Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.

Digital Object Identifier 10.1109/ACCESS.2017.DOI

Social Context-Aware Mobile DataOffloading Algorithm via Small CellBackhaul NetworksHYE-RIM CHEON1, (Student Member, IEEE) AND JAE-HYUN KIM 1, (Member, IEEE)1Department of Electrical and Computer Engineering, Ajou University, Suwon, Korea

Corresponding author: Jae-Hyun Kim (e-mail: jkim@ajou.ac.kr).

This work was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by theMinistry of Science, ICT and future Planning (NRF-2014R1A2A2A01002321)

ABSTRACT In recent years, total mobile traffic has increased explosively, and this has led to severetraffic load on the mobile network operator (MNO)’s core network. Mobile data offloading is a promisingsolution to alleviate the core network’s load. Through such mobile data offloading, MNO can reroute themobile traffic to other access networks using various radio access technologies, such as WiFi, opportunisticcommunications, and etc. In addition, social networking traffic has risen sharply due to the popularityof online social networking services, such as Facebook and Twitter. Thus, we need to consider a socialcontext for effective mobile data offloading. In this paper, in order to apply the social context to mobiledata offloading, we model the social context in terms of two aspects: the user’s social relationships andapplication’s popularity. In addition, we propose a social context-aware mobile data offloading algorithmto maximize the quality of service (QoS) of users in a small cell backhaul offloading environment. Theperformance evaluation results demonstrate that the proposed algorithm outperforms the other algorithmswithout considering the social context.

INDEX TERMS Mobile data offloading, small cell backhaul networks, social context

I. INTRODUCTION

DUE to the proliferation of smart mobile devices (e.g.smartphones, tablets, etc.), with advanced computing

and multimedia capabilities, highly evolved mobile applica-tions have been introduced. These include complicated com-putations that generate large amounts of traffic which haveled to the explosive growth of total mobile traffic. Accordingto Cisco, smart devices accounted for 89 percent of the totalmobile data traffic generated despite only accounting for 36percent of all mobile devices; further, total mobile trafficincreased 74 percent [1]. In mobile applications, the use ofsocial networking services (SNSs) such as Facebook andTwitter is increasing sharply, and the corresponding trafficaccounts for the second largest contribution to total mobiledata traffic, and the use of embedded video (video contents orvideo links shared by SNS users) in SNSs is growing as well[2]. This considerable traffic from social networking mayplace a considerable load on the mobile network operator(MNO)’s core network and lead to degradation in the qualityof service (QoS) in terms of mobile users. Therefore, studies

are necessary to distribute the traffic load considering socialcontext.

Mobile data offloading is one of the most promising solu-tions to alleviate the traffic load in the MNO’s core network.According to Cisco, 51 percent of the total mobile datatraffic in 2015 was offloaded onto the fixed network throughWi-Fi or femtocell [1]. Many researchers have studied theoffloading schemes based on the various offloading networks,such as WiFi, opportunistic communication, and small cellbackhaul networks. In WiFi offloading, the MNO can offloadthe data to free unlicensed bands. However, from the user’sperspective, they may experience poor QoS due to the lim-itation of the WiFi coverage area [3]. Another approach tomobile data offloading is opportunistic communication. Inopportunistic offloading, the MNO might deliver the data toonly target users, and then, target users can disseminate thedata to other users who can communicate with target users viaopportunistic communications. However, this is unreliablefor mobile data offloading due to the battery and storagelimitation of mobile devices, along with the low incentives

VOLUME 4, 2016 1

H. R. Cheon et al.: Social Context-Aware Mobile Data Offloading Algorithm via Small Cell Backhaul Networks

for the users of these mobile devices involved in offloading[3]. The other approach to mobile data offloading is to offloadthe traffic via small cells, such as femtocells and picocells.In small cell offloading, mobile users can directly route thetraffic to the same residential/enterprise network or Internetthrough small cell backhaul without traversing the mobilecore network. 3GPP proposed the local IP access (LIPA) andselected IP traffic offload (SIPTO) for traffic offloading [4].According to the [5], small cell offloading is effective forthe following reasons: the MNO can offload heavy users viasmall cells (81 percent of data usage occurs mainly indoor)and provide a seamless service to users through an operator-deployed and managed small cell. Thus, in this paper, wefocus on the small cell offloading.

As mentioned previously, social networking traffic is in-creasing, and this exceedingly large traffic is from the re-peated downloading of the same popular content [6]. Inaddition, the popularity of the content is highly affectedby the user’s social relationships [7]. Thus, applying socialcontexts to traffic offloading is crucial. Some works exploitsocial context to offload the traffic and most of these worksare even based on opportunistic communications [8]– [10],but social context is still not yet widely applied to small celloffloading. As mentioned previously, opportunistic commu-nication for mobile data offloading is limited due to severalreasons. Therefore, in order to fully exploit social context inmobile data offloading, we propose a mobile data offloadingalgorithm considering social context based on the small celloffloading, which offloads the data traffic via small cell fixedbackhaul networks.

In order to analyze the social context and apply this tothe offloading decision, sufficient storage and sophisticatedcalculation capabilities without high latency are required.Thus, we consider multi-access (or mobile) edge computing(MEC) for mobile data offloading. MEC provides a cloudcomputing-based IT service environment at the edge of themobile network that is within the radio access network(RAN) [11] and can ensure ultra-low latency and high band-width. However, the effect of MEC (e.g. low latency, and etc)is beyond the scope of this paper.

The main contributions of this paper include:

• New social context modeling: We model the socialcontext to propose the mobile data offloading algo-rithm, which take advantage of the user’s social rela-tionship and the popularity of mobile application ser-vices. Through social context modeling, we estimate theprobability that each user will select a mobile applica-tion service. For convenience, we will shorten ‘mobileapplication services’ to ‘applications’ in this paper.

