Optimizing Offline Access to Social Network

Content on Mobile Devices

Ngoc Do1 , Ye Zhao1 , Shu-Ting Wang2 , Cheng-Hsin Hsu2 , and Nalini Venkatasubramanian1
1

Department of Information and Computer Science, University of California, Irvine, USA

2
Department of Computer Science, National Tsing Hua University, Hsin-Chu, Taiwan

AbstractIn this paper, we explore the problem of supporting

efficient access to social media contents on social network sites
for mobile devices without requiring mobile users to be online
all the time. We propose and implement a broker/proxy based
architecture that stages data at a broker/proxy, and selectively
downloads to the mobile device only those contents that have
a high likelihood of being viewed. The system determines the
relevance of social media updates that continuously arrive (e.g.,
Facebook friend updates) for each user. Using knowledge of
this relevance and current network/system conditions, we develop
scheduling algorithms that determine which social contents are
sent to the devices. We develop an Android app providing offline
access to Facebook. Our experimental results indicate that our
system is energy efficient, which saves energy by 6.9 times for
WiFi and 9.1 times for cellular connections. We also use data traces
gathered from our app to further drive extensive simulation based
evaluations which show that our proposed algorithms provide
efficient facilities for tuning the systems performance.1
Index TermsSocial networks, mobile devices, optimization,
energy consumption, offline access.

I. I NTRODUCTION
The phenomenal popularity of social networks, such as
Facebook, Twitter, LinkedIn, Google+, and Instagram, has
changed the way people interact today. Indeed, many people
rely on these social networks to communicate with their friends,
family, and community on a day to day basis. The ability to
continue these interactions anytime, anywhere seamlessly is
quickly becoming commonplace, and users on modern mobile
devices expect to not just to access social networks but also
exchange rich media contents, such as video, audio, and images,
for an enhanced user experience. It is reported that 93% of
Android smartphone users in India use social networks on their
smartphones [1] and often this is the reason why they purchase
smartphones in the first place. In North America, a recent IDC
report on smartphone users indicates that 70% of them access
Facebook via smartphones, and more strikingly, 40% of users
feel connected when using Facebook, only trailing 43% for
making voice calls and 49% for texting [2]. In fact, the main
finding of the IDC report is mobile+social=connectedness,
i.e., people feel isolated without mobile access to social networks.
To ensure this constant connectedness, mobile users subscribe (and pay for) 3G/4G data plans that are often expensive
and do not work for a host of reasons including: (a) wireless
network availability is sporadic (accessibility of WiFi access
points, unpredictable data rates in 3G networks), (b) mobile
1 This research has been supported in part by NSF grants 1063596, 1059436,
1352727 and 1143705.

devices are battery-powered with stringent energy budgets that

are easily depleted from constant connectivity to and interaction
with WiFi/3G networks, (c) the shared network bandwidth is
limited in public locations (where users want to access this
information), and (d) dataplans are becoming volume-driven
and hence costly.
We observe that the basic need (root cause) of alwayson connectedness stems from the assumption that existing
mobile social mobile apps, such as mobile Facebook app,
expect always-on connectivity. The modus operandi today is
that these apps synchronize with social networks when mobile
users launch the apps and mobile devices are connected to
the Internet. We believe that offering offline access for these
apps is necessary for mobile users in order to interact with
social contents when Internet access is not available. There are
many scenarios in which the offline access feature is useful.
For example, a student takes a subway to his or her school. In
the subway, the student does not typically have mobile Internet
connections to social network sites, and thus the offline access
feature is the only option allowing the student to view and
interact with updates from his or her friends that are prefetched
before the student steps into the subway.
The problem of optimally prefetching social media content
from social media providers to mobile users is more complex
than it seems at first glance. Multiple factors contribute to
the non-uniform nature of the problem at mobile users, contents, and system levels. More specifically, mobile users have
different/personalized viewing needs and preferences; social
media contents are diverse in size and importance; and network
conditions may be highly dynamic due to fading, shadowing,
interference, and congestion on wireless/wired links. All of the
above factors make prefetching social media contents on mobile
devices quite challenging.
In this paper, we envision a distributed broker/proxy solution,
in which broker/proxy nodes in the network are employed to
help mobile devices to optimally prefetch social media contents.
Fig. 1 illustrates the considered broker/proxy architecture where
mobile users and social media providers exchange data by
staging relevant content intelligently with the supports from the
broker/proxy nodes. The goal of the broker/proxy architecture is
to maximize the viewing likelihood and quality of experience of
the prefetched social media contents while maintaining certain
energy level at mobile devices for user daily activities. The
main contributions of this paper are:
1) We design a broker/proxy based architecture and framework to prefetch media contents from online social networks and deliver the contents efficiently to mobile users

Figure 1.

Considered system architecture.

through the broker/proxy nodes (Section II).

