Social Network Aware Device-to-Device Communication ... - IEEE Xplore

16 downloads 937 Views 2MB Size Report
easily estimated. To this end, for the users within a considered. OffSN, we define an online social network (OnSN) which reflects the users' online social ties and ...
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 1, JANUARY 2015

177

Social Network Aware Device-to-Device Communication in Wireless Networks Yanru Zhang, Student Member, IEEE, Erte Pan, Student Member, IEEE, Lingyang Song, Senior Member, IEEE, Walid Saad, Member, IEEE, Zaher Dawy, Senior Member, IEEE, and Zhu Han, Fellow, IEEE

Abstract—Device-to-device (D2D) communication is seen as a major technology to overcome the imminent wireless capacity crunch and to enable new application services. In this paper, a novel social-aware approach for optimizing D2D communication by exploiting two layers, namely the social network layer and the physical wireless network layer, is proposed. In particular, the physical layer D2D network is captured via the users’ encounter histories. Subsequently, an approach, based on the so-called Indian Buffet Process, is proposed to model the distribution of contents in the users’ online social networks. Given the social relations collected by the base station, a new algorithm for optimizing the traffic offloading process in D2D communications is developed. In addition, the Chernoff bound and approximated cumulative distribution function (cdf) of the offloaded traffic are derived and the validity of the bound and cdf is proven. Simulation results based on real traces demonstrate the effectiveness of our model and show that the proposed approach can offload the network’s traffic successfully. Index Terms—Device-to-Device communication, social network, Indian buffet process.

I. I NTRODUCTION

T

HE recent proliferation of smartphones and tablets has led to the introduction of truly pervasive, anytime, anywhere wireless communications [1], [2]. The rise of online services, such as Facebook and YouTube, has significantly increased the frequency of the users’ online activities. Due to this continuously increasing demand for wireless access, a tremendous amount of data is circulating over today’s wireless networks. This increase in demand is straining current cellular systems, thus requiring novel approaches for network design [3]. Manuscript received February 10, 2014; revised June 11, 2014; accepted June 20, 2014. Date of publication July 2, 2014; date of current version January 7, 2015. This work was made possible by NPRP Grant 4-347-2-127 from the Qatar National Research Fund (a member of Qatar Foundation). The statements made herein are solely the responsibility of the authors. The associate editor coordinating the review of this paper and approving it for publication was Q. Li. Y. Zhang, E. Pan, and Z. Han are with the Department of Electrical and Computer Engineering, University of Houston, Houston, TX 77004 USA (e-mail: [email protected]; [email protected]; [email protected]). L. Song is with the School of Electrical Engineering and Computer Science, Peking University, Beijing 100871, China (e-mail: lingyang.song@ pku.edu.cn). W. Saad is with the Wireless@Virginia Tech, Bradley Department of Electrical and Computer Engineering, Blacksburg, VA 24061 USA (e-mail: walids@ vt.edu). Z. Dawy is with the Electrical and Computer Engineering Department, American University of Beirut, Beirut 1107 2020, Lebanon (e-mail: zd03@ aub.edu.lb). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TWC.2014.2334661

In order to cope with this wireless capacity crunch, device-todevice (D2D) communication underlaid over cellular systems has recently emerged as a promising technique that can significantly boost the performance of wireless networks [4]. In D2D communication, user equipments (UEs) transmit data signals to each other over a direct link instead of through the wireless infrastructure, i.e., the cellular network’s base stations (BSs). The key idea is to allow direct D2D communication over the licensed band and under the control of the cellular system’s operator [5]. Furthermore, as D2D often occurs over shorter distances, it is expected to yield higher data rates for the UEs than infrastructure-based communications. D2D communication is thus regarded as a promising technology for improving the spectral utilization of wireless systems and for introducing new applications such as proximity services or public safety applications. Recent studies have shown that the majority of the traffic in cellular pertains to the download of popular content such as videos or mobile applications [6], [7]. Given the commonality of such downloads, offloading it to the D2D tier can reduce the load on the cellular network’s infrastructure. In practice, popular contents, such as certain YouTube videos, are requested much more frequently than others. As a result, the BSs often end up serving different mobile users with the same contents using multiple duplicate transmissions. In this case, following the BS’s first transmission of the popular content, such content is now locally accessible to others users in the same area, if cellular UEs’ resource blocks (RBs) can be shared with others. Newly arriving users that are within the transmission distance can receive the “old” contents directly from those users through D2D communication. Here, the BS needs to only serve those users that request “new” content, which has never been downloaded before or not currently available from other users. Through this D2D communication, one can reduce considerable redundant requests to the BS thus alleviating its traffic load. The main contribution of this paper is to propose a novel approach to deploy reliable D2D communication mechanism that allows to exploit the social network characteristics so as to improve packet transmission and reduce the load on the network’s infrastructure. To achieve this goal, first, the suitable UEs must be chosen so as to setup the D2D communication links that can sustain the data transmission. Since D2D communication occurs between individual users, the connectivity among users can be intermittent. As a result, the communication may drop due to users’ mobility which can impact the users’ qualityof-service while also reducing the efficiency of traffic offload. In contrast, the social relations between individuals tend to be