• Social context-aware small cell mobile data offload-ing algorithm: We propose a social context-aware mo-bile data offloading algorithm to maximize the QoS ofsmall cell users, which is associated with a small cell,in the small cell mobile data offloading. The proposedalgorithm determines an offloading weighting factor for

each application of each user by estimating the applica-tion selection probability and network utilization.

• Performance evaluation: We conduct simulations todemonstrate the good performance of the proposed al-gorithm compared to other algorithms without consid-ering social context modeling.

The remainder of the paper is organized as follows. InSection II, we review the related works. In Section III, wedescribe the system model including social context modeling.In Section IV, we describe the proposed social context-aware small cell mobile data offloading algorithm in detail.In Section V, we analyze the performance of the proposedalgorithm. Finally, we conclude this paper in Section VI.

II. RELATED WORKSA. MOBILE DATA OFFLOADINGAs previously stated, mobile data offloading is a promisingsolution to alleviate traffic load. Many works have inves-tigated mobile data offloading. In [3], they examined thestate-of-the-art works in mobile data offloading in terms ofboth technologies and business. A few works have beenfocused on WiFi offloading [12], [13]. In [12], they presentedquantitative research into the performance of mobile dataoffloading via WiFi networks. In [13], they proposed WiFioffloading mechanisms for vehicular cellular traffic, which isoffloaded via carrier-WiFi networks.

A few existing works have investigated small cell offload-ing, such as LIPA/SIPTO [14], [15]. In [14], a bearer-basedoffloading algorithm was proposed to support a LIPA/SIPTOsolution in the MNO’s core network. In [15], they discussedthe main requirement to support LIPA/SIPTO and provided asurvey on the aspects of management and service continuityin LIPA/SIPTO.

B. SOCIAL CONTEXT IN COMMUNICATIONSA number of works have examined the social context incommunications [16], [17]. In [16], they introduced the basicconcepts of socially aware networking (SAN) and presenteda survey of state-of-the-art works in SAN. In [17], they pro-posed a combined social and communication network modeland analyzed the average delay and success probability inthis combined model. In addition, many works have studiedthe exploitation of social context in communications andnetworks [18]- [21]. In [18], they provided a survey on state-of-the-art social-aware routing protocols in delay tolerantnetworks. In [19], they proposed an algorithm to optimizeresource allocation in device-to-device (D2D) wireless smallcell networks, which jointly exploits the wireless and socialcontext of users. In [20], they proposed a framework toimprove the quality of recommendations in location-basedsocial networks, which exploits social context as well asspatial and temporal context.

C. SOCIAL CONTEXT IN MOBILE DATA OFFLOADINGA few works have considered social context in mobile dataoffloading [8]- [10], [22]. In [8], they investigated the target-

2 VOLUME 4, 2016

J. H. Kim et al.: Social Context-Aware Mobile Data Offloading Algorithm via Small Cell Backhaul Networks

set selection problem in mobile traffic offloading via oppor-tunistic communications. In addition, through a trace-drivensimulation study, they showed that social participation is acritical factor in opportunistic-communication-based mobiledata offloading. In [9], they proposed a traffic offloadingframework via opportunistic sharing, which depends on theuser’s content spreading impacts in online SNSs as wellas the mobility patterns in offline mobile social networks.In [10], they proposed a novel traffic offloading frameworkvia D2D sharing, which considers the tags on users andcontents. In [22], they proposed two social-aware incentivemechanisms in a WiFi access point (AP) based offloading,which is based on the modeling of social relationships ofmobile users by a PageRank algorithm.

As shown here, most of offloading algorithms consideringsocial context have been based on opportunistic commu-nications. However, as mentioned in the Introduction, weconsider the small cell mobile data offloading using the socialcontext. In this paper, we take into consideration the two as-pects of social contexts, such as the user’s social relationshipand the application’s popularity. In addition, we investigatehow the social context affects the performance of the mobiledata offloading.

III. SYSTEM MODELA. NETWORK ARCHITECTUREWe consider a mobile data offloading network architecture asshown in Fig. 1, which consists of a small cell, a MEC server,an MNO’s core network, a small cell backhaul network, andmultiple mobile device users. We assume that all mobiledevices are associated with the small cell and that the smallcell type is for residential and enterprise use, which have thewireline backhaul network (e.g. broadband Internet service)and can directly connect the Internet. We assume that theMEC server is located in close proximity to the small cell.The MEC server plays a key role in mobile data offloading,which collects the information to analyze the social contextand network utilization, and decides the amount of offloadingthe mobile traffic for each application of each user. After theoffloading decision, the offloading traffic is routed from theMNO’s core network to the small cell’s backhaul network.In order not to consider a user’s mobility in this small celloffloading, we assume that all users are fixed at the samedistance from a small cell. We assume that the users havethe social relationships with each other through the onlinesocial network activities and that the application’s selectionbehavior is affected by these social relationships.

B. MODELING THE SOCIAL CONTEXTBased on the [6] and [7], we model the social context usingthe following attributes:• User’s social impact: According to [7], the social re-

lationships are one of the primary schemes for whichthe users find and select applications. In addition, someusers with a high social impact have a strong influenceon the application popularity. Thus, we assume that a

MEC server

Internet

Residential/enterprise network

MNOcore network

S-GW P-GW

Backhaul

Mobile application service server

Physical domain

Social domain

User 1

User 3

User 2

User 4

User 5

User 6

User 1

User 2User 5

User 6

User 4User 3

Online social network

Small cell

Core networktraffic

Small cell backhaul offloading traffic

FIGURE 1. The network architecture for traffic offloading considering socialimpact in a social network.

user’s application selection behavior is affected by theselection behavior of other neighbor users, where theneighbor users refer to the users who have some socialrelationship with a specific user. In addition, we assumethat the most popular application of the most sociallyinfluential user will be frequently selected.