2) We develop fine-grained control mechanisms for optimizing data transfer to mobile devices based on the Lyapunov
optimization framework that maximizes the probability of
content viewing while maintaining a guaranteed residual
energy level at mobile devices. We design and develop
an optimal scheduling technique and a near-optimal technique (that achieves an approximation factor of 2) using
the above framework to deliver media contents to mobile
users - our scheme accounts for the relevance of contents
to the user and is aware of energy consumption at mobile
devices (Section III).
3) We evaluate our system using a testbed of Android
phones. We compare our broker/proxy based system
against an offline pull based system that executes completely at mobile devices. Our results illustrate the broker/proxy based system saves energy 6.9 times for WiFi
connections and 9.1 times for cellular connections. We
also collect data traces from our users for simulation
based evaluations. The results from the simulations confirm the correctness and efficiency of our proposed algorithms (Section IV).
II. A B ROKER /P ROXY F RAMEWORK FOR S OCIAL C ONTENT
D ELIVERY
A straightforward approach to enabling offline social network
access is to program mobile devices to periodically check the
social networks for new updates and download relevant updates.
This approach, however, is expensive both in terms of the
consumed data plan quota and mobile device energy, which
is caused primarily by the limitations in the current APIs of
social network providers. In order to determine whether there
are new updates2 , mobile device has to set up connections, and
send requests for new updates. If there are no new updates,
the requests are useless, and device resources (i.e., energy)
consumed for sending the requests are wasteful. In some cases,
mobile devices even have to pull/retrieve multiple data items
(e.g., posts, comments/likes of Facebook posts), and analyze the
data to detect new updates. Executing these processes on mobile
devices periodically is taxing on resource-constrained mobile
devices. A better approach is to employ a system of brokers
2 Social network sites, such as Facebook, provide limited notification service,
that is only for updates related directly to users, but not for updates from friends,
e.g., items in Facebook News Feed page.

Figure 2.

Major software modules in our system.

and proxies in the network, in order to download, analyze, and

deliver new updates from social network sites to mobile users.
Our proposed approach consists of a system of brokers/proxies that analyzes new updates and downloads the
most relevant content from social network sites to mobile
users. In this system, the key coordination is performed at a
broker social media users on mobile devices register with
the system through this brokerage service. The broker service
periodically checks and tracks social media updates for each
mobile user, calculates the relevance of every update to the
user and generates an annotated list of updates, enhanced with
the calculated relevance factors. A broker may manage one or
more proxies, and each proxy in turn serves one or multiple
mobile devices. The broker passes the annotated list of new
contents to each mobile devices proxy and also initiates the
actual downloads of the updated contents onto the respective
network proxies (usually, in close vicinity of the client). In
general, proxies may be installed by mobile users, cellular
carriers, content distribution networks, social networks, and
third-party companies.
In this paper, we assume that proxy selection is done by
individual mobile users, who may have lists of trustworthy
proxies in mind or prefer to use closer proxies for higher
network bandwidth. More sophisticated policies for proxy selection are possible [3] - for example multiple brokers/proxies
may be employed for scalability and mobility handling - the
design of such policies are out of the scope of this paper. We
aim to design, implement, and evaluate a broker/proxy solution
that schedules mobile prefetches to achieve these goals and
relieve the mobile device from the associated computation and
communication overhead.
Fig. 2 illustrates the software modules of the proposed
system. The broker/proxy periodically checks for updates from
social network sites and efficiently delivers the updates to
mobile users. The crux of the system resides in two key
modules that are described below: (i) user-content relevance
evaluator, which estimates the likelihood of a content being
watched by a user (based on social connections) and (ii)
scheduling algorithm, which performs the actual delivery of
contents to mobile users such that the benefits of prefetching
contents is maximized under the resource constraints.
User-content relevance evaluator. We present only briefly
here how we evaluate user-content relevance using a learning
technique. The reader can find more details in our previous

work [4]. This module estimates the likelihood a social content

will be viewed by a user using social network information. The
resulting user-content relevance provides hints to the scheduling and delivery modules in the proxy for resource-efficient
delivery. Existing approaches to calculate content relevance in
social networks use: (1) content based approaches [5][7] that
exploit correlation among content attributes such as keywords,
(2) collaborative filtering approaches [8][10] that capture user
similarity and use this to determine the user interest in a
content based on similar users interests on the same content,
and (3) social network based approaches [11][13] that exploit
social relations to determine whether a user is likely to view a
content. We follow the approaches [11][13] in the last category
to rank social media contents based on social relations and
interactions. However, unlike prior techniques, our relevance
evaluation scheme is designed to have low complexity and be
scalable to frequent arrival of new contents.
There are two steps in our relevance evaluation scheme: (a)
social-based feature construction and (b) viewing likelihood
prediction. In the first step, we construct social features and
interactions to capture user interests on a content. Four features
of mobile social networks are constructed:
1) Post interactions, which capture the interactions between
a user and his/her friends via posts, comments, and likes.
2) Private message exchanges, which capture the private
message interactions among users.
3) User interests, which capture the number of view clicks
of a user on contents posted by another user.
4) Post popularity, which capture the number of comments
and likes of a post for the global popularity of a post.
The user-content relevance evaluator collects all the features
and applies a training process in the next step.
The second step employs a logistic regression classifier to
build a prediction model and evaluate the viewing likelihood
for a newly arriving media content and user pair. In particular,
the logistic regression classifier evaluates and calculates proper
coefficients for the impact of all four features. The learning
model is then used to evaluate the viewing likelihood for a
newly arriving media content. Our experiments show that the
model parameters are calculated in a lightweight and quick
manner. The maximum running time to build the model in our
experiments is only a few hundreds of milliseconds. Therefore,
the user-content relevance evaluator can frequently update the
model to reflect current behaviors of the user.
Scheduling algorithm. The second key component is the
Scheduling Algorithm module that executes on the proxy. The
challenge here is to utilize output (relevance) from the relevance
evaluator module to efficiently schedule the delivery of contents
from the proxy to the mobile device based on the available
resources at the mobile device, e.g., network bandwidth and
energy level. Updates from social network sites for a user are
downloaded to the selected proxy via the Download and Storage module. Decisions are made by the scheduling algorithm
running on the proxy to perform the prefetch of relevant updates
under good network conditions and battery levels at the mobile
device. These system conditions are captured via a Device
Monitor at the proxy and the Delivery module implements the

