) and
an independent feature (continuous feature of categorical). The main
linear regression methods are GLM, Ridge, Lasso, Elastic Net, etc.
Survival regression method Applied to time-event-related censored data, where the dependent
feature is categorical and the independent feature is continuous.
Time series analysis method Applied to uncovering patterns in sequential data and forecasting
future instances. Principal time series models include the ARIMA
model, SARIMA, Additive model, etc.
Exploring Nonlinear Models

Nonlinear (classification) methods differentiate classes (also called categories or
labels)—see Table 1-2. When there are two classes in the dependent feature, you’ll
implement binary classification methods. When there are more than two classes, you’ll
implement multiclass classification methods. There are innumerable functions for
classification, including sigmoid, tangent hyperbolic, and kernel, among others. Their
application depends on the context. Subsequent chapters will cover some of these
functions.
2
Table 1-2. Varying Nonlinear Models

Model Description
Binary classification method Applied when the categorical feature has only two possible
outcomes. The popular binary classifier is the logistic regression
model.
Multiclass classification Applied when the categorical feature has more than two
method possible outcomes. The main multi-class classifier is the
Linear Discriminant Analysis model (which can also be used for
dimension reduction).
Survival classification Applied when you’re computing the probabilities of an event
occurring using a categorical feature.
Exploring Ensemble Methods

Ensemble methods enable you to uncover linearity and nonlinearity. The main
ensembler is the random forest trees, which is often computationally demanding. In
addition, its performance depends on a variety of factors like the depth, iterations, data
splitting, and boosting.
Exploring Unsupervised Methods

In contrast to supervised machine learning, unsupervised learning contains no ground
truth. You expose the model to all the sample data, allowing it to guess about the pattern
of the data. The most common unsupervised methods are cluster-related methods.
Exploring Cluster Methods

Cluster methods are convenient for assembling common values in the data; they enable
you to uncover patterns in both structured and unstructured data. Table 1-3 highlights a
few cluster methods that you can use to guess the pattern of the data.
3
Table 1-3. Varying Cluster Models

Techniques Description
Centroid clustering Applied to determine the center of the data and draw data points toward
the center. The main centroid clustering method is the k-means method.
Density clustering Applied to determine where the data is concentrated. The main density
clustering model is the DBSCAN method.
Distribution clustering Identifies the probability of data points belonging to a cluster based on
some distribution. The main distribution clustering method is the Gaussian
Mixture method.
Exploring Dimension Reduction

Dimension reducers help determine the extent to which factors or components elucidate
related changes in data. Table 1-4 provides an overview of the chief dimension reducers.
Table 1-4. Main Dimension Reducers

Technique Description
Factor analysis Applied to determine the extent to which factors elucidate related changes
of features in the data.
Principal component Applied to determine the extent to which factors elucidate related changes
analysis of features in the data.
Exploring Deep Learning

Deep learning extends machine learning in that it uses artificial neural networks to
solve complex problems. You can use deep learning to discover patterns in big data.
Deep learning uses artificial neural networks to discover patterns in complex data.
The networks use an approach similar to that of animals, whereby neurons (nodes)
that receive input data from the environment scan the input and pass it to neurons in
successive layers to arrive at some output (which is understanding the complexity in the
data). There are several artificial neural networks that you can use to determine patterns
of behavior of a phenomenon, depending on the context. Table 1-5 highlights the chief
neural networks.
4
Table 1-5. Varying Neural Networks

Network Description
Restricted Boltzmann Machine (RBM) The most common neural network that contains only the
hidden and visible layers.
Multilayer Perceptron (MLP) A neural network that prolongs a restricted Boltzmann
machine with input, hidden and output layers.
Recurrent Neural Network (RNN) Serves as a sequential modeler.
Convolutional Neural Network (CNN) Serves as a dimension reducer and classifier.
Conclusion
This chapter covered two ways in which machines learn—via supervised and
unsupervised learning. It began by explaining supervised machine learning and
discussing the three types of supervised learning methods and their applications. It then
covered unsupervised learning techniques, dimension reduction, and cluster analysis.
5
CHAPTER 2
Big Data, Machine

Learning, and Deep
Learning Frameworks
This chapter carefully presents the big data framework used for parallel data processing
called Apache Spark. It also covers several machine learning (ML) and deep learning
(DL) frameworks useful for building scalable applications. After reading this chapter, you
will understand how big data is collected, manipulated, and examined using resilient
and fault-tolerant technologies. It discusses the Scikit-Learn, Spark MLlib, and XGBoost
frameworks. It also covers a deep learning framework called Keras. It concludes by
discussing effective ways of setting up and managing these frameworks.
Big data frameworks support parallel data processing. They enable you to contain big
data across many clusters. The most popular big data framework is Apache Spark, which
is built on the Hadoop framework.
B
ig Data
Big data means different things to different people. In this book, we define big data as
large amounts of data that we cannot adequately handle and manipulate using classic
methods. We must undoubtedly use scalable frameworks and modern technologies to
process and draw insight from this data. We typically consider data “big” when it cannot
fit within the current in-memory storage space. For instance, if you have a personal
computer and the data at your disposal exceeds your computer’s storage capacity, it’s big
data. This equally applies to large corporations with large clusters of storage space. We
often speak about big data when we use a stack with Hadoop/Spark.
7
Chapter 2 Big Data, Machine Learning, and Deep Learning Frameworks
Big Data Features

The features of big data are described as the four Vs—velocity, volume, variety, and
veracity. Table 2-1 highlights these features of big data.
Table 2-1. Big Data Features

Element Description
Velocity Modern technologies and improved connectivity enable you to generate data at an
unprecedented speed. Characteristics of velocity include batch data, near or real-time
data, and streams.
Volume The scale at which data increases. The nature of data sources and infrastructure
influence the volume of data. Characteristics of the volume include exabyte,
zettabyte, etc.
Variety Data can come from unique sources. Modern technological devices leave digital
footprints here and there, which increase the number of sources from which businesses
and people can get data. Characteristics of variety include the structure and complexity
of the data.
Veracity Data must come from reliable sources. Also, it must be of high quality, consistent, and
complete.
Impact of Big Data on Business and People

Without a doubt, big data affects the way we think and do business. Data-driven
organizations typically establish the basis for evidence-based management. Big data
involves measuring the key aspects of the business using quantitative methods. It helps
support decision-making. The next sections discuss ways in which big data affects
businesses and people.
Better Customer Relationships

Insights from big data help manage customer relationships. Organizations with big data
about their customers can study customers’ behavioral patterns and use descriptive
analytics to drive customer-management strategies.
8
Refined Product Development

Data-driven organizations use big data analytics and predictive analytics to drive
product development. and management strategies. This approach is useful for
incremental and iterative delivery of applications.
I mproved Decision-Making
When a business has big data, it can use it to uncover complex patterns of a
phenomenon to influence strategy. This approach helps management make well-
informed decisions based on evidence, rather than on subjective reasoning. Data-driven
organizations foster a culture of evidence-based management.
We also use big data in fields like life sciences, physics, economics, and medicine.
There are many ways in which big data affects the world. This chapter does not consider
all factors. The next sections explain big data warehousing and ETL activities.
Big Data Warehousing

Over the past few decades, organizations have invested in on-premise databases,
including Microsoft Access, Microsoft SQL Server, SAP Hana, Oracle Database, and many
more. There has recently been widespread adoption of cloud databases like Microsoft
Azure SQL and Oracle XE. There are also standard big data (distributed) databases like
Cassandra and HBase, among others. Businesses are shifting toward scalable cloud-
based databases to harness benefits associated with increasing computational power,
fault-tolerant technologies, and scalable solutions.
Big Data ETL

Although there have been significant advances in database management, the way that
people manipulate data from databases remains the same. Extracting, transforming, and
loading (ETL) still play an integral part in analysis and reporting. Table 2-2 discusses ETL
activities.
9
Table 2-2. ETL Activities

Activity Description
Extract Involves getting data from some database.

Transforming Involves converting data from a database into a suitable format for analysis and
reporting
Loading Involves warehousing data in a database management system.
To perform ETL activities, you must use a query language. The most popular query
language is SQL (Standard Query Language). There are other query languages that
emerged with the open source movement, such as HiveQL and BigQuery. The Python
programming language supports SQL. Python frameworks can connect to databases by
implementing libraries, such as SQLAlchemy, pyodbc, SQLite, SparkSQL, and pandas,
among others.
Big Data Frameworks

Big data frameworks enable developers to collect, manage, and manipulate distributed
data. Most open source big data frameworks use in-memory cluster computing. The
most popular frameworks include Hadoop, Spark, Flink, Storm, and Samza. This book
uses PySpark to perform ETL activities, explore data, and build machine learning
pipelines.
Apache Spark
Apache Spark executes in-memory cluster computing. It enables developers to
build scalable applications using Java, Scala, Python, R, and SQL. It includes cluster
components like the driver, cluster manager, and executor. You can use it as a standalone
cluster manager or on top of Mesos, Hadoop, YARN, or Baronets. You can use it to access
data in the Hadoop File System (HDFS), Cassandra, HBase, and Hive, among other data
sources. The Spark data structure is considered a resilient distributed data set. This book
introduces a framework that integrates both Python and Apache Spark (PySpark). The
book uses it to operate Spark MLlib. To understand this framework, you first need to
grasp the idea behind resilient distributed data sets.
10
Resilient Distributed Data Sets

Resilient Distributed Data Sets (RDDs) are immutable elements for parallelizing data
or for transforming existing data. Chief RDD operations include transformation and
actions. We store them in any storage supported by Hadoop. For instance, in a Hadoop
Distributed File System (HDF), Cassandra, HBase, Amazon S3, etc.
S
park Configuration
Areas of Spark configuration include Spark properties, environment variables, and
logging. The default configuration directory is SPARK_HOME/conf.
You can install the findspark library in your environment using pip install
findspark and install the pyspark library using pip install pyspark.
Listing 2-1 prepares the PySpark framework using the findspark framework.
Listing 2-1. Prepare the PySpark Framework
import findspark as initiate_pyspark

initiate_pyspark.init("filepath\spark-3.0.0-bin-hadoop2.7")
Listing 2-2 stipulates the PySpark app using the SparkConf() method.
Listing 2-2. Stipulate the PySpark App
from pyspark import SparkConf

pyspark_configuration = SparkConf().setAppName("pyspark_linear_method").
setMaster("local")
Listing 2-3 prepares the PySpark session with the SparkSession() method.
Listing 2-3. Prepare the Spark Session
from pyspark.sql import SparkSession

pyspark_session = SparkSession(pyspark_context)
S
park Frameworks
Spark frameworks extend the core of the Spark API. There are four main Spark
frameworks—SparkSQL, Spark Streaming, Spark MLlib, and GraphX.
11
SparkSQL
SparkSQL enables you to use relational query languages like SQL, HiveQL, and Scala. It
includes a schemaRDD that has row objects and schema. You create it using an existing
RDD, parquet file, or JSON data set. You execute the Spark Context to create a SQL
context.
Spark Streaming
Spark streaming is a scalable streaming framework that supports Apache Kafka, Apache
Flume, HDFS, and Apache Kensis, etc. It processes input data using DStream in small
batches you push using HDFS, databases, and dashboards. Recent versions of Python
do not support Spark Streaming. Consequently, we do not cover the framework in this
book. You can use a Spark Streaming application to read input from any data source and
store a copy of the data in HDFS. This allows you to build and launch a Spark Streaming
application that processes incoming data and runs an algorithm on it.
Spark MLlib
MLlib is an ML framework that allows you to develop and test ML and DL models.
In Python, the frameworks work hand-in-hand with the NumPy framework. Spark
MLlib can be used with several Hadoop data sources and incorporated alongside
Hadoop workflows. Common algorithms include regression, classification, clustering,
collaborative filtering, and dimension reduction. Key workflow utilities include feature
transformation, standardization and normalization, pipeline development, model
evaluation, and hyperparameter optimization.
GraphX
GraphX is a scalable and fault-tolerant framework for iterative and fast graph parallel
computing, social networks, and language modeling. It includes graph algorithms
such as PageRank for estimating the importance of each vertex in a graph, Connected
Components for labeling connected components of the graph with the ID of its lowest-
numbered vertex, and Triangle Counting for finding the number of triangles that pass
through each vertex.
12
M
L Frameworks
To solve ML problems, you need to have a framework that supports building and scaling
ML models. There is no shortage of ML models – there are innumerable frameworks
for ML. There are several ML frameworks that you can use. Subsequent chapters cover
frameworks like Scikit-Learn, Spark MLlib, H2O, and XGBoost.
S
cikit-Learn
The Scikit-Learn framework includes ML algorithms like regression, classification, and
clustering, among others. You can use it with other frameworks such as NumPy and
SciPy. It can perform most of the tasks required for ML projects like data processing,
transformation, data splitting, normalization, hyperparameter optimization, model
development, and evaluation. Scikit-Learn comes with most distribution packages that
support Python. Use pip install sklearn to install it in your Python environment.
H
2O
H2O is an ML framework that uses a driverless technology. It enables you to accelerate
the adoption of AI solutions. It is very easy to use, and it does not require any technical
expertise. Not only that, but it supports numerical and categorical data, including text.
Before you train the ML model, you must first load the data into the H2O cluster. It
supports CSV, Excel, and Parquet files. Default data sources include local file systems,
remote files, Amazon S3, HDFS, etc. It has ML algorithms like regression, classification,
cluster analysis, and dimension reduction. It can also perform most tasks required
for ML projects like data processing, transformation, data splitting, normalization,
hyperparameter optimization, model development, checking pointing, evaluation, and
productionizing. Use pip install h2o to install the package in your environment.
Listing 2-4 prepares the H2O framework.
Listing 2-4. Initializing the H2O Framework
import h2o
h2o.init()
13
X
GBoost
XGBoost is an ML framework that supports programming languages, including Python.
It executes gradient-boosted models that are scalable, and learns fast parallel and
distributed computing without sacrificing memory efficiency. Not only that, but it is
an ensemble learner. As mentioned in Chapter a, ensemble learners can solve both
regression and classification problems. XGBoost uses boosting to learn from the errors
committed in the preceding trees. It is useful when tree-based models are overfitted. Use
pip install xgboost to install the model in your Python environment.
D
L Frameworks
DL frameworks provide a structure that supports scaling artificial neural networks. You
can use it stand-alone or with other models. It typically includes programs and code
frameworks. Primary DL frameworks include TensorFlow, PyTorch, Deeplearning4j,
Microsoft Cognitive Toolkit (CNTK), and Keras.
K
eras
Keras is a high-level DL framework written using Python; it runs on top of an ML
platform known as TensorFlow. It is effective for rapid prototyping of DL models. You
can run Keras on Tensor Processing Units or on massive Graphic Processing Units. The
main Keras APIs include models, layers, and callbacks. Chapter 7 covers this framework.
Execute pip install Keras and pip install tensorflow to use the Keras framework.
14
CHAPTER 3
Linear Modeling
with Scikit-Learn,
PySpark, and H2O
This introductory chapter explains the ordinary least-squares method and executes it
with the main Python frameworks (i.e., Scikit-Learn, Spark MLlib, and H2O). It begins by
explaining the underlying concept behind the method.
Exploring the Ordinary Least-Squares Method