• Application’s popularity: According to [6], the pop-ularity distribution of contents shows power-law withtruncated tails, and this means that the most popularapplications are frequently selected. Thus, we assumethat a user’s application selection is affected by thenumber of times the application is selected.

First, we model social context in terms of each user’s socialimpact. As mentioned previously, we consider a user i, wherei ∈ I = {1, 2, ..., N}. The social relationship in the socialnetwork is modeled as a social graph using an undirectedgraph, in which G = {V,E} where V is the set of verticesand E is the set of edges, and edge (vi, vi′) ∈ E is identicalto the edge (vi′ , vi) ∈ E. Vertex vi denotes user i ∈ I andedge (vi, vi′) denotes a social relationship between user i anduser i′. Then, based on the social graph, we construct thesocial adjacency matrix, A =

[aii′]N×N , where N denotes

the total number of users, and the entity aii′ in A indicatesthat whether user i and user i′ have a social relationship ornot and it is given by

aii′ =

{0, (vi, vi′) /∈ E, (1a)1, (vi, vi′) ∈ E. (1b)

VOLUME 4, 2016 3

H. R. Cheon et al.: Social Context-Aware Mobile Data Offloading Algorithm via Small Cell Backhaul Networks

FIGURE 2. The example of a social graph generated by NodeXL

We use a degree centrality in the social graph as the user’ssocial impact in a social network as follows.

DegCi =deg(i)

n− 1, (2)

where deg(i) is the number of directly connected neighborsof user i and n is the maximum number of possible connectededges in the social graph [16]. Fig. 2 shows an example of asocial graph from the data set of ego-Facebook in SNAP [23],and the social graph is generated by NodeXL [24]. As shownin Fig. 2, the blue diamond and the black circle refer to theusers in the social network, and the solid line refers to thesocial relationship between the users. In particular, the bluediamond means one of the high degree centrality users, i.e.this blue diamond refers to the high social impact user in thesocial network.

Second, we model the social context in terms of theapplication’s popularity. We consider the case in which auser i selects application j, where j ∈ J = {1, 2, ...,M}.Based on the application’s selection frequency, we constructthe application selection history matrix, H =

[hij]N×M ,

where M denotes the total number of applications, and theentity hij refers to the number of times that user i selects theapplication j. In addition, we calculate the rank of applicationpopularity, rij , which indicates the rank for application j ofuser i. When the rank for application j of user i is smallest,this indicates that application j is most popular with user i.

Finally, considering both the user’s social impact andapplication popularity, we define the social impact-awarepopularity factor, which is given by

SIij = DegCi ·1/rij√∑j(1/rij)

2. (3)

This means that the highest rank application for the highestsocial impact user will affect the other user’s application se-lection, and this will apply to the estimation of the applicationselection probability in the next subsection.

C. ESTIMATION OF SOCIAL CONTEXT-AWAREAPPLICATION SELECTION PROBABILITYIn order to estimate the user’s application selection proba-bility, we use the Indian buffet process (IBP) model, whichis a stochastic process model that defines a probability dis-tribution over binary matrices by specifying a procedureby which customers (objects) choose dishes (features) [25].Then, in order to apply a user’s social impact and applicationpopularity to the estimation of the user’s application selec-tion, we define the social context-aware application selectionprobability based on the social context modeling described inthe previous subsection.

First, when a user does not have any social relationshipswith others, i.e. aii′ = 0,∀i′ ∈ Ii, where Ii is the set ofusers that may have a social relationship with user i, the useri’s application j selection probability follows the Poissondistribution and it is given by

Pij = SIij ·

[(α/i)m

i0

mi0!· e−αi

], (4)

where α is a Poisson distribution parameter and mi0 is the

number of new selected applications by user i. This meansthat the user is not affected by the other user’s selection dueto no social relationship.

By contrast, when a user has any social relationships withothers, i.e. aii′ = 1,∀i′ ∈ Ii, the user i’s application jselection probability follows the IBP model but is modifiedby whether or not the application j has previously beenselected at least once by user i and this is given by

Pij =

SIij ·

{ (α/i)mi0

mi0!· e−αi

}, hij = 0, (5a)

SIij ·{mi′

j

i

}, hij ≥ 1, (5b)

where mi′

j is the total sum of the number of selection forapplication j by other users that have a social relationshipwith user i and it is defined as

mi′

j =∑i′

hi′j ,∀i′ ∈ Ii. (6)

mi′

j is the principal difference from the original IBP model.In the original IBP model, they simply take account of thetotal number of application selections without consideringwhether the user really knows the number of applicationselections for other users.

IV. PROPOSED SOCIAL CONTEXT-AWARE SMALL CELLMOBILE DATA OFFLOADING ALGORITHMIn this section, we propose the social context-aware mobiledata offloading algorithm through small cell backhaul net-works. The objective of the mobile data offloading algorithmis to determine the offloading weighting factor to maximize

4 VOLUME 4, 2016

J. H. Kim et al.: Social Context-Aware Mobile Data Offloading Algorithm via Small Cell Backhaul Networks

TABLE 1. Notations

Notations ExplanationI Set of usersJ Set of applicationsi User indexj Application indexN Number of usersM Number of applicationsA Social adjacency matrixH Application selection history matrixaii′ Social relationship indicator between user i and i′

hij Number of times the user i selects the application jDegCi Degree centrality of user irij Rank of application for application j of user iSIij Social impact-aware popularity factor for applica-

tion j of user iIi Set of users that have a social relationship with user

iPij Social context-aware application selection probabil-

ity for application j of user imi

0 Number of new selected application by user imi′

j Total sum of the number of selection for applicationj by other users that have a social relationship withuser i

wij Selected offloading weighting factor for applicationj of user i

Wij(u) Selectable offloading weighting factors for applica-tion j of user i

UBN , UCN Network utilization of backhaul and core networkCBN , CCN Network capacity of backhaul and core networkmi

j Number of time the user i selects the application j

Rij , Dij , P eij Expecte data rate, transmission delay, and packeterror loss rate for application j of user i