content transfer. The details of the scheduling algorithm are

presented in the next section.
III. C ONTENT S CHEDULING A LGORITHM
In this section, we formulate and solve the scheduling problem of prefetching social contents to mobile device optimally.
A. System Models and Problem Formulation
We divide time into slots. Every time slot, the scheduling
algorithm selects which contents will be delivered to mobile
device.
Content arrival. A new content c generated by a friend of u
is fetched by the proxy and stored in the proxys database before
it can be delivered to the user. Content c includes metadata
(e.g., post id, author id, description, source links of images
and videos, and etc.) and media content (images and videos).
Let us denote the size of content c (including metadata and
media content) as sc , and the total size of all contents for the
user fetched by the proxy in time slot t as (t). We have (t) =
P
ca(t) sc , where a(t) is a set of contents arriving to the proxy
in t. Assume that (t) is finite and limited by max , i.e., (t)
max t. The time averaged fetched data amount is defined as:
t1
X
, lim 1

E{( )}.
(1)
t t
=0
Content departure and dropping. Every time slot, the
scheduling algorithm chooses a subset of the fetched contents
to deliver to the user. Let r(t) be a set of contents scheduled to
be delivered in time slot
P t, and (t) be the total delivered data
amount, i.e., (t) = cr(t) sc . We assume (t) max t.
It is not hard to see that (t) is limited by the mobile devices
downlink bandwidth at t, defined as (t), i.e., (t) (t). We
also assume (t) max t. The time averaged data amount
delivered to the user is defined as:
t1
1X
E{( )}.
(2)

, lim
t t
=0
Some fetched contents are dropped by the proxy due to low
viewing probability or large size (hence, high energy consumption). Let d(t) and (t) be a set of dropped contentsP
and the total dropped data amount in time slot t, i.e., (t) = cd(t) sc .
The time averaged data amount dropped from the fetched
content list is:
t1
1X

, lim
E{( )}.
(3)
t t
=0
Energy budget and consumption. In our system, each user
sets an energy threshold et and nothing is prefetched if the
current energy level, denoted as e(t) is less than et . Moreover,
we allow mobile users to set energy budget for prefetching
contents, in order to preserve battery levels for other daily
activities, such as making phone calls and browsing Web pages.
The energy budget is set for user-specified time periods, e.g.,
a user may set the energy budget to 10% of the current energy
for a period between 8 a.m. to 6 p.m., and to 30% between 6
p.m. to 8 a.m. In particular, we let ts and te be the starting and
ending times of a period. Assume period [ts , tf ] consists of a
set of time slots {t0 , t1 ,...}. Let us define ea as the available
energy amount at the device at ts , and eu as the usable energy
percentage for the system to download social contents in the

period. The energy budget at the beginning of the period is

ep = [(ea et )eu ]+ , where [x]+ , max(0, x). For every
content download, the consumed energy is deducted from the
energy budget. The energy budget may increase if the phone is
charged. In time slot t, if the phone is charged and the current
energy level is larger than et , eu percentage of the charged
energy will be added to the energy budget. Let w(t) be the
charged energy in time slot t. The energy added to the budget
is (t) = w(t)eu . We write (t0 ) = ep + w(t0 )eu for t0 , and
define the time averaged energy added to the budget as:
t1
1X
, lim
E{( )}.
(4)
t t
=0
We next model the energy consumed by mobile device for
receiving contents from the broker/proxy. Let c (t) be the
energy amount used for receiving content c in time slot t. We
employ the energy models from [14] to estimate c (t), based
on the content, network type, and network condition.
P The total
energy consumption in time slot t is (t) =
cr(t) c (t),
where r(t) is the contents delivered to the mobile device in t.
The time averaged energy consumption amount is written as:
t1
1X
E{( )}.
(5)
, lim
t t
=0
Utility model. Let pc be the probability a user u will view
c, which is between 0 and 1. The closer to 1 pc is, the more
likely u views c. We consider time decay of contents, i.e., a
more recent content interests users more. Let Uc (t) = f (pc , t)
be the time decay utility function that models the benefit to
deliver content c to the user in time slot t, which is a decreasing
function on pc and t. In a time slot, some contents are delivered
to the user while some are dropped, or the proxy may be totally
filled up. Lets denote Ur (t) and Ud (t) as the total utilities of
contents which are P
delivered and dropped, respectively.
MatheP
matically, Ur (t) = cr(t) Uc (t) and Ud (t) = cd(t) Uc (t).
Problem formulation. The scheduling algorithm selects a
set of contents r(t) to deliver to the mobile device and a set of
contents d(t) to drop from the proxy based on the following
formulation.

t1 
1X

max: U = lim
E Ur (t) Ud (t) ;
(6a)
t t
=0

st:
+ ;

(6b)
;
X

(6c)
sc (t);

(6d)

c (t) min(e(t) et , P (t)).

(6e)

cr(t);r(t)f (t)

X
cr(t);r(t)f (t)