The ordinary least-squares method is used with data that has an output feature that is
not confined (it’s continuous). This method expects normality and linearity, and there
must be an absence of autocorrelation in the error of terms (also called residuals) and
multicollinearity. It is also highly prone to abnormalities in the data, so you might want
to use alternative methods like Ridge, Lasso, and Elastic Net if this method does not
serve you well.
Listing 3-1 attains the necessary data from a Microsoft CSV file.
Listing 3-1. Attain the Data
import pandas as pd
df = pd.read_csv(r"filepath\WA_Fn-UseC_-Marketing_Customer_Value_
Analysis.csv")
Listing 3-2 stipulates the names of columns to drop and then executes the drop()
method. It then stipulates axes as columns in order to drop the unnecessary columns in
the data.
15
Chapter 3 Linear Modeling with Scikit-Learn, PySpark, and H2O
Listing 3-2. Drop Unnecessary Features in the Data
drop_column_names = df.columns[[0, 6]]

initial_data = df.drop(drop_column_names, axis="columns")
Listing 3-3 attains the dummy values for the categorical features in this data.
Listing 3-3. Attain Dummy Features
initial_data.iloc[::, 0] = pd.get_dummies(initial_data.iloc[::, 0])

Listing 3-4 outlines the independent and dependent features.
Listing 3-4. Outline the Features
import numpy as np
int_x = initial_data.iloc[::,0:19]
fin_x = initial_data.iloc[::,19:21]
x_combined = pd.concat([int_x, fin_x], axis=1)
x = np.array(x_combined)
y = np.array(initial_data.iloc[::,19])
16
S
cikit-Learn in Action
Listing 3-5 randomly divides the dataframe.
Listing 3-5. Randomly Divide the Dataframe
from sklearn.model_selection import train_test_split

x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.2,
random_state=0)
Listing 3-6 scales the independent features.
Listing 3-6. Scale the Independent Features
from sklearn.preprocessing import StandardScaler

sk_standard_scaler = StandardScaler()
sk_standard_scaled_x_train = sk_standard_scaler.fit_transform(x_train)
sk_standard_scaled_x_test = sk_standard_scaler.transform(x_test)
Listing 3-7 executes the Scikit-Learn ordinary least-squares regression method.
Listing 3-7. Execute the Scikit-Learn Ordinary Least-Squares Regression Method
from sklearn.linear_model import LinearRegression

sk_linear_model = LinearRegression()
sk_linear_model.fit(sk_standard_scaled_x_train, y_train)
Listing 3-8 determines the best hyperparameters for the Scikit-Learn ordinary least-
squares regression method.
Listing 3-8. Determine the Best Hyperparameters for the Scikit-Learn Ordinary
Least-Squares Regression Method
from sklearn.preprocessing import GridSearchCV
sk_linear_model_param = {'fit_intercept':[True,False]}
sk_linear_model_param_mod = GridSearchCV(estimator=sk_linear_model,
param_grid=sk_linear_model_param, n_jobs=-1)
sk_linear_model_param_mod.fit(sk_standard_scaled_x_train, y_train)
17
print("Best OLS score: ", sk_linear_model_param_mod.best_score_)

print("Best OLS parameter: ", sk_linear_model_param_mod.best_params_)
Best OLS score: 1.0

Best OLS parameter: {'fit_intercept': True}
Listing 3-9 executes the Scikit-Learn ordinary least-squares regression method.
Listing 3-9. Execute the Scikit-Learn Ordinary Least-Squares Regression Method

sk_linear_model = LinearRegression(fit_intercept= True)
sk_linear_model.fit(sk_standard_scaled_x_train, y_train)
sk_linear_model_param_mod.best_params_)
Listing 3-10 computes the Scikit-Learn ordinary least-squares regression method’s

intercept.
Listing 3-10. Compute the Scikit-Learn Ordinary Least-Squares Regression

Method’s Intercept
print(sk_linear_model.intercept_)
433.0646521131769
Listing 3-11 computes the Scikit-Learn ordinary least-squares regression method’s

coefficients.

Method’s Coefficients
print(sk_linear_model.coef_)
[-6.15076155e-15 2.49798076e-13 -1.95573220e-14 -1.90089677e-14
-5.87187344e-14 2.50923806e-14 -1.05879478e-13 1.53591400e-14
-1.82507711e-13 -7.86327034e-14 4.17629484e-13 1.28923537e-14
6.52911311e-14 -5.28069778e-14 -1.57900159e-14 -6.74040176e-14
-9.28427833e-14 5.03132848e-14 -8.75978166e-15 2.90235705e+02
-9.55950515e-14]
Listing 3-12 computes the ordinary least-squares regression method’s predictions.
18

Method’s Predictions
sk_yhat = sk_linear_model.predict(sk_standard_scaled_x_test)
Listing 3-13 assesses the Scikit-Learn ordinary least-squares method (see Table 3-1).
Listing 3-13. Assess the Scikit-Learn Ordinary Least-Squares Method
from sklearn import metrics

sk_mean_ab_error = metrics.mean_absolute_error(y_test, sk_yhat)
sk_mean_sq_error = metrics.mean_squared_error(y_test, sk_yhat)
sk_root_sq_error = np.sqrt(sk_mean_sq_error)
sk_determinant_coef = metrics.r2_score(y_test, sk_yhat)
sk_exp_variance = metrics.explained_variance_score(y_test, sk_yhat)
sk_linear_model_ev = [[sk_mean_ab_error, sk_mean_sq_error, sk_root_sq_
error, sk_determinant_coef, sk_exp_variance]]
sk_linear_model_assessment = pd.DataFrame(sk_linear_model_ev, index =
["Estimates"], columns = ["Sk mean absolute error", "Sk mean squared error",
"Sk root mean squared error", "Sk determinant
coefficient", "Sk variance score"])
sk_linear_model_assessment
y = np.array(initial_data.iloc[::,19])
Table 3-1. Assessment of the Scikit-Learn Ordinary Least-Squares Method

Sk Mean Sk Mean Sk Root Mean Sk Determinant Sk Variance
Absolute Error Squared Error Squared Error Coefficient Score
Estimates 9.091189e-13 1.512570e-24 1.229866e-12 1.0 1.0
Table 3-1 shows that the Scikit-Learn ordinary least-squares method explains the
entire variability.
19
P
ySpark in Action
This section executes and assesses the ordinary least-squares method with the PySpark
framework. Listing 3-14 prepares the PySpark framework with the findspark framework.

Listing 3-15 stipulates the PySpark app with the SparkConf() method.

pyspark_configuration = SparkConf().setAppName("pyspark_linear_method").
setMaster("local")
Listing 3-16 prepares the PySpark session with the SparkSession() method.

Listing 3-17 changes the pandas dataframe created earlier in this chapter to a
PySpark dataframe using the createDataFrame() method.
Listing 3-17. Change Pandas Dataframe to a PySpark Dataframe
pyspark_initial_data = pyspark_session.createDataFrame(initial_data)
Listing 3-18 creates a list for independent features and a string for the dependent
feature. It converts data using the VectorAssembler() method for modeling with the
PySpark framework.
Listing 3-18. Transform the Data

x_list = list(x_combined.columns)
y_list = initial_data.columns[19]
from pyspark.ml.feature import VectorAssembler
20
pyspark_data_columns = x_list
pyspark_vector_assembler = VectorAssembler(inputCols=pyspark_data_columns,
outputCol="variables")
pyspark_data = pyspark_vector_assembler.transform(pyspark_initial_data)
Listing 3-19 divides the data using the randomSplit() method.
Listing 3-19. Divide the Dataframe

(pyspark_training_data, pyspark_test_data) = pyspark_data.
randomSplit([.8,.2])
Listing 3-20 executes the PySpark ordinary least-squares regression method.
Listing 3-20. Execute the PySpark Ordinary Least-Squares Regression Method
from pyspark.ml.regression import LinearRegression

pyspark_linear_model = LinearRegression(labelCol=y_list,
featuresCol=pyspark_data.columns[-1])
pyspark_fitted_linear_model = pyspark_linear_model.fit(pyspark_training_data)
Listing 3-21 computes the PySpark ordinary least-squares regression method’s

predictions.
Listing 3-21. Compute the PySpark Ordinary Least-Squares Regression Method’s

Predictions
pyspark_yhat = pyspark_fitted_linear_model.transform(pyspark_test_data)
Listing 3-22 assesses the PySpark ordinary least-squares method.
Listing 3-22. Assess the PySpark Ordinary Least-Squares Method
pyspark_linear_model_assessment = pyspark_fitted_linear_model.summary
print("PySpark root mean squared error", pyspark_linear_model_assessment.
rootMeanSquaredError)
print("PySpark determinant coefficient", pyspark_linear_model_assessment.r2)
PySpark root mean squared error 2.0762306526480097e-13

PySpark determinant coefficient 1.0
21
H
2O in Action
This section executes and assesses the ordinary least-squares method with the H2O
framework.
Listing 3-23. Prepare the H2O Framework
import h2o as initialize_h2o

initialize_h2o.init()
Listing 3-24 changes the pandas dataframe to an H2O dataframe.
Listing 3-24. Change the Pandas Dataframe to an H2O Dataframe
h2o_data = initialize_h2o.H2OFrame(initial_data)
Listing 3-25. Outline Features
y = y_list
x = h2o_data.col_names
x.remove(y_list)
Listing 3-26 randomly divides the data.
h2o_training_data, h2o_validation_data, h2o_test_data = h2o_data.split_

frame(ratios=[.8,.1])
Listing 3-27 executes the H2O ordinary least-squares regression method.
Listing 3-27. Execute the H2O Ordinary Least-Squares Regression Method
from h2o.estimators import H2OGeneralizedLinearEstimator

h2o_linear_model = H2OGeneralizedLinearEstimator(family="gaussian")
h2o_linear_model.train(x=x,y=y,training_frame=h2o_training_data,validation_
frame=h2o_validation_data)
22
Listing 3-28 computes the H2O ordinary least-squares method’s predictions.
Listing 3-28. H2O Ordinary Least-Squares Method Executed Predictions
h2o_yhat = h2o_linear_model.predict(h2o_test_data)
Listing 3-29 computes the H2O ordinary least-squares method’s standardized

coefficients (see Figure 3-1).
Listing 3-29. H2O Ordinary Least-Squares Method’s Standardized Coefficients
h2o_linear_model_std_coefficients = h2o_linear_model.std_coef_plot()
h2o_linear_model_std_coefficients
Figure 3-1. H2O ordinary least-squares method’s standardized coefficients
23
Listing 3-30 computes the H2O ordinary least-squares method’s partial dependency
(see Figure 3-2).
Listing 3-30. H2O Ordinary Least-Squares Method’s Partial Dependency
h2o_linear_model_dependency_plot = h2o_linear_model.partial_plot
(data = h2o_data, cols = list(initial_data.columns[[0,19]]), server=False,
plot = True)
h2o_linear_model_dependency_plot
24
Figure 3-2. H2O ordinary least-squares method’s partial dependency