RjBN , Rj

CN Backhaul network and core network supportabledata rate for application j

DjBN , Dj

CN Backhaul network and core network transmissiondelay for application j

PejBN , P ejCN Backhaul network and core network packet errorloss rate for application j

Lj Average packet size for application jO Objective functionwR, wD, wPe QoS weighting factors for data rate, transmission

delay, and packet error loss ratewijmax Offloading weighting factor for application j of user

i to maximize the QoSTLij Traffic volume for application j of user iTLoff

ij Offloading traffic volume for application j of user iWij Set of the offloading weighting factor values for

application j of user iwk

ij kth offloading weighting facor value for applicationj of user i

K Number of intervalsRk

ij , Dkij , P ekij The kth expecte data rate, transmission delay, and

packet error loss rate for application j of user iOij Set of the objective function values for application

j of user iokij kth objective function value for application j of

user iJ−j Set of applications except application jwijmax Final offloading weighting factor for application j

of user i to maximize the QoS

the QoS of users and the detailed algorithm is stated inAlgorithm 1.

First, we calculate the selectable offloading weighting fac-tors based on (4), (5), and network utilization. These factors

enable to offload more traffic for the application with highersocial context-aware probability, and to offload more trafficto the backhaul network when the core network utilization isrelatively higher than the backhaul network utilization. Basedon this, we define the selectable offloading weighting factorsWij(u) as follows.

Wij(u) =Pij√∑j P

2ij

· u

UCN + UBN, (7)

where u ∈ [UCN, UCN + UBN], and UBN and UCN are theutilization of the backhaul and core network, respectively, asgiven by

UBN =

∑i

∑j wij · Lj ·mi

j

CBN, (8)

UCN =

∑i

∑j(1− wij) · Lj ·mi

j

CCN, (9)

where mij is the number of times user i selects application

j, and CBN and CCN are the backhaul and core networkcapacities, respectively.

Next, we calculate the selectable expected QoS values,Rij

is the expected data rate, Dij is the expected transmissiondelay, and P eij is the expected packet error loss rate forapplication j of user i when the offloading weighting factoris wij , which is selected from Wij(u). These QoS values aregiven by

Rij = wijRjBN + (1− wij)R

jCN, (10)

Dij = wijDjBN + (1− wij)D

jCN, (11)

P eij = wijP ejBN + (1− wij)P ejCN, (12)

where RjBN and Rj

CN are the backhaul network and corenetwork supportable data rate for application j, respectively;this means that the network can actually support data ratesin the current network condition, and they are described indetail [26]. Dj

BN and DjCN are the average transmission delay

for application j in the backhaul network and core network,respectively, and these are defined as

DjBN =

Lj

RjBN

, (13)

DjCN =

Lj

RjCN

, (14)

where Lj is the average packet size for application j. P ejBN,P ejCN are the packet error loss rate for application j in thebackhaul network and core network, respectively, which varyaccording to the network utilization.

Next, we define the objective function to maximize theQoS of users and this is defined as

O =∑i

∑j

[wR

Rij√∑j R

2ij

+ wD1/Dij√∑j(1/Dij)2

+ wPe1/P eij√∑j(1/P eij)2

],

(15)

VOLUME 4, 2016 5

H. R. Cheon et al.: Social Context-Aware Mobile Data Offloading Algorithm via Small Cell Backhaul Networks

Algorithm 1 Social context-aware small cell offloading al-gorithm

1: Phase1 Estimate the application selection probability

2:Using the social context model, estimate the appli-cation selection probability by (4) and (5)

3: Phase 2 Find the selectable offloading weighting factors4: for i = 1 to N do5: for j = 1 to M do

6:Calculate the selectable offloading weightingfactors, Wij(u) by (7)

7: end for8: end for9: Phase 3 Calculate the selectable expected data rate,

delay, packet error loss rate10: for i = 1 to N do11: for j = 1 to M do12: Calculate Rij , Dij , P eij13: end for14: end for15: Phase 4 Determine the maximum offloading weighting

factor16: repeat

17:

Calculate the objective functionO and determinethe wijmax value for each application j of eachuser i from the Wij(u)

18: until When O is maximum19: Phase 5 Calculate the offloading volume20: Calculate the offloading traffic volume for each ap-

plication of each user, TLoffij = wijmax · TLij

wherewR, wD, andwPe are the QoS weighting factors for datarate, delay and packet error loss rate, respectively, and it de-pends on each user’s QoS parameter preference, which refersto how important each QoS parameter is. Then, we calculatethe objective function for each expected QoS value and repeatthis until the objective function reaches its maximum. Whenthe objective function is maximum, we finally determine themaximum offloading weighting factor for each applicationj of user i, wijmax from the selectable offloading weightingfactors to maximize the QoS of total small cell users.