The objective function in Eq. (6a) maximizes the long term

utility for content delivery and minimizes the long term utility
for content dropping. The proxy drops some contents to stabilize the system as shown in constraint in Eq. (6b). Weight
( 1) is added to the dropping utility to force the system
to consider more seriously when making a dropping decision.
Constraint Eq. (6c) ensures that the long term energy budget
is enough to perform scheduled transfers. Constraint Eq. (6d)
ensures that bandwidth is enough for data transfer at any time
slot. And the last constraint in Eq. (6e) makes sure that energy

is enough for data receipt from the broker/proxy (note that,

e(t) et is the difference of the current energy level with the
power threshold, and P (t) is the remaining amount in energy
budget that will be modeled in the next part).
B. An Optimal Algorithm
We design an algorithm to solve the problem in Eqs. (6a)
(6e) using the fine-grained queue control theory Lyapunov
optimization [15], which is useful for solving long-term maximization problems. We start by presenting our design principles, and then the detailed algorithm.
Algorithm design principles. We use a real queue Q and a
virtual queue P to represent constraints in Eqs. (6b) and (6c)
for data and energy, respectively. Q stores incoming contents
downloaded to the proxy. The contents in Q are taken out of
the queue for either delivery or drop. The evolution of Q over
time is written as:
Q(t + 1) , [Q(t) (t) (t) + (t)]+ .
(7)
We define the stability of the Q as:
t1
X
, lim sup 1
Q
E{Q( )} < .
(8)
t
t =0
At t = 0, we set Q(t) = 0. The constraint in Eq. (6b) is
satisfied, as long as Q is stable.
Virtual queue P is just a counter in memory, which indicates
the energy budget for the mobile device to prefetch contents.
The evolution of P over time is:
P (t + 1) , [P (t) (t) + (t)]+ .
(9)
At t = 0, we set P (t) = 0. In Section III-C, we show that P (t)
is deterministically upper bounded. This therefore satisfies the
constraint in Eq. (6c).
We define vector (t) = [Q(t); P (t)] and a quadratic
Lyapunov function:


1
2
2
L((t)) ,
Q(t) + (P (t) ) ,
(10)
2
where is a constant. Intuitively, if the system is stabilized, i.e.,
L((t)) is maintained small, P (t) is accumulated close to .
Therefore, by choosing a reasonable , we control the available
energy amount for the mobile device to prefetch contents. A
one-step Lyapunov drift is defined as below:
((t)) , E{L((t + 1)) L((t))|(t)}.
(11)
The intuition is if ((t)) is maintained negatively over time,
the system is stable.
Lemma 1 (Lyapunov drift): The one-step conditional Lyapunov drift satisfies the following constraint at any time slot
under any algorithms used to control the system:
4((t)) = E{L((t + 1)) L((t))|(t)}
E{Q(t)(t) + (P (t) )(t)|(t)}
E{Q(t)(t)|(t)}
(12)
+ E{Q(t)(t)|(t)}
+ E{(P (t) )(t)|(t)},
where is a carefully chosen constant.
Our goals are to minimize the drift and maximize the
systems utility. Thereby, we obtain:
min : 4((t)) V E{Ur (t) Ud (t)|(t)},
(13)
where V is a weight that controls how important the utility is,
compared to the system stability.

By applying Lemma 1 to Eq. (13), we maximize the sum

(t) = 1 (t) + 2 (t), where:
1 (t) = E{V Ur (t) + Q(t)(t) + (P (t) )(t)|(t)}
+ E{Q(t)(t) V Ud (t)|(t)};
(14)
2 (t) = E{( P (t))(t)|(t)}.
The proposed algorithm. Our proposed algorithm is to minimize Eq. (13) by maximizing 1 (t) and 2 (t). The algorithm
consists of two procedures: (1) Content Selection to maximize
1 , and (2) Energy Control to maximize 2 . Algorithm 1
presents the algorithm in details.
Algorithm 1 Social Content Scheduling Algorithm
1) Content Selection. In time slot t, from a set of social
contents f (t) fetched to the proxy so far, the proxy selects
a set r(t) to deliver to the mobile user and a set d(t) to
drop such that
Xthe following objective
X is satisfied:
r1 (c, t) +

max:

cr(t);r(t)f (t)

st:

d1 (c, t)

cd(t),d(t)f (t)

sc (t);

(15)

cr(t);r(t)f (t)

c (t) min(e(t) et , P (t)), where

cr(t);r(t)f (t)
r1 (c, t) = V Uc (t) + Q(t)sc +
d1 (c, t) = Q(t)sc V Uc (t).

(P (t) )c (t);

(16)

2) Energy Control. In time slot t, the proxy adds (t) to

P (t) if P (t) .
The Content Selection procedure selects r(t) and d(t) to
maximize the objective function in Eq. (15) with two constraints. The first constraint guarantees that the devices downlink data rate in time slot t is enough to deliver the selected
contents, and the second constraint ensures the energy is
sufficient for the mobile device to receive the contents. This
problem is in fact a knapsack problem that can be solved
by optimization tools, e.g., CPLEX [16]. The Energy Control
procedure keeps track of the virtual queue P (t), and only adds
(t) to P (t) if P (t) does not overreach yet.
C. Performance Analysis
We present several theory bounds of the proposed algorithm3 .
Let us denote U as the maximum long term utility of the problem in Eqs. (6a)(6e) achieved by any stationary randomized
algorithms. Notice that such an algorithm requires to know
about future. Our proposed algorithm achieves a long term
arbitrarily close to U without knowing about future.
utility U
Theorem 1 (Utility lower bound): Our algorithm achieves a
lower bound of the time average utility:
T 1
1 X

U = lim
E{Ur ( ) Ud ( )} U , (17)
T T
V
=0
where is a constant defined in Lemma 1.
The theorem shows that by choosing an arbitrarily large V , the
achieved utility is arbitrarily close to the maximum value.
Theorem 2 (Real queue Q upper bound): Our algorithm restricts the storage limit for real queue Q as:
Q(t) V + max t.
(18)
3 The

detailed proofs are omitted due to the space limitations.