25
Listing 3-31 arranges the features that are the most important to the H2O ordinary
least-squares method in ascending order (see Figure 3-3).
Listing 3-31. H2O Ordinary Least-Squares Method’s Feature Importance
h2o_linear_model_feature_importance = h2o_linear_model.varimp_plot()
h2o_linear_model_feature_importance
Figure 3-3. H2O ordinary least-squares method’s feature importance
Listing 3-32 assesses the H2O ordinary least-squares method.
Listing 3-32. Assess H2O Ordinary Least-Squares Method
h2o_linear_model_assessment = h2o_linear_model.model_performance()
print(h2o_linear_model_assessment)
ModelMetricsRegressionGLM: glm
** Reported on train data. **
MSE: 24844.712331260016
26
RMSE: 157.6220553452467
MAE: 101.79904883889066
RMSLE: NaN
R^2: 0.7004468136072375
Mean Residual Deviance: 24844.712331260016
Null degrees of freedom: 7325
Residual degrees of freedom: 7304
Null deviance: 607612840.7465751
Residual deviance: 182012362.53881088
AIC: 94978.33944003603
Listing 3-33 improves the performance of the H2O ordinary least-squares method by
specifying remove_collinear_columns as True.
Listing 3-33. Improve the Performance of the Ordinary Least-Squares Method
h2o_linear_model_collinear_removed = H2OGeneralizedLinearEstimator(family="
gaussian", lambda_ = 0,remove_collinear_columns = True)
h2o_linear_model_collinear_removed.train(x=x,y=y,training_frame=h2o_
training_data,validation_frame=h2o_validation_data)
Listing 3-34 assesses the H2O ordinary least-squares method.
Listing 3-34. Assess the H2O Ordinary Least-Squares Method
h2o_linear_model_collinear_removed_assessment = h2o_linear_model_collinear_
removed.model_performance()
print(h2o_linear_model_collinear_removed)
MSE: 23380.71864337616
RMSE: 152.9075493341521
MAE: 102.53007935777588
RMSLE: NaN
R^2: 0.7180982143647627
Null deviance: 607612840.7465751
27

AIC: 94533.40762597627
ModelMetricsRegressionGLM: glm
** Reported on validation data. **
MSE: 25795.936313899092
RMSE: 160.6111338416459
MAE: 103.18677222520363
RMSLE: NaN
R^2: 0.7310558588001701
Null deviance: 84181020.04623385
AIC: 11430.364002305443
Conclusion
This chapter executed three key machine learning frameworks (Scikit-Learn, PySpark,
and H2O) to model data and spawn a continuous output feature using a linear method.
It also explored ways of assessing that method.
28
CHAPTER 4
Survival Analysis with

PySpark and Lifelines
This chapter describes and executes several survival analysis methods using the main
Python frameworks (i.e., Lifelines and PySpark). It begins by explaining the underlying
concept behind the Cox Proportional Hazards model. It then introduces the accelerated
failure time method.
Exploring Survival Analysis

Survival methods are common in the manufacturing, insurance, and medical
science fields. They are convenient for properly assessing risk when an independent
investigation is carried out over long periods and subjects enter and leave at given times.
These methods are employed in this chapter to determine the probabilities of a
machine failing and patients surviving an illness, among other applications. They are
equally suitable for missing values (censored data).
Exploring Cox Proportional Hazards Method

The Cox Proportional Hazards method is the best survival method for handling censored
data with subjects that have related changes. It is similar to the Mantel-Haenszel method
and expects the hazard rate to be a function of time. It states that the rate is a function of
covariates. Equation 4-1 defines the Cox Proportional Hazards method.
t ,x ,x
1 2 ,,x p t exp b1 X 1 b2 X 2 bp X p (Equation 4-1)
Listing 4-1 attains the necessary data from a Microsoft Excel file.
29
Chapter 4 Survival Analysis withPySpark and Lifelines
import pandas as pd
initial_data = pd.read_excel(r"filepath\survival_data.xlsx", index_col=[0])
Listing 4-2 finds the ratio for dividing the data.
Listing 4-2. Find the Ratio for Dividing the Data
int(initial_data.shape[0]) * 0.8
345.6
Listing 4-3 divides the data.
Listing 4-3. Divide the Data
lifeline_training_data = initial_data.loc[:346]
lifeline_test_data = initial_data.loc[346:]
L ifeline in Action
This section executes and assesses the Cox Proportional Hazards method with the
Lifeline framework. Listing 4-4 executes the Lifeline Cox Proportional Hazards method.
Listing 4-4. Execute the Lifeline Cox Proportional Hazards Method
from lifelines import CoxPHFitter

lifeline_cox_method = CoxPHFitter()
lifeline_cox_method.fit(lifeline_training_data, initial_data.columns[0],
initial_data.columns[1])
Listing 4-5 computes the test statistics (see Table 4-1) and assesses the Lifeline
Cox Proportional Hazards method with a scaled Schoenfeld, which helps disclose any
abnormalities in the residuals (see Figure 4-1).
30
Listing 4-5. Compute the Lifeline Cox Proportional Hazards Method’s Test
Statistics and Residuals
lifeline_cox_method_test_statistics_schoenfeld = lifeline_cox_method.
check_assumptions(lifeline_training_data, show_plots=True)
lifeline_cox_method_test_statistics_schoenfeld
Table 4-1. Test Statistics for the Lifeline Cox Proportional Hazards Method
Test Statistic p
Age km 10.53 <0.005

rank 10.78 <0.005
Fin km 0.12 0.73
rank 0.14 0.71
Mar km 0.18 0.67
rank 0.20 0.66
Paro km 0.13 0.72
rank 0.11 0.74
Prio km 0.49 0.48
rank 0.47 0.49
Race km 0.34 0.56
rank 0.37 0.54
Wexp km 11.91 <0.005
rank 11.61 <0.005
31
Figure 4-1. Scaled Schoenfeld residuals of age
Listing 4-6 determines the Lifeline Cox Proportional Hazards method’s assessment
summary (see Table 4-2).
Listing 4-6. Compute the Assessment Summary
lifeline_cox_method_assessment_summary = lifeline_cox_method.
print_summary()
lifeline_cox_method_assessment_summary
32
Table 4-2. Summary of the Cox Proportional Hazards

Coef Exp(coef) Se(coef) Coef Coef Exp(coef) Exp(coef) Z P -log2(p)
Lower Upper Lower Upper 95%
95% 95% 95%
Fin -0.71 0.49 0.23 -1.16 -0.27 0.31 0.77 -3.13 <0.005 9.14
Age -0.03 0.97 0.02 -0.08 0.01 0.93 1.01 -1.38 0.17 2.57
Race 0.39 1.48 0.37 -0.34 1.13 0.71 3.09 1.05 0.30 1.76
Wexp -0.11 0.90 0.24 -0.59 0.37 0.56 1.44 -0.45 0.65 0.62
Mar -1.15 0.32 0.61 -2.34 0.04 0.10 1.04 -1.90 0.06 4.11
Paro 0.07 1.07 0.23 -0.37 0.51 0.69 1.67 0.31 0.76 0.40
Prio 0.10 1.11 0.03 0.04 0.16 1.04 1.17 3.24 <0.005 9.73
Listing 4-7 determines the log test confidence interval for each feature in the data
(see Figure 4-2).
Listing 4-7. Execute the Lifeline Cox Proportional Hazards Method
lifeline_cox_log_test_ci = lifeline_cox_method.plot()
lifeline_cox_log_test_ci
33
Figure 4-2. Log test confidence interval
Exploring the Accelerated Failure Time Method

The accelerated failure time method models the censored data with a log-linear function
to describe the log of the survival time. Likewise, it also assumes each instance is
independent.
P
ySpark in Action
This section executes the accelerated failure time method with the PySpark framework.
Listing 4-8 runs the PySpark framework with the findspark framework.

34

pyspark_configuration = SparkConf().setAppName("pyspark_aft_method").
setMaster("local")
Listing 4-10 prepares the PySpark session using the SparkSession() method.

Listing 4-11. Change the Pandas Dataframe to a PySpark Dataframe
feature. It the converts the data using the VectorAssembler() method for modeling with
the PySpark framework.
x_list = list(initial_data.iloc[::, 0:9].columns)

Listing 4-13 executes the PySpark accelerated failure time method.
35
Listing 4-13. Execute the PySpark Accelerated Failure Time Method
from pyspark.ml.regression import AFTSurvivalRegression

pyspark_accelerated_failure_method = AFTSurvivalRegression(censorCol=pyspa
rk_data.columns[1], labelCol=pyspark_data.columns[0],featuresCol="variables",)
pyspark_accelerated_failure_method_fitted = pyspark_accelerated_failure_
method.fit(pyspark_data)
Listing 4-14 computes the PySpark accelerated failure time method’s predictions.
Listing 4-14. Compute the PySpark Accelerated Failure Time Method’s

Predictions
pyspark_accelerated_failure_method_fitted.transform(pyspark_data).
select(pyspark_data.columns[1],"prediction")
pyspark_yhat.show()
+------+------------------+
|arrest| prediction|
+------+------------------+
| 1|18.883982665910125|
| 1| 16.88228128814963|
| 1|22.631360777172517|
| 0|373.13041474613107|
| 0| 377.2238319806288|
| 0| 375.8326538406928|
| 1| 20.9780526816987|
| 0| 374.6420738270714|
| 0| 379.7483494080467|
| 0| 376.1601473382181|
| 0| 377.1412349521787|
| 0| 373.7536844216336|
| 1| 36.36443059383637|
| 0|374.14261327949384|
| 1| 22.98494042401171|
| 1| 50.61463874375869|
36
| 1| 25.56399364288275|
| 0|379.61997114629696|
| 0| 384.3322960430372|
| 0|376.37634062210844|
+------+------------------+
Listing 4-15 computes the PySpark accelerated failure time method’s coefficients.
Listing 4-15. Compute the PySpark Accelerated Failure Time Method’s

Coefficients
pyspark_accelerated_failure_method_fitted.coefficients
DenseVector([0.0388, -1.7679, -0.0162, -0.0003, 0.0098, -0.0086, -0.0026,

0.0115, 0.0003])
C
onclusion
This chapter executed two key machine learning frameworks (Lifeline and PySpark) to
model censored data with the Cox Proportional Hazards and accelerated failure time
methods.
37
CHAPTER 5
Nonlinear Modeling
With Scikit-Learn,
PySpark, and H2O
This chapter executes and appraises a nonlinear method for binary classification (called
logistic regression) using a diverse set of comprehensive Python frameworks (i.e., Scikit-
Learn, Spark MLlib, and H2O). To begin, it clarifies the underlying concept behind the
sigmoid function.
Exploring the Logistic Regression Method

The logistic regression method unanimously accepts values and then models them
by executing a function (sigmoid) to anticipate values of a categorical output feature.
Equation 5-1 defines the sigmoid function, which applies to logistic regression (also see
Figure 5-1).
L ex
Sx (Equation 5-1)
1 ex e x 1
Both Equation 5-1 and Figure 5-1 suggest that the function produces binary output
values.
39
Chapter 5 Nonlinear Modeling With Scikit-Learn, PySpark, and H2O
Figure 5-1. Sigmoid function
Listing 5-1 attains the necessary data from a Microsoft CSV file using the pandas
framework.

import pandas as pd
df = pd.read_csv(r"C:\Users\i5 lenov\Downloads\banking.csv")
Listing 5-2 stipulates the names of columns to drop and then executes the drop()
method. It stipulates axes as columns in order to drop the unnecessary columns in the
data.
Listing 5-2. Drop Unnecessary Features in the Data
drop_column_names = df.columns[[8, 9, 10]]

40
Listing 5-3 attains dummy values for categorical features in the data.
Listing 5-3. Attain Dummy Features

Listing 5-4 drops the null values.
Listing 5-4. Drop Null Values
initial_data = initial_data.dropna()
S
This section executes and assesses the logistic regression method with the Scikit-Learn
framework. Listing 5-5 outlines the independent and dependent features.
import numpy as np
x = np.array(initial_data.iloc[::, 0:17])
y = np.array(initial_data.iloc[::,-1])

random_state=0)
Listing 5-7 scales the independent features.
41
Listing 5-7. Scale Independent Features

Listing 5-8 executes the Scikit-Learn logistic regression method.
Listing 5-8. Execute the Scikit-Learn Logistic Regression Method
from sklearn.linear_model import LogisticRegression

sk_logistic_regression_method = LogisticRegression()
sk_logistic_regression_method.fit(sk_standard_scaled_x_train, y_train)
Listing 5-9 determines the best hyperparameters for the Scikit-Learn logistic
regression method.
Listing 5-9. Determine the Best Hyperparameters for the Scikit-Learn Logistic
Regression Method
from sklearn.model_selection import GridSearchCV
sk_logistic_regression_method_param = {"penalty":("l1","l2")}
sk_logistic_regression_method_param_mod = GridSearchCV(estimator=sk_
logistic_regression_method, param_grid=sk_logistic_regression_method_param,
n_jobs=-1)
sk_logistic_regression_method_param_mod.fit(sk_standard_scaled_x_train, y_train)
print("Best logistic regression score: ", sk_logistic_regression_method_
param_mod.best_score_)
print("Best logistic regression parameter: ", sk_logistic_regression_
method_param_mod.best_params_)
Best logistic regression score: 0.8986039453717755
Best logistic regression parameter: {'penalty': 'l2'}
Listing 5-10 executes the logistic regression method with the Scikit-Learn framework.
Listing 5-10. Execute the Scikit-Learn Logistic Regression Method
sk_logistic_regression_method = LogisticRegression(penalty="l2")
sk_logistic_regression_method.fit(sk_standard_scaled_x_train, y_train)
42
Listing 5-11 computes the logistic regression method’s intercept.
Listing 5-11. Compute the Logistic Regression Method’s Intercept
print(sk_logistic_regression_method.intercept_)
[-2.4596243]
Listing 5-12 computes the coefficients.
Listing 5-12. Compute the Logistic Regression Method’s Coefficients
print(sk_logistic_regression_method.coef_)
[[ 0.03374725 0.04330667 -0.01305369 -0.02709009 0.13508899 0.01735913
0.00816758 0.42948983 -0.12670658 -0.25784955 -0.04025993 -0.14622466
-1.14143485 0.70803518 0.23256046 -0.02295578 -0.02857435]]
Listing 5-13 computes the Scikit-Learn logistic regression method’s confusion

matrix, which includes two forms of errors—false positives and false negatives and true
positives and true negatives (see Table 5-1).
Listing 5-13. Compute the Scikit-Learn Logistic Regression Method’s Confusion