Next, we calculate the offloading ratio for each applicationof each user, TLoff

ij = wijmax · TLij and offload the trafficfor each application of each user to the backhaul network byTLoff

ij .However, finding the offloading weighting factor to max-

imize the objective function for all applications of all usersis difficult because every possible combination of offload-ing weighting factors for all applications of all users mustbe compared heuristically. Thus, we propose the offloadingweighting factor search algorithm as given in Algorithm 2.Using this algorithm, we find the offloading weighting factorfor a specific application of a specific user. In order todetermine this, we compare the objective function values ofa specific user when the offloading weighting factor value

Algorithm 2 Offloading weighting factor search algorithm1: for i = 1 to N do2: for j = 1 to M do3: for k = 1 to K do4: Calculate wk

ij ∈ Wij

5: end for6: end for7: end for8: for i = 1 to N do9: for j = 1 to M do

10: for k = 1 to K do11: Calculate Rk

ij′ , Dkij′ , and P ekij′

12: end for13: end for14: end for15: for i = 1 to N do16: for j = 1 to M do17: for k = 1 to K do18: Calculate the objective function value okij19: end for20: wijmax = arg max∀wkij∈Wij

Oij

21: end for22: end for23: return W = [wijmax]N×M

for the specific application varies whereas the offloadingweighting factor values for other applications are fixed.

First, let Wij = {w1ij , w

2ij , ..., w

kij , ..., w

Kij } denote the

set of the offloading weighting factor values, where K isthe number of intervals, and we calculate the kth offloadingweighting factor value for application j of user i as follows.

wkij = min(Wij)+ (k− 1) ·

max(Wij

)−min

(Wij

)K

. (16)

Next, we calculate the kth expected data rate, delay, andpacket error loss rate for application j of user i as follows.

Rkij = wk

ijRjBN + (1− wk

ij)RjCN, (17)

Dkij = wk

ijDjBN + (1− wk

ij)DjCN, (18)

Pekij = wkijP ejBN + (1− wk

ij)P ejCN. (19)

Let Oij = {o1ij , o2ij , ..., okij , ..., oKij } is the set of the objec-tive function values for application j of user i, and we definethe kth objective function value for application j of user i as

6 VOLUME 4, 2016

J. H. Kim et al.: Social Context-Aware Mobile Data Offloading Algorithm via Small Cell Backhaul Networks

follows

okij = wRRk

ij√∑∀j′∈J−j

(R1

ij′

)2+(Rk

ij

)2+ wD

1/Dkij√∑

∀j′∈J−j

(1/D1

ij′

)2+(1/Dk

ij

)2+ wPe

1/P ekij√∑∀j′∈J−j

(1/P e1ij′

)2+(1/P ekij

)2 ,(20)

where j′ ∈ J−j refers to one of the applications in the setof applications except application j. okij is obtained from theexpected QoS values for application j of user i when theoffloading weighting factor value for application j is wk

ij

from Wij while the offloading weighting factor values forother applications except application j are fixed to w1

ij .Then, we determine the final offloading weighting factor

value for application j of user i, wijmax, as follows.

wijmax = arg max∀wkij∈WijOij . (21)

This value means the offloading weighting factor value forthe maximum objective function value for application j ofuser i.

V. PERFORMANCE EVALUATIONIn this section, we evaluate the performance of the proposedalgorithm through the extensive simulations. We comparethe performance of the proposed algorithm with those oftwo other algorithms: the algorithm using the social contextmodel without considering the user’s social impact (withoutcentrality), and the algorithm not using the social contextmodel (without centrality & history).

A. EVALUATION ENVIRONMENTAs mentioned in Section III-A, we assume the small cellmobile data offloading architecture, which consists of users,a small cell, a MEC server, a MNO’s core network, and asmall cell backhaul network. We set the number of users asN = 60 and assume that all users are associated with thesmall cell. We assume that the backhaul network initiallyhas a slightly bad condition compared to the core network,but this network condition will vary according to networkutilization. Based on this assumption, we set the networkcapacity of the backhaul and the core network as CBN = 800Mbps and CCN = 1300 Mbps, respectively [27]. In addition,the packet error loss rate is randomly selected from a uniformdistribution over [10−3, 10−2] for the backhaul network and auniform distribution over [10−4, 10−3] for the core network[28] and they vary according to network utilization. We setthe number of applications as M = 100, and the traffic ofeach application is generated into four classes as determinedby [29], with the details shown in Table 2. We assume thatthis packet error loss rate varies according to the network

TABLE 2. Traffic for each application parameters

Parameter Data Rate RatioWeb Uniform(150,449) kbps 36%

Music/Audio Uniform(450,749) kbps 8%SD Video Uniform(750,1149) kbps 20%HD Video Uniform(1150,2300) kbps 36%

0 20 40 60 80 1000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Rank

App

licat

ion

Sel

ectio

n P

roba

bilit

y

IBP generationProposed(w/ centrality)w/o centralityw/o centrality & history

FIGURE 3. Application selection probability vs rank

0 20 40 60 80 1000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Rank

Off

load

ing

Wei

ghtin

g F

acto

r

Proposed(w/ centrality)w/o centralityw/o centrality & history

FIGURE 4. Offloading weighting factor vs rank

utilization. We use the ego-Facebook data set from [23] asthe social network. As provided in [30], we first perform theIBP for 10,000 rounds with 100 applications. Then, based onthe social context modeling, each user selects applicationsfrom 100 applications.

B. APPLICATION SELECTION PROBABILITYFigs. 3, 4 show the estimated application selection prob-ability and the offloading weighting factor based on theapplication’s rank. In Fig. 3, we compare the applicationselection probability of the proposed model with those of the

VOLUME 4, 2016 7

H. R. Cheon et al.: Social Context-Aware Mobile Data Offloading Algorithm via Small Cell Backhaul Networks

other two models to the IBP-generated data, because the IBPmodel is verified by [30], which is similar to real data. Weobserve that the curve of the model without centrality showsthe most similar tendency to that of the IBP curve, but thecurve of the proposed model is similar to the IBP curve onthe whole. In particular, when the application’s popularityrank is high, the probability value of the proposed model ismost similar to the value of IBP. In addition, as shown inFig. 4, the curve of the proposed algorithm and the algorithmwithout centrality show a similar tendency to that of thecurve in Fig. 3. Furthermore, we observe that the proposedalgorithm determines the offloading weighting factor to be ahigher value as the rank of application is higher. However,the curve of the algorithm without centrality and history doesnot show a similar tendency to the curve of Fig. 3 because thisalgorithm does not use the social context model to determinethe offloading weighting factor. From Fig. 3, 4, we canobserve that due to the social context modeling from twoaspects, the proposed algorithm can estimate the applicationselection probability similar to the real data and determinethe offloading weighting factor of high-rank application tobe higher.