1536-1276 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

178

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 1, JANUARY 2015

stable over time. Such social ties can be utilized to achieve a stable transmission link via D2D communication. This social relation assisted data transmission wireless network is referred to as the offline social network (OffSN). Then, we investigate the probability that a certain content is requested, which can be used to assess the amount of traffic that will eventually be offloaded via the D2D network. This probability is affected by both external (external influence from media or friends) and internal (user’s own interests) factors. While the users’ interests are difficult to predict a priori, the external influence which is based on users’ selections can be easily estimated. To this end, for the users within a considered OffSN, we define an online social network (OnSN) which reflects the users’ online social ties and influence to one another. To model the social influence in the OnSN, we introduce a novel approach based on the so-called Indian Buffet learning process. Subsequently, the learned influence is combined with practical OffSN metrics within a well-defined utility function that is then used to develop a novel D2D traffic offload algorithm. In order to evaluate the traffic offloading performance of our algorithm, we will derive the Chernoff bound on the number of old contents user selects. To make the analysis more accurate, we also derive the approximated cumulative density function (cdf) of the offloaded traffic. Our simulation results show that our model can fit the real scenarios effectively. Under certain circumstances, our algorithm can increase a considerable amount of data rate in the network. Additional simulations based on real traces further corroborate the performance gains of our proposed approach. The rest of this paper is organized as follows. Section II provides a detailed literature survey. The system model is provided in Section III for OffSN and OnSN. The traffic offloading algorithm is presented in Section IV, followed by the performance evaluation in Section V. The simulation results are shown in Section VI. Finally, conclusions are drawn in Section VII. II. L ITERATURE R EVIEW D2D communication has been recently considered as a promising method for improving frequency resource usage efficiency and for offloading cellular network traffic [1]. D2D communication shares the same frequency band as cellular communication, so proper resource allocation mechanisms are necessary [4], [5], [8]. Several recent works studied the use of D2D as a means to optimize resource usage and maintain an efficient co-existence between the D2D services and the main cellular network [4], [8]. For example, in [5], the authors proposed an auction-based approach for resource management in cellular networks with underlaid D2D communication. Based on experimental results, the work in [7] shows that more and more users access contents from their mobile devices and increase the load on the current wireless infrastructure. In [9], the authors show that D2D allows a significant reduction of traffic on the main network if the UEs within a cluster can share data and resources. The work in [10] proposes the use of network-wide clustering for increasing the reuse of wireless frequencies. This idea is similar to our work of formulating OffSNs in each cellular network. However, in [10], the authors