Matrix
sk_logistic_regression_method_assessment_1 = pd.DataFrame(metrics.
confusion_matrix(y_test, sk_yhat), index=["Actual: Deposit","Actual: No
deposit"], columns=("Predicted: deposit","Predicted: No deposit"))
print(sk_logistic_regression_method_assessment_1)
Table 5-1. Scikit-Learn Logistic Regression Method’s Confusion Matrix

Predicted: Deposit Predicted: No Deposit
Actual: Deposit 7230 95

Actual: No Deposit 711 202
Listing 5-14 computes the appropriate classification report (see Table 5-2).
43
Listing 5-14. Compute the Scikit-Learn Logistic Regression Method’s

Classification Report
sk_logistic_regression_method_assessment_2 = pd.DataFrame(metrics.
classification_report(y_test, sk_yhat, output_dict=True)).transpose()
print(sk_logistic_regression_method_assessment_2)
Table 5-2. Scikit-Learn Logistic Regression Method’s Classification Report

Precision Recall F1-score Support
0 0.910465 0.987031 0.947203 7325.000000

1 0.680135 0.221249 0.333884 913.000000
Accuracy 0.902161 0.902161 0.902161 0.902161
Macro Avg 0.795300 0.604140 0.640544 8238.000000
Weighted Avg 0.884938 0.902161 0.879230 8238.000000
Listing 5-15 arranges the Scikit-Learn logistic regression method’s receiver operating
characteristics curve. The goal is to condense the arrangement of the true positive rate (the
proclivity of the method to correctly differentiate positive classes) and the false positive
rate (the proclivity of the method to correctly differentiate negative classes). See Figure 5-2.
Listing 5-15. Receiver Operating Characteristics Curve for the Scikit-Learn

Logistic Regression Method
sk_yhat_proba = sk_logistic_regression_method.predict_proba(sk_standard_
scaled_x_test)[::,1]
fpr_sk_logistic_regression_method, tprr_sk_logistic_regression_method, _ =
metrics.roc_curve(y_test, sk_yhat_proba)
area_under_curve_sk_logistic_regression_method = metrics.roc_auc_score(y_
test, sk_yhat_proba)
plt.plot(fpr_sk_logistic_regression_method, tprr_sk_logistic_regression_
method, label="AUC= "+ str(area_under_curve_sk_logistic_regression_method))
plt.xlabel("False Positive Rate (FPR)")
plt.ylabel("True Positive Rate (TPR)")
plt.legend(loc="best")
plt.show()
44
Figure 5-2. Receiver operating characteristics curve for the Scikit-Learn logistic
regression method
Listing 5-16 arranges the Scikit-Learn logistic regression method’s precision-recall

curve to condense the arrangement of the precision and recall (see Figure 5-3).
Listing 5-16. Precision-Recall Curve for the Scikit-Learn Logistic Regression

Method
p_sk_logistic_regression_method, r__sk_logistic_regression_method, _ =
metrics.precision_recall_curve(y_test, sk_yhat)
weighted_ps_sk_logistic_regression_method = metrics.roc_auc_score(y_test,
sk_yhat)
plt.plot(p_sk_logistic_regression_method, r__sk_logistic_regression_method,
label="WPR= " +str(weighted_ps_sk_logistic_regression_method))
plt.xlabel("Recall")
plt.ylabel("Precision")
plt.show()
45
Figure 5-3. Precision-recall curve for the Scikit-Learn logistic regression

method
Listing 5-17 arranges the learning curve for the Scikit-Learn logistic regression
method to disclose the variations in weighted training and cross-validation accuracy
(see Figure 5-4).
Listing 5-17. Learning Curve for the Logistic Regression Method Executed by
Scikit-Learn
from sklearn.model_selection import learning_curve
train_port_sk_logistic_regression_method, trainscoresk_logistic_regression_
method, testscoresk_logistic_regression_method = learning_curve(sk_
logistic_regression_method, x, y,
cv=3, n_jobs=-5, train_sizes=np.linspace(0.1,1.0,50))
trainscoresk_logistic_regression_method_mean = np.mean(trainscoresk_
logistic_regression_method, axis=1)
46
testscoresk_logistic_regression_method_mean = np.mean(testscoresk_logistic_
regression_method, axis=1)
plt.plot(train_port_sk_logistic_regression_method, trainscoresk_logistic_
regression_method_mean, label="Weighted training accuracy")
plt.plot(train_port_sk_logistic_regression_method, testscoresk_logistic_
regression_method_mean, label="Weighted cv accuracy Score")
plt.xlabel("Training values")
plt.ylabel("Weighted accuracy score")
plt.show()
Figure 5-4. Learning curve for the logistic regression method executed by
Scikit-Learn
47
P
ySpark in Action
This section executes and assesses the logistic regression method with the PySpark
framework.


pyspark_configuration = SparkConf().setAppName("pyspark_logistic_
regression_method").setMaster("local")

Listing 5-22 creates a list of independent features and a string for the dependent
feature. It then converts the data using the VectorAssembler() method for modeling
with the PySpark framework.

y_list = str(initial_data.columns[-1])
48
outputCol="features")
Listing 5-23 divides the data using the randomSplit() method.

(pyspark_training_data, pyspark_test_data) = pyspark_data.randomSplit([.8,.2])
Listing 5-24 executes the PySpark logistic regression method.
Listing 5-24. Execute the PySpark Logistic Regression Method
from pyspark.ml.classification import LogisticRegression

pyspark_logistic_regression_method = LogisticRegression(labelCol = y_list,
featuresCol = "features")
pyspark_logistic_regression_method_fitted = pyspark_logistic_regression_
method.fit(pyspark_training_data)
Listing 5-25 computes the PySpark logistic regression method’s predictions.
Listing 5-25. Logistic Regression Method Predictions (Method Executed with

PySpark Framework)
pyspark_yhat = pyspark_logistic_regression_method_fitted.transform(pyspark_
test_data)
Listing 5-26 arranges the PySpark logistic regression method’s receiver operating
characteristics curve to condense the arrangement of the precision and recall (see
Figure 5-5).
Listing 5-26. Receiver Operating Characteristics Curve for the PySpark Logistic
Regression Method
pyspark_logistic_regression_method_assessment = pyspark_logistic_
regression_method_fitted.summary
pyspark_logistic_regression_method_roc = pyspark_logistic_regression_
method_assessment.roc.toPandas()
pyspark_logistic_regression_method_auroc = pyspark_logistic_regression_
method_assessment.areaUnderROC
49
plt.plot(pyspark_logistic_regression_method_roc["FPR"], pyspark_logistic_
regression_method_roc["TPR"],
label="AUC= "+str(pyspark_logistic_regression_method_auroc))
plt.legend(loc=4)
plt.show()
Figure 5-5. Receiver operating characteristics curve for the PySpark logistic
regression method
Listing 5-27 arranges the PySpark logistic regression method’s precision-recall curve
to condense the arrangement of the precision and recall (Figure 5-6).
50
Listing 5-27. Precision-Recall Curve for the PySpark Logistic Regression Method
pyspark_logistic_regression_method_assessment = pyspark_logistic_
regression_method_fitted.summary
pyspark_logistic_regression_method_assessment_pr = pyspark_logistic_
regression_method_assessment.pr.toPandas()
pyspark_logistic_regression_method_assessment_wpr = pyspark_logistic_
regression_method_assessment.weightedPrecision
plt.plot(pyspark_logistic_regression_method_assessment_pr["precision"],
pyspark_logistic_regression_method_assessment_pr["recall"],
label="WPR: "+str(pyspark_logistic_regression_method_assessment_
wpr))
plt.xlabel("Precision")
plt.ylabel("Recall")
plt.show()
Figure 5-6. Precision-recall curve for the PySpark logistic regression method
51
H
2O in Action
This section executes and assesses the logistic regression method using the H2O
framework. Listing 5-28 prepares the H2O framework.

Listing 5-29 changes the pandas dataframe to the H2O dataframe.
Listing 5-29. Change the Pandas Dataframe to H2O Dataframe

y = y_list
x.remove(y_list)

52
Listing 5-32 executes the H2O logistic regression method.
Listing 5-32. Execute the H2O Logistic Regression Method
from h2o.estimators.glm import H2OGeneralizedLinearEstimator

h2o_logistic_regression_method = H2OGeneralizedLinearEstimator(family=
"binomial")
h2o_logistic_regression_method.train(x = x, y = y, training_frame =
h2o_training_data, validation_frame = h2o_validation_data)
Listing 5-33 computes the H2O logistic regression method’s predictions.
Listing 5-33. Compute the H2O Logistic Regression Method’s Predictions
h2o_yhat = h2o_logistic_regression_method.predict(h2o_test_data)
Listing 5-34 computes the H2O logistic regression method’s predictions (see
Figure 5-7).
Listing 5-34. Compute the H2O Logistic Regression Method’s Standardized

Coefficients
h2o_logistic_regression_method_std_coefficients = h2o_logistic_regression_
method.std_coef_plot()
h2o_logistic_regression_method_std_coefficients
53
Figure 5-7. H2O logistic regression method’s standardized coefficients
Listing 5-35 computes the H2O logistic regression method’s partial dependence
(see Figure 5-8).
Listing 5-35. Compute the H2O Logistic Regression Method’s Partial

Dependency
h2o_logistic_regression_method_dependency_plot = h2o_logistic_regression_
method.partial_plot(data = h2o_data, cols = list(initial_data.columns
[[0, 17]]), server=False, plot = True)
h2o_logistic_regression_method_dependency_plot
54
Figure 5-8. H2O logistic regression method’s standardized coefficients

55
Listing 5-36 arranges the features that are most important to the H2O logistic
regression method in ascending order (see Figure 5-9).
Listing 5-36. Compute the H2O Logistic Regression Method’s Variance

Importance
h2o_logistic_regression_method_feature_importance = h2o_logistic_
regression_method.varimp_plot()
h2o_logistic_regression_method_feature_importance
Figure 5-9. H2O logistic regression method’s variance importance
Listing 5-37 arranges the H2O logistic regression method’s receiver operating
characteristics curve to condense the arrangement of the true positive rate and false
positive rate (see Figure 5-10).
56
Listing 5-37. Receiver Operating Characteristics Curve for the H2O Logistic
Regression Method
h2o_logistic_regression_method_assessment = h2o_logistic_regression_method.
model_performance()
Figure 5-10. Receiver operating characteristics curve for the H2O logistic
regression method
Conclusion
and H2O) to model data and spawn a categorical output feature with two classes using
the logistic regression method.
57
CHAPTER 6
Tree Modeling and

Gradient Boosting with
Scikit-Learn, XGBoost,
PySpark, and H2O
This chapter executes and appraises a tree-based method (the decision tree method)
and an ensemble method (the gradient boosting trees method) using a diverse set of
comprehensive Python frameworks (i.e., Scikit-Learn, XGBoost, PySpark, and H2O). To
begin, the chapter clarifies how decision trees compute the probabilities of classes.
D
ecision Trees
The decision tree is an elementary, non-parametric method suitable for linear and
nonlinear modeling. It executes decision rules and develops a tree-like structure that
divides values into varying groups with the least depth (see a Figure 6-1).
59
Chapter 6 Tree Modeling and Gradient Boosting with Scikit-Learn, XGBoost, PySpark, and H2O
Figure 6-1. Decision tree
P
reprocessing Features
This chapter manipulates the data from Chapter 5, so it does not sequentially cover the
preprocessing tasks. Listing 6-1 executes all the preprocessing tasks.
Listing 6-1. Preprocess Features
from sklearn.tree import DecisionTreeClassifier

import numpy as np
import pandas as pd
df = pd.read_csv(r"filepath\banking.csv")
60

x = np.array(initial_data.iloc[::, 0:17])
y = np.array(initial_data.iloc[::, -1])
random_state=0)
S
This section executes and assesses the decision tree method with the Scikit-Learn
framework. Listing 6-2 executes the decision tree method with the Scikit-Learn
framework.
Listing 6-2. Execute the Scikit-Learn Decision Tree Method
from sklearn.tree import DecisionTreeClassifier

sk_decision_tree_method = DecisionTreeClassifier()
sk_decision_tree_method.fit(sk_standard_scaled_x_train, y_train)
Listing 6-3 determines the best hyperparameters for the Scikit-Learn decision tree
method.
Listing 6-3. Determine the Best Hyperparameters for the Scikit-Learn Decision
Tree Method
from sklearn.model_selection import GridSearchCV
sk_decision_tree_method_parameters = {"criterion":("gini","entropy"),
"max_depth":[1, 2, 3, 4, 5, 6]}
sk_decision_tree_method_g_search = GridSearchCV(estimator = sk_decision_
tree_method, param_grid = sk_decision_tree_method_parameters)
sk_decision_tree_method_g_search.fit(sk_standard_scaled_x_train, y_train)
print("Best decision tree regression score: ", sk_decision_tree_method_g_
search.best_score_)
print("Best decision tree parameter: ", sk_decision_tree_method_g_search.
best_params_)
61
Best decision tree regression score: 0.900030349013657