C. INFLUENCE OF SOCIAL CONTEXTIn this subsection, we set the QoS weighting factors to thesame values as wR = 0.33, wD = 0.33, and wPe = 0.33 soas to only investigate the influence of social context. Fig. 5shows the expected QoS value according to the number ofiterations, which are averaged by all users and applications.We observe that the proposed algorithm outperforms theother algorithms in terms of data rate and transmission delay.In data rate, the performance of the proposed algorithm isimproved by about 28.6 percent compared to the value of thealgorithm without centrality and history, and in transmissiondelay, the performance of the proposed algorithm is improvedby about 37.5 percent compared to that of the algorithmwithout centrality. This indicates that the proposed algorithmcan determine the appropriate offloading weighting factor foreach application of each user due to the two aspects of socialcontext modeling, which improves the performance of datarate and delay much more and leads to an enhanced totalQoS of the user. However, in terms of packet error loss rate,the performance of the proposed algorithm is degraded byabout 6 percent. Since we assume a bad backhaul networkcondition compared to that of the core network, especiallyin terms of packet error loss rate, the proposed algorithmdetermines the offloading weighting factor, which resultsin the enhancement of the user’s total QoS despite slightdegradation of the packet error loss rate.

In order to demonstrate the QoS performance from anotherangle, we consider the QoS value according to the rank,which is averaged by all users. As shown in Fig. 6, in the pro-posed algorithm, the QoS value of the high-rank applicationis much better compared to other algorithms, especially fordata rate and transmission delay. In the proposed algorithm,in order to maximize the total QoS of users, this might set

0 10 20 30 40 505

6

7

8

9

10

11

12x 10

4

Number of Iterations

Dat

a R

ate

(bps

)

Proposed(w/ centrality)w/o centralityw/o centrality & history

(a)

0 10 20 30 40 500.2

0.25

0.3

0.35

0.4

0.45

0.5

Number of Iterations

Tra

nsm

issi

on D

elay

(se

c)

Proposed(w/ centrality)w/o centralityw/o centrality & history

(b)

0 10 20 30 40 50

6.5

6.6

6.7

6.8

6.9

7

7.1

7.2

7.3

7.4x 10

-4

Number of Iterations

Pac

ket

Err

or L

oss

Rat

e

Proposed(w/ centrality)w/o centralityw/o centrality & history

(c)FIGURE 5. Average expected QoS value vs. iterations (a) data rate (b)transmission delay (c) packet error loss rate

8 VOLUME 4, 2016

J. H. Kim et al.: Social Context-Aware Mobile Data Offloading Algorithm via Small Cell Backhaul Networks

0 20 40 60 80 1000.5

1

1.5

2

2.5

3

3.5

4x 10

5

Rank

Dat

a R

ate

(bps

)

Proposed(w/ centrality)w/o centralityw/o centrality & history

(a)

0 20 40 60 80 1000.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

0.55

Rank

Tra

nsm

issi

on D

elay

(se

c)

Proposed(w/ centrality)w/o centralityw/o centrality & history

(b)

0 20 40 60 80 1000.6

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4

1.5

1.6x 10

-3

Rank

Pac

ket

Err

or L

oss

Rat

e

Proposed(w/ centrality)w/o centralityw/o centrality & history

(c)FIGURE 6. Average expected QoS value vs. rank (a) data rate (b)transmission delay (c) packet error loss rate

0 10 20 30 40 500

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

Number of Iterations

Util

izat

ion(

Bac

khau

l Net

wor

k)

Proposed(w/ centrality)w/o centralityw/o centrality & history

(a)

0 10 20 30 40 500.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Number of Iterations

Util

izat

ion(

Cor

e)

Proposed(w/ centrality)w/o centralityw/o centrality & history

(b)FIGURE 7. Network utilization vs. iterations (a) backhaul network (b) corenetwork

the offloading weighting factors to high values, which is fora few high-rank applications of high social impact users.Otherwise, in the algorithm without centrality and history,the QoS values according to the rank do not show a similartendency to those of other algorithms, as it shows goodperformance in data rate and transmission delay, and badperformance in packet error loss rate at the higher rank.Because they do not use the social context modeling, thisalgorithm is not affected by the rank of application.

Fig. 7 shows the network utilization of the backhaul andcore network. The proposed algorithm has the lowest value inthe backhaul network and the second lowest value in the corenetwork as shown in Fig. 7. This means that the proposedalgorithm can suitably distribute the core network’s trafficload as well as maximize the QoS of users although they maynot offload a lot of traffic to the backhaul network. Otherwise,in the algorithm without centrality, this can substantiallyalleviate the core network’s traffic load, but result in a slight

VOLUME 4, 2016 9

H. R. Cheon et al.: Social Context-Aware Mobile Data Offloading Algorithm via Small Cell Backhaul Networks

Scenario 1 Scenario 2 Scenario 30

2

4

6

8

10

12

x 104

QoS Weighting Scenarios

Dat

a R

ate

(bps

)

Proposed(w/ centrality)w/o centralityw/o centrality & history

(a)

Scenario 1 Scenario 2 Scenario 30

0.5

1

1.5

2

2.5

3

QoS Weighting Scenarios

Tra

nsm

issi

on D

elay

(se

c)

Proposed(w/ centrality)w/o centralityw/o centrality & history

(b)

Scenario 1 Scenario 2 Scenario 31

2

3

4

5

6

7

8

9

10x 10

-4

QoS Weighting Scenarios

Pac

ket

Err

or L

oss

Rat

e

Proposed(w/ centrality)w/o centralityw/o centrality & history

(c)FIGURE 8. Average expected QoS value vs. QoS weighting scenarios (a)data rate (b) transmission delay (c) packet error loss rate

QoS degradation compared to that of the proposed algorithm.In addition, in the algorithm without centrality and history,this algorithm does not have any advantages in terms ofalleviating the core network load and enhancing the QoS ofusers.