The theorem indicates the real queue is bounded by V +max

in any time slot. The larger V leads to a larger restriction on
Q. That is, there is a trade-off O(V, 1/V ) for the size of the
real queue and the achieved long term utility. That is, queue
size increases by V leads to long term utility increase by V1 .
An interesting observation from this analysis is that we
can control the storage size, which is not unlimited at the
broker/proxy. Another interesting observation from this theorem
is that we can control the time average queue delay. Queue
delay is the time difference between a content is created and
it is delivered to the mobile device. We define time average
= Q V +max where r is time
queue delay as follows: D
r
r
average data rate from the broker/proxy to the mobile device.
It is clear that queue delay can be controlled by V .
Theorem 3 (Energy upper bound): Our algorithm ensures
that the energy consumed in any time slot is bounded by:
P (t) + max t.
(19)
The theorem shows that our algorithm ensures energy budget
at any time is not more than a threshold. We know that energy
budget P (t) shared for the prefetching system may increase
when the device is charged. This theorem infers that the energy
budget is controlled even when the device is charged fully for
a long time.
D. A Near-Optimal Algorithm
We next present an efficient algorithm for content selection
that runs in O(nlog(n)) with an approximation factor of 2. That
opt
1 opt
is, the algorithm achieves apx
1 (t) 2 1 (t), where 1 (t)
is the optimal solution for the problem in Eqs. (15)(16) and
apx
1 (t) is the near-optimal solution.
Algorithm 2 A Near-Optimal Algorithm for Content Selection
Step 1: Given a set f (t) of contents fetched by the proxy so
far, select r(t) to maximize:
X
max:
(r1 (c, t) [d1 (c, t)]+ )
cr(t);r(t)f (t)

st:

(20)

sc (t).

cr(t);r(t)f (t)

This is done by:

r (c,t)[d (c,t)]+

1
1) For each c f (t), calculate wc = 1
.
sc
2) Add cP
in f (t) to r(t) in decreasing order of wc if wc > 0
until cr(t) sc (t) does not hold.
Step 2: Select d(t) from the set f (t)r(t) by iterating through
each content c in f (t) r(t), and add c to d(t) if d1 (c, t) > 0.

We first convert the problem in Eqs. (15)(16) to a new

problem. The energy consumption for receiving contents is
linear to data size [14], i.e., (t) = a (t)s(t) + b (t), where
s(t) is the total amount of contents transferred in time slot
t, and a (t) and b (t) are constants in each time slot, which
are functions of the network type and condition. Therefore, the
second
constraint of the problem in Eqs. (15)(16) becomes
P
min(e(t)et ,P (t))b (t)
, which leads to the recr(t) sc
a (t)
laxation of one of the two constraints. Let us define (t) =
(t))b (t)
min((t), (e(t)et,Pa (t)
). The problem in Eqs. (15)(16)

is equivalent
Xto the following problem:

(r1 (c, t) [d1 (c, t)]+ ) +

max:

cr(t)
r(t)f (t)

st:

[d1 (c, t)]+ ;

cf (t)

(21)
sc (t).

Facebook account, and present the results in Table I. This table

shows that using the broker/proxy saves significant energywe
notice a 6.9-fold reduction with WiFi and a 9.1-fold reduction
with cellular networks. The experimental results demonstrate
the value of even a simplistic broker/proxy solution.

cr(t);r(t)f (t)
P
Notice that in the new objective function cf (t) [d1 (c, t)]+ is
independent to r(t). This allows us to select the contents for
delivering r(t) first, and then derive the contents for dropping
d(t). We solve the problem in Eq. (21) in two steps, as
presented in Algorithm 2. The following theorems analyze the
performance of the near-optimal algorithm.
Lemma 2: Using Algorithm 2 for Content Selection achieves
a bound on 1 (t):
1 opt
(22)
apx
1 (t) 1 (t)
2
Theorem 4 (Utility bound): By using Algorithm 2 for Content Selection, the system achieves the following bound on the
long term utility:
apx 1 U apx .
(23)
U
2
V
We note that using Algorithm 2 achieves the same bounds on
the real queue and virtual queue as presented in Theorems 2
and 3.

IV. P ERFORMANCE E VALUATION

A. Prototype System and Evaluation

Figure 3.

Table I
E NERGY CONSUMPTION (J) AT DEVICES FOR OFFLINE ACCESS TO
FACEBOOK IN 30 HOURS
Approach
Without Broker/Proxy
With Broker/Proxy

WiFi
970.9 (6.9X)
140.2

Cellular
16,783.4 (9.1X)
1,843.1

We illustrate the feasibility of our system by developing a

prototype consisting of a broker/proxy program and an Android
app, called Offline Facebook [17]. This app can download
updates on users News Feeds page at Facebook including the
feed contents (texts, images, and videos), comments and likes,
and provides a friendly GUI for the users to view as well
as interact (e.g., comments and likes) with the downloaded
contents offline without network connectivity (an example is
presented in Fig. 3). The app can also receive updates from the
broker/proxy program.
To quantify the benefit of the broker/proxy solution, we conducted the following experiments: We install the broker/proxy
on an external Ubuntu machine, and the app on two Nexus
One phones running Android 2.3.6. We configure the apps
on the two phones to run with and without the broker/proxy,
respectively. Without the broker/proxy, the phone checks with
Facebook servers for the latest updates once every 5 minutes;
in the other case, the broker/proxy performs the checking of
updates and signals the phone using Google Cloud Messaging,
if any relevant updates appear. Upon being signaled the phone
establishes a connection to the broker/proxy and receives the
updates. We consider two access networks: WiFi and cellular
(T-Mobile), and use PowerTutor [18] to measure the power consumption of our app. We run 30-hr experiments using the same

GUI for our data collection Android app [17].