Best decision tree parameter: {'criterion': 'entropy', 'max_depth': 5}
Listing 6-4 executes the decision tree method using the Scikit-Learn framework.
Listing 6-4. Execute the Scikit-Learn Decision Tree Method
sk_decision_tree_method = DecisionTreeClassifier(criterion = "entropy",

max_depth = 5)
sk_decision_tree_method.fit(sk_standard_scaled_x_train, y_train)
Listing 6-5 arranges the Scikit-Learn decision tree method’s classification report (see
Table 6-1).
Listing 6-5. Arrange the Scikit-Learn Decision Tree Method’s Classification

Report
sk_yhat = sk_decision_tree_method.predict(sk_standard_scaled_x_test)
sk_decision_tree_method_assessment_2 = pd.DataFrame(metrics.classification_
report(y_test, sk_yhat, output_dict=True)).transpose()
sk_decision_tree_method_assessment_2
Table 6-1. Scikit-Learn Decision Tree Method’s Classification Report

0 0.914696 0.982253 0.947271 7325.000000

1 0.650538 0.265060 0.376654 913.000000
Accuracy 0.902768 0.902768 0.902768 0.902768
Macro Avg 0.782617 0.623656 0.661963 8238.000000
Weighted Avg 0.885420 0.902768 0.884031 8238.000000
Listing 6-6 arranges the decision tree method’s receiver operating characteristics
curve to condense the arrangement of the true positive rate and the false positive rate
(see Figure 6-2).
62
Listing 6-6. Receiver Operating Characteristics Curve for the Decision Tree
Method (Executed by the Scikit-Learn Framework)
import matplotlib.pyplot as plt
%matplotlib inline
sk_yhat_proba = sk_decision_tree_method.predict_proba(sk_standard_scaled_
x_test)[::,1]
fpr_sk_decision_tree_method, tprr_sk_decision_tree_method, _ = metrics.
roc_curve(y_test, sk_yhat_proba)
area_under_curve_sk_decision_tree_method = metrics.roc_auc_
score(y_test, sk_yhat_proba)
plt.plot(fpr_sk_decision_tree_method, tprr_sk_decision_tree_method,
label="AUC= "+ str(area_under_curve_sk_decision_tree_method))
plt.show()
Figure 6-2. Receiver operating characteristics curve for the Scikit-Learn decision
63
Listing 6-7 arranges the Scikit-Learn decision tree method’s receiver operating
Figure 6-3).
Listing 6-7. Precision-Recall Curve for the Scikit-Learn Decision Tree Method
p_sk_decision_tree_method, r__sk_decision_tree_method, _ = metrics.

precision_recall_curve(y_test, sk_yhat)
weighted_ps_sk_decision_tree_method = metrics.roc_auc_score(y_test, sk_yhat)
plt.plot(p_sk_decision_tree_method, r__sk_decision_tree_method,
label="WPR= " +str(weighted_ps_sk_decision_tree_method))
plt.show()
Figure 6-3. Precision-recall curve for the Scikit-Learn framework decision tree
method
64
Listing 6-8 arranges the Scikit-Learn decision tree method’s learning curve (see
Figure 6-4).
Listing 6-8. Learning Curve for the Decision Tree Method Executed by the
Scikit-Learn Framework
train_port_sk_decision_tree_method, trainscore_sk_decision_tree_method,
testscore_sk_decision_tree_method = learning_curve(sk_decision_tree_method,
x, y, cv=3, n_jobs=-5, train_sizes=np.
linspace(0.1,1.0,50))
trainscoresk_decision_tree_method_mean = np.mean(trainscore_sk_decision_
tree_method, axis=1)
testscoresk_decision_tree_method_mean = np.mean(testscore_sk_decision_tree_
method, axis=1)
plt.plot(train_port_sk_decision_tree_method, trainscoresk_decision_tree_
method_mean, label="Weighted training accuracy")
plt.plot(train_port_sk_decision_tree_method, testscoresk_decision_tree_
method_mean, label="Weighted cv accuracy Score")
plt.show()
65
Figure 6-4. Learning curve for the decision tree method executed by Scikit-Learn
G
radient Boosting
Gradient boosting methods inherit values of input features and then execute countless
tree models to halt the loss function (i.e., the mean absolute error for linear modeling).
It does this by assimilating weak models, then incrementally and iteratively models
weighing data, diligently accompanied by an election of a weak model with the best
performance.
X
GBoost in Action
This section executes and assesses the decision tree method with the XGBoost
framework. Listing 6-9 executes the XGBoost gradient boosting method.
Listing 6-9. Execute the XGBoost Gradient Boosting Method

from xgboost import XGBClassifier
xgb_gradient_boosting_method = XGBClassifier()
xgb_gradient_boosting_method.fit(sk_standard_scaled_x_train, y_train)
66
Listing 6-10 arranges the XGBoost gradient boosting method’s classification report
(see Table 6-2).
Listing 6-10. Arrange the XGBoost Gradient Boosting Method’s Classification

Report
sk_yhat_xgb_gradient_boosting_method = sk_yhat_xgb_gradient_boosting_method
= xgb_gradient_boosting_method.predict(sk_standard_scaled_x_test)
xgb_gradient_boosting_method_assessment_2 = pd.DataFrame(metrics.
classification_report(y_test, sk_yhat_xgb_gradient_boosting_method, output_
dict=True)).transpose()
print(xgb_gradient_boosting_method_assessment_2)
Table 6-2. XGBoost Gradient Boosting Method’s Classification Report

0 0.913410 0.976382 0.943847 7325.00000

1 0.575980 0.257393 0.355791 913.000000
Accuracy 0.896698 0.896698 0.896698 0.896698
Macro Avg 0.744695 0.616888 0.649819 8238.000000
Weighted Avg 0.876013 0.896698 0.878674 8238.000000
Listing 6-11 arranges the XGBoost gradient boosting method’s receiver operating
Figure 6-5).
Listing 6-11. Receiver Operating Characteristics Curve for the XGBoost Gradient
Boosting Method
yhat_proba_xgb_gradient_boosting_method = xgb_gradient_boosting_method.
predict_proba(sk_standard_scaled_x_test)[::,1]
fpr_xgb_gradient_boosting_method, tprr_xgb_gradient_boosting_method, _ =
metrics.roc_curve(y_test, yhat_proba_xgb_gradient_boosting_method)
area_under_curve_xgb_gradient_boosting_method = metrics.roc_auc_score(y_
test, yhat_proba_xgb_gradient_boosting_method)
67
plt.plot(fpr_xgb_gradient_boosting_method, tprr_xgb_gradient_boosting_
method, label="AUC= "+ str(area_under_curve_xgb_gradient_boosting_method))
plt.show()
Figure 6-5. Receiver operating characteristics curve for the XGBoost gradient
boosting method
Listing 6-12 arranges the XGBoost gradient boosting method’s precision-recall curve
to condense the arrangement of the precision and recall (see Figure 6-6).
Listing 6-12. Precision-Recall Curve for the Scikit-Learn Decision Tree Method
sk_yhat_xgb_gradient_boosting_method)
weighted_ps_xgb_gradient_boosting_method = metrics.roc_auc_score(y_test,
sk_yhat_xgb_gradient_boosting_method)
68
plt.plot(p_xgb_gradient_boosting_method, r__xgb_gradient_boosting_method,
label="WPR= " +str(weighted_ps_xgb_gradient_boosting_method))
plt.show()
Figure 6-6. Precision-recall curve for the Scikit-Learn decision tree method
P
ySpark in Action
This section executes and assesses the gradient boosting method with the PySpark
framework. Listing 6-13 prepares the PySpark framework using the findspark framework.

69

pyspark_configuration = SparkConf().setAppName("gradient_boosting_method").
setMaster("local")

feature. It then converts the data using the VectorAssembler() method to model with

Listing 6-18 divides the data with the randomSplit() method.
70
(pyspark_training_data, pyspark_test_data) = pyspark_data.

randomSplit([.8,.2])
Listing 6-19 executes the PySpark gradient boosting regression method.
Listing 6-19. Execute the PySpark Gradient Boosting Method
from pyspark.ml.classification import GBTClassifier

pyspark_gradient_boosting_method = GBTClassifier(labelCol = y_list,
featuresCol = "features")
pyspark_gradient_boosting_method_fitted = pyspark_gradient_boosting_method.
fit(pyspark_training_data)
Listing 6-20 computes the gradient boosting regression method’s predictions using
Listing 6-20. Gradient Boosting Method Predictions (Method Executed with

PySpark Framework)
pyspark_yhat = pyspark_gradient_boosting_method_fitted.transform(pyspark_
test_data)
H
2O in Action
This section executes and assesses the principal component method with the H2O
framework. Listing 6-21 prepares the H2O framework.

71
Listing 6-23 outlines the features.

y = y_list
x.remove(y_list)

Listing 6-25 executes the H2O gradient boosting method.
Listing 6-25. Execute the H2O Gradient Boosting Method
from h2o.estimators import H2OGradientBoostingEstimator

h2o_gradient_boosting_method = H2OGradientBoostingEstimator(nfolds=3)
h2o_gradient_boosting_method.train(x = x, y = y, training_frame = h2o_
training_data, validation_frame = h2o_validation_data)
Listing 6-26 assesses the H2O gradient boosting method (see Table 6-3).
Listing 6-26. Assess the H2O Gradient Boosting Method
h2o_gradient_boosting_method_history = h2o_gradient_boosting_method.
scoring_history()
print(h2o_gradient_boosting_method_history.head(5))
72
Chapter 6
Table 6-3. H2O Gradient Boosting Method Assessment

Timestamp Duration Number_ Training_ Training_ Training_ Validation_ Validation_ Validation_
of_trees rmse mae deviance rmse mae deviance
0 2021- 10.384 0.0 0.317003 0.200981 0.100491 0.311883 0.197762 0.097271

08-24 sec
04:04:30
1 2021- 10.496 1.0 0.309924 0.196310 0.096053 0.305219 0.193472 0.093159
08-24 sec
04:04:30
2 2021- 10.586 2.0 0.304085 0.192122 0.092468 0.299762 0.189611 0.089857
08-24 sec
04:04:30
3 2021- 10.659 3.0 0.299238 0.188351 0.089543 0.295379 0.186199 0.087249
08-24 sec
04:04:30
4 2021- 10.745 4.0 0.295198 0.184939 0.087142 0.291736 0.183116 0.085110
08-24 sec
04:04:30
73
Tree Modeling and Gradient Boosting with Scikit-Learn, XGBoost, PySpark, and H2O
Conclusion
This chapter executed four key machine learning frameworks (Scikit-Learn, XGBoost,
PySpark, and H2O) to model data and spawn a categorical output feature with two
classes using the decision tree and gradient boosting methods.
74
CHAPTER 7
Neural Networks with

Scikit-Learn, Keras,
and H2O
This chapter executes and assesses nonlinear neural networks to address binary
classification using a diverse set of comprehensive Python frameworks (i.e., Scikit-Learn,
Keras, and H2O).
Exploring Deep Learning

Deep learning methods transcend machine learning methods. They handle considerable
complexity in models and methods by including extensive hidden layers. Machine
learning methods typically resemble an input-output system, where an objective
function accepts input values and transforms them with an objective function to
produce values. Deep learning broadens the processing activities by including many
layers that include nodes that repetitively and iteratively process data. The widespread
recognition of deep learning relates to its customizability; it enable developers to
customize methods based on preferences.
Multilayer Perceptron Neural Network

As the name suggests, a multilayer perceptron neural network contains multiple layers.
Moreover, the fundamental structure remains the same; there has to be one layer for
receiving input values and one layer for generating output values. Layers that model the
data to distribute differing weights and biases are treated as hidden layers.
75
Chapter 7 Neural Networks with Scikit-Learn, Keras, and H2O
Figure 7-1 shows the multilayer perceptron neural network that this chapter uses.
age
job
marital
education
default
housing
loan
Deposit
contact
campaign
pdays No deposit
previous
poutcome
emp_var_rate
cons_price_idx
cons_conf_idx
euribor3m
nr_employed
Figure 7-1. Multilayer perceptron neural network
P
This chapter manipulates the data attained in Chapter 6, so it does not describe the
preprocessing tasks in detail. Listing 7-1 executes all the preprocessing tasks.
76
import numpy as np
import pandas as pd
df = pd.read_csv(r"C:\Users\i5 lenov\Downloads\banking.csv")
x = np.array(initial_data.iloc[::,0:17])
y = np.array(initial_data.iloc[::,-1])
random_state=0)
S
This section executes and assesses a multilayer perceptron method using the Scikit-Learn
framework. Listing 7-2 executes the Scikit-Learn multilayer perceptron neural network.
Listing 7-2. Execute the Scikit-Learn Multilayer Perceptron Neural Network
from sklearn.neural_network import MLPClassifier

sk_multilayer_perceptron_net = MLPClassifier()
sk_multilayer_perceptron_net.fit(sk_standard_scaled_x_train, y_train)
77
Listing 7-3 arranges the Scikit-Learn multilayer perceptron neural network’s

classification report (see Table 7-1).
Listing 7-3. Arrange the Scikit-Learn Multilayer Perceptron Neural Network’s