From Figs. 5– 7, it can be observed that the QoS of usersmay not be improved substantially, although the traffic ofthe high popularity application is offloaded much more. Inaddition, for the effective mobile data offloading, it shouldtake account of not only the application’s popularity but alsothe user’s social impact based on the user’s social relationshipin the same way as the proposed algorithm.

D. INFLUENCE OF QOS WEIGHTING FACTOR

In this subsection, we consider three QoS weighting sce-narios: scenario 1 for high data rate weighting (wR =0.8, wD = 0.1, wPe = 0.1), scenario 2 for high transmis-sion delay weighting (wR = 0.1, wD = 0.8, wPe = 0.1),and scenario 3 for high packet error loss rate weighting(wR = 0.1, wD = 0.1, wPe = 0.8). According to these QoSweighiting scenarios, Fig. 8 depicts the average expectedQoS value. From Fig. 8, it can be observed that the QoSweighting scenarios have the least effect on the proposedalgorithm due to exploiting the two aspects of social contextmodeling, and they have the largest effect on the algorithmwithout centrality and history due to the lack of considerationof social context.

VI. CONCLUSIONIn this paper, we have proposed the social context-awaremobile data offloading algorithm via small cell backhaulnetworks to maximize the QoS of small cell users. We modelthe social context from two perspectives, the user’s socialrelationships and the application’s popularity, and estimatethe social context-aware application selection probability.Based on this, we proposed the social context-aware trafficoffloading algorithm. The results of the performance evalu-ation demonstrate that the proposed algorithm outperformsthe other algorithms without considering social context, es-pecially in terms of data rate and transmission delay. In addi-tion, the simulation results show that the proposed algorithmcan appropriately alleviate the core network’s traffic load andmaximize the QoS of users, despite offloading less traffic tothe backhaul network.

ACKNOWLEDGMENTThis work was supported by Basic Science Research Pro-gram through the National Research Foundation of Korea(NRF) funded by the Ministry of Science, ICT and futurePlanning (NRF-2014R1A2A2A01002321).

REFERENCES[1] Cisco, “Cisco Visual Networking Index: Global mobile data traffic forecast

update, 2015-2020,” White Paper , Feb. 2016.[2] Ericsson, “Ericsson mobility report,” Nov. 2017.

10 VOLUME 4, 2016

J. H. Kim et al.: Social Context-Aware Mobile Data Offloading Algorithm via Small Cell Backhaul Networks

[3] A. Aijaz, H. Aghvami and M. Amani, “A survey on mobile data offloading:technical and business perspectives,” IEEE Wireless Commun. , vol.20,no.2, pp. 104-112, April 2013.

[4] 3GPP TR 23.829 V10.0.1, “3GPP technical specification group servicesand system aspects; local IP access and selected IP traffic offload,” Oct.2011.

[5] Small Cell Forum, “Femtocells - natural solution for offload,”SCF016.07.02, Dec. 2013.

[6] M. Cha, H. Kwak, P. Rodriguez, Y. Y. Ahn, and S.Moon, “I tube, you tube,everybody tubes: Analyzing the world’s largest user generated contentvideo system,” in Proc. the 7th ACM SIGCOMM conference on Internetmeasurement (IMC), San Diego, CA, USA, Oct. 2007, pp. 1-14.

[7] C. Canali, M. Colajanni, and R. Lancellotti, “Characteristics and evolutionof content popularity and user relations in social networks,” in Proc. TheIEEE symposium on Computers and Communications(ISCC), Riccione,Italy, Jun. 2010, pp. 750-756

[8] X. Wang, M. Chen, T. T. Kwon, L. Jin, and V. C. M. Leung, “Mobiletraffic offloading by exploiting social network services and leveragingopportunistic device-to-device sharing,” IEEE Wireless Commun., vol.21,no.3, pp. 28-36 June 2014

[9] X. Wang, Z. Sheng, S. Yang, and V. C. M. Leung, “Tag-assisted social-aware opportunistic device-to-device sharing for traffic offloading in mo-bile social networks,” IEEE Wireless Commun., vol. 23, no. 4, pp. 60-67,Aug. 2016.

[10] B. Liu, W. Zhou and J. Jiang, and K. Wang, “K-Source: Multiple sourceselection for traffic offloading in mobile social networks,” in Proc. 20168th International Conference on Wireless Communications & Signal Pro-cessing (WCSP), Yangzhou, China, Oct. 2016, pp. 1-5

[11] Y. C. Hu, M. Patel, D. Sabella, N. Sprecher, and V. Young, “Mobile EdgeComputing A key technology towards 5G,” ETSI White Paper No. 11, Sep.2015.

[12] K. Lee, J. Lee, Y. Yi, I. Rhee, and S. Chong “Mobile data offloading: howmuch can WiFi deliver?,” IEEE/ACM Trans. Netw., vol. 21, no. 2, pp. 536-550 Apr. 2013.