B. Trace-driven Simulations
1) Dataset: To evaluate the performance of our proposed
algorithm, we conducted trace-driven simulations using Facebook user data collected through our Android app [17]. We
release our app to 10 users in America, Asia and Europe,
and collect their traces between May 15 and 24, 2013. With
their consensus, we log and collected information on their
social attributes (e.g., Facebook friends and groups), news
feeds attributes (e.g., their authors, created time, downloaded
time, and size), device resources (e.g., battery and network
conditions), user-app interactions (e.g., timestamps of clicks to
view, comment, and like a news feed), and etc. To protect users
privacy, all Facebook objects, such as users, friends, groups and
news feeds, are one-way encoded to prevent being traced back
to the original users and contents. Our dataset includes 12,596
multimedia contents for all 10 users. The maximum number of
contents for a user is 3,919, and the minimum number is 130.
A user has 310 friends on average. The maximum number of
friends is 503 and the minimum is 75. Each user clicks to view
contents 693 times on average. The maximum ratio between
the number of clicks and the number of contents is 95%, and
the minimum ratio is 17%.
2) Settings: We built a Java based simulator that is driven by
the collected traces: the feed stream trace was used to model
Facebooks content arrival while the network trace to model
network conditions for the simulation. Energy consumption
was simulated using PowerTutors models [18]. Optimization
solver [16] was used to solve the optimization problem in
the Content Selection procedure in Algorithm 1. We run the

simulations on a Windows 7 machine with an Intel Dual Core

1.8 GHz CPU with 4 GB RAM. We compare our algorithms,
OPT (the optimal solution presented in Section III-B) and
NOPT (the near optimal solution presented in Section III-D)
with 3 baselines algorithms:
Rand: This algorithm selects contents randomly to deliver
to user.
Size: Contents with smaller size are given higher priority
for delivery selection.
Prob: Contents with higher viewing probability are given
higher priority for delivery selection.
In all algorithms, we do not download contents when the current
energy level on the mobile device is lower than the energy
threshold. The energy threshold in the simulation is set to the
average energy level per user by default. For our algorithms,
we set parameter = 10 to place a serious consideration on
content dropping. is set to 20% of the maximum battery
capacity. All users had been set to have the same energy budget
periods in a day; more specially, a day time was divided into 3
periods: working period from 7 A.M. to 6 P.M, hangout period
from 6 P.M. to 11 P.M., and home period 11 P.M. to 7 A.M.
The corresponding usage energy percentage for these periods
are 20%, 30% and 40%.
We consider several metrics to evaluate the algorithms.
Utility: Accumulative viewing probability of all social
contents received by a user.
Queue delay: Average delay from the time a content is
downloaded by the broker/proxy to the time it is delivered
to a user.
Delivery ratio: A ratio of the number of contents delivered
to a user to the number of contents downloaded by the
broker/proxy from Facebook for the user.
Hit late ratio: The ratio of number of contents scheduled
to be delivered to the user later than their actually viewing
time to the total number of contents delivered to the user.
Effective energy: Total energy consumed by the mobile
device to download contents divided by the number of
contents actually viewed by a user.
3) Simulation Results: We evaluate our system by showing
the performance of our proposed algorithms under varying
values of parameter V . This helps us understand the proposed
algorithms behaviors. Later, we compare our algorithms with
other baseline algorithms.
Performance of the algorithms. We start with comparing
OPT and NOPTs performance under different V values. Fig. 4
presents the utility achieved by our algorithms. We observe
that the increase of V leads to the increase of the utility in
both algorithms. For example, at a very small value, V = 10,
the utility is 51.52 for NOPT while at V = 106 , the utility is
393.88. It is because at a larger V , the system becomes more
hesitating to drop contents, and pushes content transfer to the
users more strongly. A higher number of contents received by
mobile device leads to a higher total viewing probability. This
is supported by Fig. 5, which presents delivery ratio of contents
delivered to the users. At V = 10, the delivery ratio is only 9%,
but at V = 106 , mobile device receives approximately 70%.
This result proves the correctness of our performance analysis

in Section III-C.
An interesting result in Figs. 4 and 5 is the two algorithms,
OPT and NOPT, achieves very close utility and delivery ratio
(our experiments also indicate they have very similar results for
the other metrics). For example, at V = 100, NOPT achieves
the utility of 79.14 while OPT has a slightly higher value, 80.94.
The largest gap of utility between the two algorithms is only
4.36 in any V . The key difference between these algorithms
is at the running time. Using optimization solver CPLEX to
solve the OPT problems leads to the worst case running time
of 2.4 secs in our experiments. The worst case running times
for NOPT is only 78 ms. This indicates that NOPT runs much
faster than OPT while achieving a very close performance. We
thus suggest to use NOPT for the practical deployment purpose,
and from now on, we do not report experiment results for OPT.
Although larger V leads to higher utility, it also leads to
larger queue delay. For example, in Fig. 6 at V = 10, the
queue delay for NOPT is 0.5 hours while at V = 106 , the
delay is more than 8 hours. A reason for the high delay is
we set energy threshold to be high (average energy level).
But we see a trend here. With larger V , more contents in
queue are considered for being delivered to mobile devices.
This observation is inline with our analysis: queue size and
queueing delay are strongly dependent on V . Transferring a
larger number of contents causes high queue delay. Note that in
the social network context, queue delay in some sense indicates
the satisfaction of user: how soon they can view social content
updates from their friends.
We report effective energy according to V in Fig. 7. The
lower the effective energy, the better the system works, because
more energy is used for delivering right contents. Fig. 7 demonstrates that when V increases, the effective energy decreases
and then increases. For example, at V = 101 , the effective
energy is 18 J per content; at V = 104 , it decreases to 11 J per
content; it increases back to 30 J per content when V = 106 .
This trend can be explained as follows. With a small V , queue
size is dominated in making content selection for delivery
and dropping, and viewing probability does not have much
impact on decisions. Thus, the effective energy is low at small
V . With a medium V , the viewing probability plays a larger
role, and a higher number of contents clicked to view by the
user is delivered, which contributes to achieving low effective
energy. With a large V , the system delivers more contents than
necessary if network and energy conditions permit. Therefore,
there are more contents not clicked to view by the user delivered
to the user. This leads to higher effective energy.
Compared with other strategies. Fig. 8 shows the effective
energy to download contents viewed by users using each
algorithm under two energy thresholds. For fairness, we set
parameter = 10000 to prevent our algorithm from dropping
contents so that all algorithms download all contents (with
different schedules) under the same energy and bandwidth
constraints in the simulations. The proposed algorithms exhibit
better energy efficiency than the baseline algorithms: 15%30%
difference is observed.
Fig. 9 presents the hit late ratio of the downloaded contents,
which indicates the timeliness of content downloads that are