Classification Report
sk_yhat_multilayer_perceptron_net = sk_multilayer_perceptron_net.
predict(sk_standard_scaled_x_test)
sk_multilayer_perceptron_net_assessment = pd.DataFrame(metrics.
classification_report(y_test, sk_yhat_multilayer_perceptron_net,
output_dict=True)).transpose()
print(sk_multilayer_perceptron_net_assessment)
Table 7-1. Scikit-Learn Multilayer Perceptron Neural Network’s Classification

Report
0 0.917062 0.976655 0.945921 7325.000000

1 0.608696 0.291347 0.394074 913.000000
Accuracy 0.900704 0.900704 0.900704 0.900704
Macro Avg 0.762879 0.634001 0.669998 8238.000000
Weighted Avg 0.882886 0.900704 0.884761 8238.000000
Listing 7-4 arranges the multilayer perceptron neural network’s receiver operating
characteristics curve to condense the arrangement of the true positive rate and the false
positive rate (see Figure 7-2).
Listing 7-4. Arrange a Receiver Operating Characteristics Curve for the

Multilayer Perceptron Network (Executed with the Scikit-Learn Framework)
%matplotlib inline
yhat_proba_sk_multilayer_perceptron_net = sk_multilayer_perceptron_net.
predict_proba(sk_standard_scaled_x_test)[::,1]
78
fpr_sk_multilayer_perceptron_net, tprr_sk_multilayer_perceptron_net, _ =
metrics.roc_curve(y_test, yhat_proba_sk_multilayer_perceptron_net)
area_under_curve_sk_multilayer_perceptron_net = metrics.roc_auc_score(y_
test, yhat_proba_sk_multilayer_perceptron_net)
plt.plot(fpr_sk_multilayer_perceptron_net, tprr_sk_multilayer_perceptron_
net, label="AUC= "+ str(area_under_curve_sk_multilayer_perceptron_net))
plt.show()
Figure 7-2. Receiver operating characteristics curve for the multilayer perceptron
network executed with the Scikit-Learn framework
Listing 7-5 arranges the Scikit-Learn multilayer perceptron neural network’s

precision-recall curve to condense the arrangement of the precision and recall (see
Figure 7-3).
79
Listing 7-5. Precision-Recall Curve for the Scikit-Learn Multilayer Perceptron

Neural Network
p_sk_multilayer_perceptron_net, r_sk_multilayer_perceptron_net, _ =
metrics.precision_recall_curve(y_test, sk_yhat_multilayer_perceptron_net)
weighted_ps_sk_multilayer_perceptron_net = metrics.roc_auc_score(y_test,
sk_yhat_multilayer_perceptron_net)
plt.plot(p_sk_multilayer_perceptron_net, r_sk_multilayer_perceptron_net,
label="WPR= " +str(weighted_ps_sk_multilayer_perceptron_net))
plt.show()
Figure 7-3. Precision-recall curve for the Scikit-Learn multilayer perceptron

neural network
80
Listing 7-6 arranges the Scikit-Learn multilayer perceptron neural network’s learning
curve (see Figure 7-4).
Listing 7-6. Arrange a Learning Curve for the Multilayer Perceptron Network
Executed with the Scikit-Learn Framework
train_port_sk_multilayer_perceptron_net, trainscore_sk_multilayer_
perceptron_net, testscore_sk_multilayer_perceptron_net = learning_
curve(sk_multilayer_perceptron_net, x, y, cv=3, n_jobs=-5, train_sizes=np.
linspace(0.1,1.0,50))
trainscoresk_multilayer_perceptron_net_mean = np.mean(trainscore_sk_
multilayer_perceptron_net, axis=1)
testscoresk_multilayer_perceptron_net_mean = np.mean(testscore_sk_
multilayer_perceptron_net, axis=1)
plt.plot(train_port_sk_multilayer_perceptron_net, trainscoresk_multilayer_
perceptron_net_mean, label="Weighted training accuracy")
plt.plot(train_port_sk_multilayer_perceptron_net, testscoresk_multilayer_
perceptron_net_mean, label="Weighted cv accuracy Score")
plt.show()
81
Figure 7-4. Learning curve for the multilayer perceptron network executed with
the Scikit-Learn framework
K
eras in Action
This section executes and assesses a deep belief neural network using the Keras
framework. Listing 7-7 preprocesses the features.
Listing 7-7. Feature Preprocessing

random_state=0)
x_train, x_val, y_train, y_val = train_test_split(x_train,y_train,test_
size=0.2,random_state=0)
82
Listing 7-8 employs the tensorflow framework as a backend and installs the Keras
framework.
Listing 7-8. Employ the Tensorflow Framework as Backend and Install the Keras
Framework
import tensorflow as tf
from tensorflow.keras import Sequential, regularizers
from tensorflow.keras.layers import Dense
from tensorflow.keras.wrappers.scikit_learn import KerasClassifier
Listing 7-9 structures a multilayer perceptron neural network with two hidden layers
that have 17 neurons. The input layer contains a "l1" regularizer with a 0.001 rate, the
relu activation function, and binary_crossentropy loss and accuracy metrics.
Listing 7-9. Structure a Multilayer Perceptron with Keras
def keras_multilayer_perceptron_net(optimizer="adam"):
keras_multilayer_perceptron_net_model = Sequential()
keras_multilayer_perceptron_net_model.add(Dense(17, input_dim=17,
activation="sigmoid",kernel_regularizer=regularizers.l1(0.001), bias_
regularizer=regularizers.l1(0.01)))
keras_multilayer_perceptron_net_model.add(Dense(17, activation="relu"))
keras_multilayer_perceptron_net_model.compile(loss="binary_
crossentropy", optimizer=optimizer, metrics=["accuracy"])
return keras_multilayer_perceptron_net_model
Listing 7-10 binds the multilayer perceptron neural network.
Listing 7-10. Bind the Multilayer Perceptron Neural Network
keras_multilayer_perceptron_net_model = KerasClassifier(build_fn=keras_
multilayer_perceptron_net)
Listing 7-11 executes the Keras multilayer perceptron neural network with 56 epochs
and a batch_size of 14.
83
Listing 7-11. Execute the Multilayer Perceptron Neural Network with the Keras
Framework
keras_multilayer_perceptron_net_model_history = keras_multilayer_
perceptron_net_model.fit(sk_standard_scaled_x_train, y_train, validation_
data=(x_val, y_.val), batch_size=14, epochs=56)
print(keras_multilayer_perceptron_net_model_history)
Listing 7-12 arranges a classification report for the multilayer perceptron network
executed with the Keras framework (see Table 7-2).
Listing 7-12. Arrange a Classification Report for the Multilayer Perceptron

Network Executed with the Keras Framework
keras_multilayer_perceptron_net_model.predict(sk_standard_scaled_x_test)
keras_multilayer_perceptron_net_model = pd.DataFrame(metrics.
classification_report(y_test, keras_yhat_multilayer_perceptron_net,
output_dict=True)).transpose()
print(keras_multilayer_perceptron_net_model)
Table 7-2. Classification Report for the Multilayer Perceptron Network Executed
with the Keras Framework
0 0.908443 0.991536 0.948172 7325.000000

1 0.744856 0.198248 0.313149 913.000000
Accuracy 0.903617 0.903617 0.903617 0.903617
Macro Avg 0.826649 0.594892 0.630661 8238.000000
Weighted Avg 0.890313 0.903617 0.877794 8238.000000
Listing 7-13 arranges the training and CV loss for a multilayer perceptron network
executed with the Keras framework (see Figure 7-5).
84
Listing 7-13. Training and CV Loss for Multilayer Perceptron Network Executed
by the Keras Framework
plt.plot(keras_multilayer_perceptron_net_model_history.history["loss"],
label="Training loss")
plt.plot(keras_multilayer_perceptron_net_model_history.history["val_loss"],
label="CV loss")
plt.xlabel("Epochs")
plt.ylabel("Loss")
plt.show()
Figure 7-5. Training and CV loss for multilayer perceptron network executed with
the Keras framework
Listing 7-14 arranges the training and CV accuracy for the multilayer perceptron
network executed with the Keras framework (see Figure 7-6).
85
Listing 7-14. Arrange the Training and CV Accuracy for Multilayer Perceptron
Network Executed with the Keras Framework
plt.plot(keras_multilayer_perceptron_net_model_history.history["accuracy"],
label="Training accuracy score")
plt.plot(keras_multilayer_perceptron_net_model_history.history
["val_accuracy"], label="CV accuracy score")
plt.xlabel("Epochs")
plt.ylabel("Accuracy")
plt.show()
Figure 7-6. Training and CV accuracy for the multilayer perceptron network
executed with the Keras framework
86
Deep Belief Networks

Deep belief networks transcend the multilayer perceptron neural network by including
hidden layers. The network must model the data repetitively before generating an output
value.
H
2O in Action
This section executes a deep belief neural network using the H2O framework.

Listing 7-16. Change the Pandas Dataframe to the H2O Dataframe
Listing 7-17outlines the features.

y = y_list
x.remove(y_list)

Listing 7-19 executes the H2O deep belief neural network.
87
Listing 7-19. Execute the H2O Deep Belief Neural Network
from h2o.estimators.deeplearning import H2ODeepLearningEstimator

h2o_deep_belief_net = H2ODeepLearningEstimator(hidden=[5], epochs=56)
h2o_deep_belief_net.train(x= x, y= y, training_frame = h2o_training_data,
validation_frame = h2o_validation_data)h2o_training_data, validation_frame
= h2o_validation_data)
Conclusion
This chapter executed three key machine learning frameworks (Scikit-Learn, Keras, and
H2O) to model the data. It produced a binary outcome using the multilayer perceptron
and deep belief network. It also applied the binary_crossentropy loss function (to
assess the neural networks) and the adam optimizer (to enhance the neural network’s
performance).
88
CHAPTER 8
Cluster Analysis
with Scikit-Learn,
PySpark, and H2O
This chapter explains the k-means cluster method by implementing a diverse set of
Python frameworks (i.e., Scikit-Learn, PySpark, and H2O). To begin, it clarifies how the
method apportions values to clusters.
Exploring the K-Means Method

The k-means method is the most common distance-computing method. It is part of
the unsupervised machine learning family. It employs the Euclidean distance objective
function to efficiently compute the distance between values, then determines the centers
and draws values adjacent to the centers to establish clusters.
Compared to other cluster methods (i.e., spatial cluster methods, birch, and
agglomerative), this method requires you to stipulate the number of k before training.
This chapter uses the elbow curve, which discloses fluctuations in the linear altered
vectors to determine the number of k. This chapter executes the method to group
customers in clusters based on specific characteristics (i.e., age, remuneration, and
spending habits).
import pandas as pd
df = pd.read_csv(r"filepath\Mall_Customers.csv")
89
Chapter 8 Cluster Analysis with Scikit-Learn, PySpark, and H2O
Listing 8-2 drops the unnecessary features in the data.
Listing 8-2. Drop Unnecessary Features

Listing 8-3 scales the data with a standard scaler.
Listing 8-3. Scale the Data

sk_standard_scaled_data = sk_standard_scaler.fit_transform(initial_data)
Listing 8-4 executes the principal component method using the Scikit-Learn
framework.
Listing 8-4. Execute the Principal Component Method with the Scikit-Learn
Framework
from sklearn.decomposition import PCA
sk_principal_component_method = PCA(n_components=3)
sk_principal_component_method.fit(sk_standard_scaled_data)
sk_principal_component_method_transformed_data = sk_principal_component_
method.transform(sk_standard_scaled_data)
Listing 8-5 discloses the number of k using an elbow curve (see Figure 8-1).
Listing 8-5. Disclose the Number of K Using an Elbow Curve
from sklearn.cluster import KMeans

%matplotlib inline
elbow_range = range(1, 15)
sk_kmeans_method = [KMeans(n_clusters=x) for x in elbow_range]
sk_kmeans_method_score = [sk_kmeans_method[x].fit(sk_principal_component_
method_transformed_data).score(sk_principal_component_method_transformed_
data) for x in range(len(sk_kmeans_method))]
plt.plot(elbow_range, sk_kmeans_method_score)
90
plt.xlabel("Number of k")
plt.ylabel("Eigenvalues")
plt.show()
Figure 8-1. Disclose the number of k using an elbow curve
Figure 8-1 shows that the curve abruptly twists, which indicates that the k-means
method must contain three components.
S
This section executes the k-means method using the Scikit-Learn framework;
see Listing 8-6.
Listing 8-6. Execute the K-Means Method Using the Scikit-Learn Framework
sk_kmeans_method = KMeans(n_clusters=3)
sk_kmeans_method_fitted = sk_kmeans_method.fit(sk_standard_scaled_data)
91
Listing 8-7 computes the Scikit-Learn k-means method’s labels.
Listing 8-7. Compute the Scikit-Learn K-Means Method’s Labels
sk_kmeans_method_labels = pd.DataFrame(sk_kmeans_method_fitted.labels_,
columns = ["Labels"])
Listing 8-8 computes the cluster centers and determines their mean and standard
deviation (see Table 8-1).
Listing 8-8. Compute Cluster Centers and Their Mean and Standard Deviation
sk_kmeans_method = KMeans(n_clusters=3)
sk_kmeans_method_centers = sk_kmeans_method_fitted.cluster_centers_
sk_kmeans_method_centers = pd.DataFrame(sk_kmeans_method_centers, columns =
("1st center","2nd center","3rd center"))
sk_kmeans_method_centers_mean = pd.DataFrame(sk_kmeans_method_centers.
mean())
sk_kmeans_method_centers_std = pd.DataFrame(sk_kmeans_method_centers.std())
sk_kmeans_method_centers_mean_std = pd.concat([sk_kmeans_method_centers_
mean, sk_kmeans_method_centers_std], axis=1)
sk_kmeans_method_centers_mean_std.columns =["Center mean", "Center standard
deviation"]
print(sk_kmeans_method_centers_mean_std)
Table 8-1. Centroid Statistics