[13] N. Cheng, N. Lu, N. Zhang, X. Zhang, X. S. Shen, and, J. W. Mark,“Opportunistic WiFi offloading in vehicular environment: a game-theoryapproach,” IEEE Trans. Intell. Transp. Syst., vol. 17, no. 7, pp. 1944-1955,Jul. 2016.

[14] L. Ma and W. Li, “Traffic offload mechanism in EPC based on bearertype,” in Proc. 7th International Conference on Wireless Communications,Networking and Mobile Computing (WiCOM) 2011, Wuhan, China, Sep.2011.

[15] K. Samdanis, T. Taleb, and S. Schmid, “Traffic offload enhancements foreUTRAN,” IEEE Commun. Surveys Tuts., vol. 14, no. 3, pp. 884-896, 3rdQuart., 2012.

[16] F. Xia, L. Liu, J. Li, J. Ma, and A. V. Vasilakos “Socially aware network-ing: a survey,” IEEE Syst. J. , vol.9, no.3, pp. 904-921, Sep. 2015.

[17] Y. E. Sagduyu and Y. Shi, “Navigating a mobile social network,” IEEEWireless Commun., vol.22, no.5, pp. 122-128, Oct. 2015.

[18] K. Wei, X. Liang, and K. Xu, “A survey of social-aware routing protocolsin delay tolerant networks: applications, taxonomy and design-relatedissues,” IEEE Commun. Surveys Tuts., vol. 16, no. 1, pp. 556-578, pp.556-578, 1st Quart., 2014.

[19] O. Semiari, W. Saad, S. Valentin, M. Bennis, and H. V. Poor, “Context-aware small cell networks: how social metrics improve wireless resourceallocation,” IEEE Trans. Wireless Commun., vol. 14, no. 11, pp. 5927-5940, Nov. 2015.

[20] T. Stepan, J. M. Morawski, S. Dick, and J. Miller, “Incorporating spatial,temporal, and social context in recommendations for location-based socialnetworks,” IEEE Trans. Comput. Social Syst., vol. 3, no. 4, pp. 164-175,Dec. 2016.

[21] R. R. F. Lopes, “Social networks adding community-scale to context-aware connectivity management,” in Proc. 2013 IEEE Wireless Commu-nications and Networking Conference (WCNC), Shanghai, China, Apr.,2013, pp. 1615-1620

[22] F. Hou and Z. Xie, “Social-aware incentive mechanism for AP basedmobile data offloading,” IEEE Access, vol. 6, pp. 49408-49417, 2018.

[23] J. Leskovec and A. Krevl, “SNAP Datasets: Stanford large network datasetcollection,” [Online]. Available: http://snap.stanford.edu/data, Jun. 2014.

[24] NodeXL by Social Media Research Foundation, [Online]. Available:https://www.smrfoundation.org/nodexl/

[25] T. L. Griffiths and Z. Ghahramani, “The Indian buffet process: an intro-duction and review,” J. Mach. Learn. Res., vol. 12, no. 4, pp. 1185-1224,Apr. 2011.

[26] H. R. Cheon, S. Q. Lee, and J. H. Kim, “New LIPA/SIPTO offloadingalgorithm by network condition and application qos requirement,” in Proc.of 2015 International Conference on Information and CommunicationTechnology Convergence (ICTC), Oct. 2015, pp. 191-196.

[27] M. Paolini, L. Hiley, and F. Rayal, “Small-cell backhaul: industry trendsand market overview,” Senza Fili Consuling, 2013.

[28] 3GPP, “Policy and charging control architecture,” 3rd GenerationPartnership Project (3GPP), TS 23.203. [Online]. Available:http://www.3gpp.org/DynaReport/23203.htm

[29] Cisco, “Global Internet speed test (GIST) for iPhone, BlackBerry andAndroid,” [Online]. Available: http://gistdata.ciscovni.com/

[30] Y. Zhang, E. Pan, L. Song, W. Saad, Z. Dawy, and Z. Han, “Social networkaware device-to-device communication in wireless networks,” IEEE Trans.Wireless Commun., vol. 14, no. 1, pp. 177âAS190, Jan. 2015.

HYE-RIM CHEON received her B.S. degree fromDepartment of Electrical and Computer Engineer-ing, Ajou University, Suwon, Rep. of Korea in2006. She is currently pursuing her Ph.D. degreein Electrical and Computer Engineering at AjouUniversity. Her research interests include mobiledata offloading in 5G HetNet, social context-aware5G system and mobile edge computing.

JAE-HYUN KIM received the B.S., M.S., andPh.D. degrees, all in computer science and engi-neering, from Hanyang University, Ansan, Korea,in 1991, 1993, and 1996 respectively. In 1996,he was with the Communication Research Labo-ratory, Tokyo, Japan, as a Visiting Scholar. FromApril 1997 to October 1998, he was a postdoctoralfellow at the department of electrical engineer-ing, University of California, Los Angeles. FromNovember 1998 to February 2003, he worked as a

member of technical staff in Performance Modeling and QoS managementdepartment, Bell laboratories, Lucent Technologies, Holmdel, NJ. He hasbeen with the department of electrical and computer engineering, AjouUniversity, Suwon, Korea, as a professor since 2003. His research interestsinclude medium access control protocols, QoS issues, cross layer optimiza-tion for wireless communication, satellite communication, and mobile dataoffloading. He is the Center Chief of Satellite Information ConvergenceApplication Services research Center (SICAS) sponsored by Institute forInformation & Communications Technology Promotion in Korea. He isChairman of the Smart City Committee of 5G Forum in Korea since 2018.He is Executive Director of the Korea Institute of Communication and Infor-mation Sciences (KICS). He is a member of the IEEE, KICS, the Institute ofElectronics and Information Engineers, and the Korea Information ScienceSociety.

VOLUME 4, 2016 11