Delivery Ratio (%)

NOPT
OPT

Utility

300
200
100
0

102

10

103

104

105

106

70
60
50
40
30
20
10
0

10

NOPT
OPT

Queue Delay (hrs)

400

102

10

103

Figure 4.

Utility.

Figure 5.

20
15
10
5
102

10

103

Delivery ratio.

104

105

20
10
0

106

Rand

Effective energy.

Figure 8.

Size
Prob
Algorithms

3
68

4
82

5
77

6
65

7
62

8
70

9
67

NOPT

Effective energy.

Table II
V IEWING PREDICTION ACCURACY (AUC) IN % OF EACH USER . T HE
AVERAGE ACCURACY IS 72%.
2
77

103

Figure 6.

30

viewed by users. We observe that our proposed algorithm

achieves the best hit late ratio. Counter intuitively, the Prob
algorithm performs worst in hit late ratio. This is because the
algorithm fails to consider content size and resource conditions
to download, although it prioritizes the contents on viewing
probability. Downloading a long video may significantly delay
the downloads of other smaller contents with similar viewing
probability. In contrast, our algorithm jointly considers the
viewing probability, content size, and resource conditions.

1
69

102

10
81

User-content relevance evaluator. We briefly present the

results for our user-content relevance evaluator. We evaluate
this component using a well-known metric AUC (the area
under the ROC curve), which indicates the prediction accuracy.
AUC varies from 0 to 100%. The closer to 100% AUC is,
the higher accuracy the evaluator achieves. AUC is a good
quality indicator for skewed distribution data set. We report
AUC using 5-fold cross validation. More specifically, the data
set is randomly divided into 5 partitions with equal size. 4
of 5 partitions were used in the training process, and the last
partition was used to test learning model. The process was
repeated 5 times, each using a different partition for testing.
The AUC achieved by our evaluator varies from 62% to 81%
with the average of 72%. Table II presents AUC in details for all
considered users, which is fairly good: always 62%. Another
observation is that the training overhead is low showing the
efficiency of the proposed evaluation algorithm. The same

104

105

106

40

Figure 7.

2
10

Hit Late Ratio (%)

25

106

50
Effective Energy

Effective Energy

105

30

User
AUC

104

Queue delay.

1.5
1.2
0.9
0.6
0.3
0
Rand

Size
Prob
Algorithms

Figure 9.

NOpt

Hit late ratio.

Windows PC described above is used for the training process

using MATLAB with 8 day data (i.e., including 4 partitions
of the data traces). The average running time is only 0.006
seconds while the maximum is 0.138 seconds.
In summary, through real-world experiments and simulation
studies, we have shown that:
The broker-proxy approach is meaningful and yields significant performance benefits over a non-proxy based
solution.
The near-optimal solution is remarkably close in performance to the optimal solution that takes a long time to
complete/converge.
The near-optimal solution outperforms other state-of-theart approaches.
Carefully tuning the control parameter V significantly
improves the performance.
V. R ELATED W ORK
Content prefetching has been studied in several domains,
such as in processors that fetch data and instructions from
RAM or disks in advance [19], and in distributed systems that
fetch files from remote servers before receiving requests from
applications [20]. Prefetching also attracts many attentions in
mobile systems in which resources such as energy and network
connectivity are constrained. Recent studies for prefetching in
mobile systems can be classified into two categories: clientbased [21][23] and broker/proxy-based approaches [24], [25].
In client-based approaches [21][23], mobile devices make
prefetching decisions by themselves. Balasubramanian et
al. [21] attempt to reduce cellular data traffic and support delay
tolerant applications by exploiting intermittent WiFi connectivity. In [22], Schulman et al. predict time periods when cellular
signal strength is strong, and transfer data in those periods to