Center Mean Center Standard Deviation
First Center -0.157489 0.942836

Second Center 0.129950 0.854012
Third Center 0.222984 0.891719
Listing 8-9 determines the scaled data and labels that the Scikit-Learn k-means
method computed (see Figure 8-2).
92
Listing 8-9. Determine the PySpark K-Means Method ‘s Labels
plt.scatter(sk_principal_component_method_transformed_data[:,0],
sk_principal_component_method_transformed_data[:, 1],
c=sk_kmeans_method_fitted.labels_,cmap="coolwarm", s=120)
plt.xlabel("y")
plt.show()
Figure 8-2. Scikit-Learn k-means method labels
P
ySpark in Action
This section executes and assesses the k-means method using the PySpark framework.
93


pyspark_configuration = SparkConf().setAppName("pyspark_kmeans_method").
setMaster("local")

Listing 8-14 creates a list for the independent features and a string for the dependent

94
Listing 8-15 executes the k-means method using the PySpark framework.
Listing 8-15. Execute the K-Means Method Using the PySpark Framework
from pyspark.ml.clustering import KMeans

pyspark_kmeans_method = KMeans().setK(3).setSeed(1)
pyspark_kmeans_method_fitted = pyspark_kmeans_method.fit(pyspark_data)
Listing 8-16 computes the k-means method’s labels and passes them to a pandas
dataframe.
Listing 8-16. Compute the PySpark K-Means Method ‘s Labels
pyspark_yhat = pyspark_kmeans_method_fitted.transform(pyspark_data)
pyspark_yhat_pandas_df = pyspark_yhat.toPandas()
Listing 8-17 shows the scaled data and labels that the PySpark k-means method
computed (see Figure 8-3).
Listing 8-17. Disclose the PySpark K-Means Method’s Labels
from mpl_toolkits.mplot3d import Axes3D

figure = plt.figure(figsize=(12, 12))
ax = Axes3D(figure)
ax.scatter(pyspark_yhat_pandas_df.iloc[::, 0],
pyspark_yhat_pandas_df.iloc[::, 1],
pyspark_yhat_pandas_df.iloc[::, 2],
c = pyspark_yhat_pandas_df.iloc[::, 4],
cmap = "coolwarm",
s = 120)
ax.set_xlabel(initial_data.columns[0])
ax.set_ylabel(initial_data.columns[1])
ax.set_zlabel(initial_data.columns[2])
figure.show()
95
Figure 8-3. PySpark k-means method labels
Listing 8-18 computes the PySpark k-means method’s centers.
Listing 8-18. Compute the PySpark K-Means Method’s Centers
pyspark_kmeans_method_centers = pyspark_kmeans_method_fitted.
clusterCenters()
print("K-Means method cluster centers")
for pyspark_cluster_center in pyspark_kmeans_method_centers:
print(pyspark_cluster_center)
K-Means method cluster centers
96
[26.5952381 33.14285714 65.66666667]
[45.30769231 60.52991453 34.47008547]
[32.97560976 88.73170732 79.24390244]
Listing 8-19 assesses the PySpark k-means method using the Silhouette method.
Listing 8-19. Assess the PySpark K-Means Method Using the Silhouette Method
from pyspark.ml.evaluation import ClusteringEvaluator

pyspark_kmeans_method_assessment = ClusteringEvaluator()
pyspark_kmeans_method_silhoutte = pyspark_kmeans_method_assessment.
evaluate(pyspark_yhat)
print("Silhouette score = " + str(pyspark_kmeans_method_silhoutte))
Silhouette score = 0.45097978411312134
H
2O in Action
This section executes and assesses the k-means method using the H2O framework.

Listing 8-21 changes the pandas dataframe to an H2O dataframe .
y = y_list
x.remove(y_list)
Listing 8-23 reserves data for training and validating the H2O k-means method.
97
h2o_training_data, h2o_validation_data = h2o_data.split_frame(ratios=[.8])
Listing 8-24 executes the H2O k-means method.
Listing 8-24. Execute the K-Means Method Using the H2O Framework
from h2o.estimators import H2OKMeansEstimator

h2o_kmeans_method = H2OKMeansEstimator(k=3)
h2o_kmeans_method.train(x = x_list, training_frame = h2o_training_data,
validation_frame = h2o_validation_data)
Listing 8-25 computes the H2O k-means method’s labels.
Listing 8-25. Compute the H2O K-Means Method Labels
h2o_yhat = h2o_kmeans_method.predict(h2o_validation_data)
Listing 8-26 assesses the H2O k-means method (see Table 8-2).
Listing 8-26. Assess the K-Means Method Executed with the H2O Framework
from h2o.estimators import H2OKMeansEstimator

h2o_kmeans_method_assessment = h2o_kmeans_method.model_performance()
h2o_kmeans_method_assessment
ModelMetricsClustering: kmeans
MSE: NaN
RMSE: NaN
Total Within Cluster Sum of Square Error: 238.6306804611187
Total Sum of Square Error to Grand Mean: 479.99999503427534
Between Cluster Sum of Square Error: 241.36931457315666
98
Table 8-2. PySpark Centroid Statistics

Centroid Size Within_cluster_sum_of_squares
0 1.0 78.0 134.983658

1 2.0 51.0 69.748489
2 3.0 32.0 33.898533
Conclusion
and H2O) in order to group data into three clusters. You saw how to identify the number
of k using the elbow curve. To assess the method, you employed the Silhouette method
and looked at the sum of squared errors.
99
CHAPTER 9
Principal Component
Analysis with Scikit-Learn,
PySpark, and H2O
This chapter executes a simple dimension reducer (a principal component method) by
implementing a diverse set of Python frameworks (Scikit-Learn, PySpark, and H2O). To
begin, it clarifies how the method computes components.
Exploring the Principal Component Method

The principal component method is a simple dimension reducer. It carries out linear
transformations on the entire data set to attain vectors (identified as eigenvalues), then
identifies incremental changes in the data. It enables you to detect changes in features
in the data. This method helps determine the ratio of each feature so you can decide
which features to preserve for modeling. It is suitable when your data contains countless
features, because it helps you identify features that are imperative to the model. It is
important to note that, although I call it a “method” in this chapter, that is not strictly the
case. Compared to other methods, it does not consider any plausible hypotheses; rather,
you execute it prior to executing linear, nonlinear, and cluster methods, among others.
This chapter uses the data from the previous chapter, which covered the k-means
method.
101
Chapter 9 Principal Component Analysis with Scikit-Learn, PySpark, and H2O
import pandas as pd
Listing 9-2 drops any unnecessary features in the data.
Listing 9-2. Drop Unnecessary Features

S
Listing 9-3 scales the whole data set with a standard scaler.
Listing 9-3. Scale the Whole Data Set

sk_standard_scaled_data = sk_standard_scaler.fit_transform(initial_data))
Listing 9-4 executes the principal component method with the Scikit-Learn
framework. It then discloses the variance that each component clarifies (see Figure 9-1).
Listing 9-4. Arrange the Explained Variance from the Scikit-Learn Principal
Components Method
%matplotlib inline
from sklearn.decomposition import PCA
sk_principal_component_method = PCA(n_components=3)
sk_principal_component_method.fit_transform(sk_standard_scaled_data)
sk_principal_component_method_variance = sk_principal_component_method.
explained_variance_
plt.bar(range(3), sk_principal_component_method_variance, label="Variance")
plt.legend()
102
plt.ylabel("Variance ratio")
plt.xlabel("Components")
plt.show()
Figure 9-1. Variance of the Scikit-Learn principal components method
Figure 9-1 shows all that components disclose over 0.5 of variance. Therefore, all the
principal component methods will have three components.
Listing 9-5 displays the variance from the Scikit-Learn principal component method.
Listing 9-5. Print the Variance that the Scikit-Learn Principal Component
Method Discloses
print("SciKit-Learn PCA explained variance ratio", sk_principal_component_
method_variance)
SciKit-Learn PCA explained variance ratio [1.33465831 1.00427272

0.67614435]
103
Listing 9-6 executes the principal component method with the Scikit-Learn
framework and shows fewer dimensions (see Figure 9-2).
Listing 9-6. Execute the Principal Component Method with the Scikit-Learn
Framework
sk_principal_component_method_ = PCA(n_components=3)
sk_principal_component_method_.fit(sk_standard_scaled_data)
sk_principal_component_method_reduced_data = sk_principal_component_
method_.transform(sk_standard_scaled_data)
plt.scatter(sk_principal_component_method_reduced_data[:,0], sk_
principal_component_method_reduced_data[:,2], c=initial_data.iloc[::, -1],
cmap="coolwarm")
plt.xlabel("y")
plt.show()
Figure 9-2. Dimensions from the Scikit-Learn principal component method

104
Figure 9-2 shows the dimensions found by the Scikit-Learn principal component
method.
P
ySpark in Action
This section executes the principal component method using the PySpark framework.


pyspark_configuration = SparkConf().setAppName("pyspark_principal_
components_analysis").setMaster("local")

Listing 9-11 creates a list for the independent features and a string for the dependent
105

Listing 9-12 scales the whole data set with the Lifeline PySpark framework.
Listing 9-12. Scale the Whole Data Set with the Lifeline PySpark Framework
from pyspark.ml.feature import StandardScaler

pyspark_standard_scaler = StandardScaler(inputCol = pyspark_data.
columns[-1], outputCol = "pyspark_scaled_features", withMean = True,
withStd = True).fit(pyspark_data)
pyspark_data_scaled_data = pyspark_standard_scaler.transform(pyspark_data)
Listing 9-13 executes the PySpark principal component method.
Listing 9-13. Execute the PySpark Principal Component Method
from pyspark.ml.feature import PCA

number_of_k = 3
pyspark_principal_components_method = PCA(k = number_of_k, inputCol =
pyspark_data_scaled_data.columns[-1], outputCol = "pca_reduced_features").
fit(pyspark_data_scaled_data)
pyspark_principal_components_method_data = pyspark_principal_components_
method.transform(pyspark_data_scaled_data)
Listing 9-14 discloses the variance that each component clarifies (see Figure 9-3).
Listing 9-14. Arrange the PySpark Principal Components Method’s Explained

Variance
pyspark_principal_component_method_variance = pyspark_principal_components_
method.explainedVariance.toArray()
plt.bar(range(3), pyspark_principal_component_method_variance,
label="Variance")
106
plt.legend()
plt.ylabel("Variance ratio")
plt.xlabel("Components")
plt.show()
Figure 9-3. Explained variance from the PySpark principal components method
Listing 9-15 prints the variance that the PySpark principal component method
reveals.
Listing 9-15. Print the Variance that the PySpark Principal Component Method
Reveals
print('Explained variance ratio', pyspark_principal_components_method.
explainedVariance.toArray())
PySpark PCA variance ratio [0.44266167 0.33308378 0.22425454]
107
The results show that the first component discloses 0.44 of the related changes in
features in the data, the second component discloses 0.33, and the third component
discloses 0.22.
Listing 9-16 reduces the data using the PySpark principal component method.
Listing 9-16. Reduce the Data with the PySpark Principal Component Method
pyspark_principal_components_method_reduced_data = pyspark_principal_
components_method_data.rdd.map(lambda row: row.pyspark_scaled_features).
collect()
pyspark_principal_components_method_reduced_data = np.array(pyspark_
principal_components_method_reduced_data)
Listing 9-17 discloses the data that the PySpark principal component method
reduced (see Figure 9-4).
Listing 9-17. Disclose the Data that the PySpark Principal Component Method
Reduced
plt.scatter(pyspark_principal_components_method_reduced_data[:,0], pyspark_
principal_components_method_reduced_data[:,1], c=initial_data.iloc[::, -1],
cmap = "coolwarm")
plt.xlabel("y")
plt.show()
108
Figure 9-4. Dimensions from the PySpark principal component method
Figure 9-4 shows the dimensions found by the PySpark principal component
method.
H
2O in Action
This section executes and assesses the principal component method using the H2O
framework.