reduce energy consumption, as low signal strength generally

leads to higher energy consumption. Higgins et al. [23] propose
to prefetch contents for applications in mobile devices based on
the tradeoff between the energy cost and delay benefit.
In broker/proxy-based approaches [24], [25], mobile devices
prefetch contents with supports from broker/proxy. In [24],
Armstrong et al. employ a proxy to fetch updates for emails
and weather reports, and batch the updates before sending them
to mobile devices. Relevant updates are delivered based on
interests expressed by users to the proxy. Mohan et al. [25]
develop an ads delivery system, in which mobile devices fetch
ads from servers via broker/proxies. They focus on the problem
of prefetching ads to mobile devices before the ads are expired.
In contrast to the prior work, we develop an offline social
network system for mobile devices to efficiently download
social media contents leveraging broker/proxy, which is challenging for several reasons. First, social network sites typically
do not offer notification services for new updates, as the
number of updates is tremendous. Second, although there is
relevance between users and contents, the arrival patterns of
social media contents are hard to predict. In the current paper,
we employ broker/proxy to optimally deliver social media
contents to mobile users by carefully analyzing the user-content
relevance and mobile device dynamics. Our system is built
on fine-grained control algorithms that select what contents to
send/drop with a strong emphasis on energy conservation. To
our best of knowledge, our paper is the first of its kinds. The
lessons learned in our work are crucial to app developers and
social network providers for realizing high-quality mobile user
experience for offline social network access.
VI. C ONCLUDING R EMARKS
Today, mobile access to current social media information
is based on the connectivity of the end-user; our aim in this
paper is to exploit the delay-tolerant nature of user access
to social media streams to prefetch social media contents
at a users mobile device when connectivity is good and
cheap. We develop and evaluate algorithms and an end-to-end
system to make the mobile social media experience seamless,
personalized, and cost-effective and efficient. The crux of our
approach is to leverage more capable resources in the network
to assist the mobile device in performing resource-intensive
storage and computation tasks, even when the user is offline.
Towards this end, we propose a broker-proxy architecture that
can efficiently sort the social media contents based on their
importance to the user; and schedule the content downloads
to the mobile device based on content relevance and current
system conditions. The feature is implemented via a mobile app
on Android platform and used to collect social network traces.
We then conducted experiments on our testbed and extensive
trace-driven simulations to evaluate our proposed approach and
system. Our evaluation results show the value of the distributed
broker/proxy/device architecture and the associated algorithms
that control the transfer of contents within this system and
outperform the baseline approaches.
As indicated, a full fledged real world deployment of the
proposed vision requires other issues that must be resolved.

These include challenges introduced by scale (large number of

mobile users), their wide distribution (global nature of social
interactions), mobility (travelling users) and security (shared
nature of wireless access). An interesting approach to address
the scaling of mobile social network applications like the one
described in this paper, is the utilization of the upcoming cloud
computing infrastructure (public, private, and edge clouds) to
implement brokers/proxies. Resource intensive techniques such
as ranking of contents across users and across social media
sites can execute in public clouds (data centers) or in regional
proxies. A structured approach to address these issues with
meaningful practical solutions are key topics being addressed
in our future research.
R EFERENCES
[1] Smart and social: Android phones in india, http://www.nielsen.com/us/
en/newswire/2012/smart-and-social-android-phones-in-india.html, May
2012.
[2] Always
connected,
https://fb-public.box.com/s/
3iq5x6uwnqtq7ki4q8wk.
[3] R. Rahimi, N. Venkatasubramanian, and A. Vasilakos, MuSIC: On
mobility-aware optimal service allocation in mobile cloud computing,
in IEEE Cloud13.
[4] Y. Zhao, N. Do, S. Wang, C. Hsu, and N. Venkatasubramanian, O2sm:
Enabling efficient offline access to online social media and social
networks, in Proc. of ACM/IFIP/USENIX Middleware, Beijing, China,
December 2013, pp. 228242.
[5] J. Ahn and et al., Open user profiles for adaptive news systems: help or
harm? in ACM WWW07.
[6] W. Chu and S. Park, Integrating tags in a semantic content-based
recommender, in ACM WWW09.
[7] M. Gemmis and et al., Personalized recommendation on dynamic content
using predictive bilinear models, in ACM RecSys08.
[8] F. Cacheda and et al., Comparison of collaborative filtering algorithms:
Limitations of current techniques and proposals for scalable, highperformance recommender systems, ACM Transactions on the Web11.
[9] B. Sarwar and et al., Item-based collaborative filtering recommendation
algorithms, in ACM WWW11.
[10] H. Ma, I. King, and M. Lyu, Effective missing data prediction for
collaborative filtering, in ACM SIGIR07.
[11] D. Liu and et al., Hybrid social media network, in ACM MM12.
[12] J. Noel and et al., New objective functions for social collaborative
filtering, in ACM WWW12.
[13] H. Ma and et al., Recommender systems with social regularization, in
ACM WSDM11.
[14] N. Balasubramanian, A. Balasubramanian, and A. Venkataramani, Energy consumption in mobile phones: A measurement study and implications for network applications, in IMC09.
[15] L. Georgiadis, M. Neely, and L. Tassiulas, Resource allocation and
cross-layer control in wireless networks, Foundations and Trends in
Networking06.
[16] IBM CPLEX optimizer, http://www-01.ibm.com/software/commerce/
optimization/cplex-optimizer/.
[17] oFacebook, http://www.ics.uci.edu/dsm/oFacebook.
[18] PowerTutor, http://ziyang.eecs.umich.edu/projects/powertutor.
[19] A. Smith, Cache memories, ACM Computing Surveys82.
[20] J. Kistler and M. Satyanarayanan, Disconnected operation in the coda
file system, ACM Transactions on Computer Systems12.
[21] A. Balasubramanian, R. Mahajan, and A. Venkataramani, Augmenting
mobile 3G using WiFi, in ACM Mobisys10.
[22] A. Schulman, V. Navda, R. Ramjee, N. Spring, P. Deshpande, and
C. Grunewald, Bartendr: A practical approach to energy aware cellular
data scheduling, in ACM Mobicom10.
[23] B. Higgins, J. Flinn, T. Giuli, B. Noble, C. Peplin, and D. Watson,
Informed mobile prefetch, in ACM Mobisys12.
[24] T. Armstrong, O. Trescases, C. Amza, and E. Lara, Efficient and transparent dynamic content updates for mobile clients, in ACM Mobisys06.
[25] P. Mohan, S. Nath, and O. Riva, Prefetching mobile ads: Can advertising
systems afford it? in ACM Eurosys13.