109
h2o_training_data, h2o_validation_data = h2o_data.split_frame(ratios=[.8])
Listing 9-21 executes the H2O principal component method.
Listing 9-21. Execute the H2O Principal Component Method
from h2o.estimators import H2OPrincipalComponentAnalysisEstimator

h2o_principal_components_method = H2OPrincipalComponentAnalysisEstimator
(k = 3, transform = "standardize")
h2o_principal_components_method.train(training_frame = h2o_training_data)
Listing 9-22 reduces the data with the H2O principal component method.
Listing 9-22. Reduce the Data with the H2O Principal Component Method
h2o_yhat = h2o_principal_components_method.predict(h2o_validation_data)
C
onclusion
and H2O) in order to condense data into a few dimensions by employing the principal
component method. As mentioned at the beginning of this chapter, the chapter did
not execute the method to test plausible hypotheses, thus there are no key metrics for
objectively assessing the method.
110
CHAPTER 10
Automating the Machine

Learning Process
with H2O
This is a short yet insightful chapter that concludes the book by explaining a
straightforward approach to automating machine learning processes with the help of a
widespread machine learning framework, known as H2O.
Exploring Automated Machine Learning

In a world saturated by innumerable machine learning frameworks, you might think that
organizations would easily adopt machine learning. Unfortunately, that is not the case.
There is a massive shortage of skills because of the considerable complexity of machine
algorithms.
To combat these challenges, H2O introduced AutoML, which attempts to automate
the machine learning process by simultaneously training and handpicking methods with
extraordinary performance.
Although the H2O AutoML function is beneficial, it does not include methods like
cluster analysis, survival analysis, and time series analysis, among others.
This chapter uses the data from Chapter 3. It also distinguishes the performance of
the leading AutoML method against the ones executed in Chapter 3.
111
Chapter 10 Automating the Machine Learning Process with H2O
P
This chapter manipulates the data attained in Chapter 3, so it does not describe the
preprocessing tasks in detail. Listing 10-1 executes all the preprocessing tasks.
import pandas as pd
df = pd.read_csv(r"filepath\WA_Fn-UseC_-Marketing_Customer_Value_Analysis.csv")
H2O AutoML in Action

This section uses the H2OAutoML() method to train multiple methods and rank their
performance. This section executes and assesses H2OAutoML.
112

Listing 10-3 changes the pandas dataframe to the H2O dataframe.
Listing 10-3. Change the Pandas Dataframe to H2O Dataframe
int_x = initial_data.iloc[::,0:19]
fin_x = initial_data.iloc[::,19:21]
x_combined = pd.concat([int_x, fin_x], axis=1)
x_list = list(x_combined.columns)
y_list = initial_data.columns[19]
y = y_list
x.remove(y_list)

The H2OAutoML() method enables us to stipulate the considerable time needed
to train each method. Specify max_runtime_secs to do so.
Listing 10-6 executes H2O’s auto machine learning method.
Listing 10-6. Execute H2O’s Auto Machine Learning
from h2o.automl import H2OAutoML

h2o_automatic_ml = H2OAutoML(max_runtime_secs = 240)
h2o_automatic_ml.train(x= x, y= y, training_frame = h2o_training_data,
validation_frame = h2o_validation_data)
113
Listing 10-7 ranks H2O’s methods in ascending order (see Table 10-1).
Listing 10-7. Rank H2O’s Methods
h2o_method_ranking = h2o_automatic_ml.leaderboard
print(h2o_method_ranking)
Table 10-1. H2O Method Ranking

Model_ID Mean_residual_ Rmse Mse Mae Rmsle
deviance
StackedEnsemble_AllModels_ 18371.3 135.541 18371.3 86.6257 0.482043

AutoML_20210824_115721
StackedEnsemble_ 18566.8 136.26 18566.8 87.7959 0.477828
BestOfFamily_AutoML_
20210824_115721
GBM_1_AutoML_20210824_ 19392.8 139.258 19392.8 90.0652 0.490441
115721
GBM_3_AutoML_20210824_ 19397.8 139.276 19397.8 90.3467 0.490855
115721
GBM_2_AutoML_20210824_ 19467.3 139.525 19467.3 89.9908 0.491887
115721
GBM_grid__1_AutoML_ 19607.9 140.028 19607.9 91.3311 0.495784
20210824_115721_model_4
GBM_4_AutoML_20210824_ 19633.4 140.119 19633.4 89.9553 0.499693
115721
XRT_1_AutoML_20210824_ 19717.6 140.42 19717.6 89.7474 0.485144
115721
GBM_grid__1_AutoML_ 19805.8 140.733 19805.8 91.7092 0.494005
20210824_115721_model_2
GBM_grid__1_AutoML_ 19858.6 140.921 19858.6 90.0839 0.521174
20210824_115721_model_9
114
Table 10-1 suggests that StackedEnsemble_AllModels_AutoML_20210824_115721 is

the highest-ranking method with the smallest root mean residual deviance (at 18371.3).
The root-mean-square error is 135.541.
In contrast, the lowest ranking method is GBM_grid__1_AutoML_20210824_11572
1_model_9, with a mean residual deviance at 19858.6 and a root-mean-square error
at 140.921. Listing 10-8 assesses the highest ranking method (StackedEnsemble_
AllModels_AutoML_20210824_115721).
Listing 10-8. Execute H2O Auto Machine Learning
highest_ranking_method = h2o_automatic_ml.leader
print(highest_ranking_method)
ModelMetricsRegressionGLM: stackedensemble
MSE: 4900.131929292816
RMSE: 70.00094234574857
MAE: 45.791336342731775
RMSLE: 0.3107457400431741
R^2: 0.9421217261180531
Null deviance: 617106545.1167994
AIC: 82648.0458245655
** Reported on validation data. **
MSE: 17967.946605287365
RMSE: 134.04456947331872
MAE: 83.66742154264892
RMSLE: 0.43883929226519075
R^2: 0.7718029372846945
115

Null deviance: 73564698.35450743
AIC: 11790.460469478105
** Reported on cross-validation data. **
MSE: 18371.339394332357
RMSE: 135.5409140973026
MAE: 86.62574587341507
RMSLE: 0.4820432203973515
R^2: 0.7830055540572105
Null deviance: 617269595.4214188
AIC: 92282.67565857261
The results suggest that the StackedEnsemble_AllModels_AutoML_20210824_115721

method explained 77% of the related changes in the test data and 78% of the related
changes in the validation.
C
onclusion
The highest-ranking H2O AutoML method (StackedEnsemble_AllModels_
AutoML_20210824_115721) pales in comparison to the ordinary least-square methods
executed in the chapter. The H20 method disclosed 77% of the related changes in the test
data. In contrast, all methods executed in Chapter 3 (with the Scikit-Learn and PySpark
frameworks) explained 100% of the related changes in the test data. Unfortunately,
the StackedEnsemble_AllModels_AutoML_20210824_115721 method does not have a
summary function. This concludes the book.
116
Index
A Deep learning (DL), 4, 7, 75
Dimension reducers, 4
Accelerated failure time method, 34
DL frameworks, Keras, 14
Apache Spark, 7, 10
AutoML method, 111
H2O, 112, 113, 115, 116 E
preprocessing tasks, 112
Elbow curve, 89, 91
Ensemble methods, 3
B
Big data, 7
F
business, 8
customer relationships, 8 findspark framework, 69
decision making, 9
ETL, 10
G
features, 8
product development, 9 Gaussian distribution, 1
warehousing, 9 Gradient boosting methods, 66
H2O framework, 71, 72
PySpark, 69, 71
C XGBoost, 66–68
Categories, 2 GraphX, 12
Cluster centers, 92
Cluster methods, 3, 4
Cox Proportional Hazards method, 29, 30
H, I, J
H2OAutoML, 112
H2O framework, 22, 23, 25–27, 52, 71
D H2O logistic regression
Data-driven organizations, 9 method, 54, 55, 57
Decision tree, 59, 60, 62 Hadoop Distributed File System (HDF), 11
Deep belief networks, 87 Hadoop File System (HDFS), 10
117
T. C. Nokeri, Data Science Solutions with Python, https://doi.org/10.1007/978-1-4842-7762-1
Index
K Principal component method, 101

H2O framework, 109
Keras, 14
PySpark framework, 105–107
k-means method, 89, 91
scales, 102
H2O framework, 97, 98
Scikit-Learn, 103
PySpark framework, 93–95, 97
PySpark framework, 20, 21, 34, 48, 69,
105–107
L PySpark logistic regression
Lifeline Cox Proportional Hazards method, 49–51
method, 30–33 PySpark principal component
Linear regression method, 1 method, 108
Logistic regression method, 39, 42
R
M Residuals, 15
Machine learning (ML), 1, 7, 111 Resilient Distributed Data
H2O, 13 Sets (RDDs), 11
Scikit-Learn framework, 13
XGBoost, 14 S
Mantel-Haenszel method, 29
Scikit-Learn decision tree
Microsoft CSV file, 15
method, 64–66
Multiclass classification methods, 2
Scikit-Learn framework, 13, 17, 19, 41,
Multilayer perceptron neural
61, 62
network, 75, 76
Scikit-Learn logistic regression method,
42, 44–47
N Scikit-Learn multilayer perceptron neural
Neural networks, 4 network, 77, 79, 80
deep belief networks, 87
H2O, 87
O keras, 82–86
Ordinary least-squares method, 1, 15 Spark configuration, 11
Spark frameworks, 11
Spark MLlib, 12
P, Q Spark streaming, 12
Preprocess features, 60 Supervised learning, 1
Preprocessing tasks, 76 Survival methods, 29
118
Index
T V, W
TensorFlow, 14 VectorAssembler() method, 35
U X, Y, Z
Unsupervised learning, 3 XGBoost, 14
119

Data Science Solutions With Python

Uploaded by

Copyright:

Available Formats

Data Science Solutions With Python

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Data Science Solutions With Python

Uploaded by

Copyright:

Available Formats

Data Science

Tshepo Chris Nokeri

Tshepo Chris Nokeri

ISBN-13 (pbk): 978-1-4842-7761-4 ISBN-13 (electronic): 978-1-4842-7762-1

Copyright © 2022 by Tshepo Chris Nokeri

About the Technical Reviewer��������������������������������������������������������������������������������� xi

Chapter 1: Exploring Machine Learning������������������������������������������������������������������� 1

Chapter 2: Big Data, Machine Learning, and Deep Learning Frameworks��������������� 7

Chapter 3: Linear Modeling with Scikit-Learn, PySpark, and H2O������������������������� 15

Chapter 4: Survival Analysis with PySpark and Lifelines��������������������������������������� 29

Chapter 5: Nonlinear Modeling With Scikit-Learn, PySpark, and H2O�������������������� 39

Chapter 6: Tree Modeling and Gradient Boosting with Scikit-Learn, XGBoost,

Chapter 7: Neural Networks with Scikit-Learn, Keras, and H2O����������������������������� 75

Chapter 8: Cluster Analysis with Scikit-Learn, PySpark, and H2O������������������������� 89

Chapter 9: Principal Component Analysis with Scikit-­Learn,

• Pandas for data structures and tools.

• PySpark for in-memory, cluster computing.

• Auto-Sklearn, Tree-based Pipeline Optimization Tool (TPOT),

• Scikit-Learn for building and validating key machine learning

• Keras for high-level frameworks for deep learning.

• H2O for driverless machine learning.

• Lifelines for survival analysis.

• NumPy for arrays and matrices.

• SciPy for integrals, solving differential equations, and optimization.

• Matplotlib and Seaborn for recognized plots and graphs.

Exploring Supervised Methods

normal distribution; the method struggles to explain the variability, resulting in an

Table 1-1. Types of Parametric Methods

Exploring Nonlinear Models

Table 1-2. Varying Nonlinear Models

Exploring Ensemble Methods

Exploring Unsupervised Methods

Exploring Cluster Methods

Table 1-3. Varying Cluster Models

Exploring Dimension Reduction

Table 1-4. Main Dimension Reducers

Exploring Deep Learning

Table 1-5. Varying Neural Networks

Big Data, Machine

Big Data Features

Table 2-1. Big Data Features

Impact of Big Data on Business and People

Better Customer Relationships

Refined Product Development

Big Data Warehousing

Big Data ETL

Table 2-2. ETL Activities

Extract Involves getting data from some database.

Big Data Frameworks

Resilient Distributed Data Sets

Listing 2-1. Prepare the PySpark Framework

import findspark as initiate_pyspark

Listing 2-2. Stipulate the PySpark App

from pyspark import SparkConf

Listing 2-3. Prepare the Spark Session

from pyspark.sql import SparkSession

Listing 2-4. Initializing the H2O Framework

Exploring the Ordinary Least-Squares Method

Listing 3-1. Attain the Data

Listing 3-2. Drop Unnecessary Features in the Data

About the Technical Reviewer�� xi

Chapter 1: Exploring Machine Learning�� 1

Chapter 2: Big Data, Machine Learning, and Deep Learning Frameworks�� 7

Chapter 3: Linear Modeling with Scikit-Learn, PySpark, and H2O�� 15

Chapter 4: Survival Analysis with PySpark and Lifelines�� 29

Chapter 5: Nonlinear Modeling With Scikit-Learn, PySpark, and H2O�� 39

Chapter 6: Tree Modeling and Gradient Boosting with Scikit-Learn, XGBoost,

Chapter 7: Neural Networks with Scikit-Learn, Keras, and H2O�� 75

Chapter 8: Cluster Analysis with Scikit-Learn, PySpark, and H2O�� 89

Chapter 9: Principal Component Analysis with Scikit-Learn,

Exploring Supervised Methods

Exploring Nonlinear Models

Exploring Ensemble Methods

Exploring Unsupervised Methods

Exploring Cluster Methods

Exploring Dimension Reduction

Exploring Deep Learning

Big Data Features

Impact of Big Data on Business and People

Better Customer Relationships

Refined Product Development

Big Data Warehousing

Big Data ETL

Big Data Frameworks

Resilient Distributed Data Sets

Exploring the Ordinary Least-Squares Method