assume that all clusters have the same (square) size and that users are uniformly distributed in the cell. Moreover, none of these existing works exploit the social dimension to optimize the efficiency of D2D communication. In recently years, OnSN has been playing a significant role in propagation of information over the Internet [11]. In particular, recent years have witnessed a dramatic rise in the number of mobile users that are connected to social networks websites, such as Facebook and Twitter [12]. The analysis of user behavior and network externality (external influence from media, friends, etc.), constitute one of the major topics studied in OnSNs [13], [14]. Many studies have proposed the use of probabilistic modeling to analyze spread of information between users, such as video re-sharing and commenting [15], [16]. The works in [13] and [14] model the network externality to users based on the measurement of social behavior histories. In [17], the authors exploit the fact that the interest in the content that spreads over OnSNs is predictable to a certain extent. Based on practical measurement, spreading impact modeling and user profiling, it is not difficult to predict the popular trends and access patterns [18], [19]. Despite the study of contents’ externality in OnSN, the learning of social relationship hidden in the network is also critical. People grouped in small networks, such as by regions and interests, are more likely to share and transfer the same information. As shown in [18], [20], both OnSN and OffSN users show clustering properties. The online impact of OnSN users highlights the homophyly and locality effect in the OffSN. Users located at close proximity often share the same interests and content access trends as shown in [19]. So the locality and redundancy characteristics of user interests can be utilized to facilitate the traffic load balancing or content delivery [21]. The content delivery in physical wireless network has been extensively studied. In [22], the authors show that content dissemination with a small number of initial seeds can guarantee the delay requirements of all users while reducing the cellular traffic load. Other related ideas are discussed in [23, 24], but most of them focus on user mobility while ignoring the friendship relations among users. The idea of combining OnSN and OffSN together to facilitate the information transmission has been studied in [25]. However, their work does not consider the variation of users’ encounter duration or the popularity trends of online content over time. The work in [20] investigates how the joint association of nodes with interest- and locality-induced social groups can be exploited to enhance content dissemination. But [20] and [25] do not account for the variations of the content distribution in OnSN over time. In summary, on the one hand, while resource allocation in D2D has been widely studied, most existing work does not account for the role of social relationships. On the other hand, while significant work studied user behavior and patterns in social networks, little of these work exploits the content dissimilation in physical layer wireless network, as done in this paper. III. S YSTEM M ODEL Consider a cellular network with one BS and multiple users. The UEs can receive signals from the BS via the cellular

ZHANG et al.: SOCIAL NETWORK AWARE DEVICE-TO-DEVICE COMMUNICATION IN WIRELESS NETWORKS

Fig. 1.

Information dissemination in both OnSN and OffSN.

network, or from the other UEs via D2D pairs using licensed spectrum resources. In this system, two network layers exist over which information is disseminated. The first layer is the OnSN. The OnSN is the platform over which users acquire the links of contents from other users. Once a link is accessed, the data packet of contents must be transmitted to the UEs over the actual physical network. Taking advantage of the social ties, the OffSN represents the physical layer network in which the requested contents to transmit. An illustration of this proposed model is shown in Fig. 1. Each active user in the OnSN corresponds to a certain UE in the OffSN. Users access content links in an increasing order of their labels. In the OnSN, each content link is spread out according to its popularity from frequent users to regular users. In particular, a group of users, which we refer to as frequent users, have a high online activity, and, thus, are the main source of influence and information dissemination. In this respect, the choices of the regular users, who access the OnSN less frequently, are usually influenced by the frequent users. In the OffSN, the first content request is served by the BS. Subsequent users can thus be served by previous users who hold the content, if they are within the D2D communication distance. In this section, we will study the properties of OffSN and OnSN, and quantify the relationship between these two networks, before developing the proposed offloading algorithm in the next section. A. Offline Social Network Model In the area covered by a BS, the distribution of the users can often be properly modeled. For instance, in public areas such as office buildings and commercial sites, the density of users is much higher than that in other locations such as sideways and open fields. In addition, users are less likely to browse the web when they are walking or driving. Indeed, the majority of the data transmissions occurs in those fixed places such as offices or homes. In such high density locations, forming D2D networks as an OffSN becomes a natural process. Thus, we can distinguish two types of areas: highly dense areas such as office buildings, and “white” areas such as open fields. In the former, we assume that D2D networks are formed based on the users’ social relations. While in the latter, due to the low density, the

179

Fig. 2. OffSN and “White” area.

users are served directly by the BS. One brief illustration is shown in Fig. 2. The proposed approach is motivated by many real-world scenarios pertaining to wireless networks. First, in contrast to WiFi, we must stress that D2D over cellular can guarantee certain levels of quality-of-service (QoS), as it operates over the licensed spectrum [1]. With this in mind, we note that, in a practical wireless network, for highly dense areas, despite the presence of WiFi connections, the unlicensed WiFi spectrum can be highly congested, thus motivating users to seek alternative technologies, such as cellular, in order to maintain a certain desirable QoS. In such scenarios, users can have an incentive to switch to D2D over cellular, which can guarantee that they achieve their desired QoS levels, as opposed to the best effort nature of WiFi. From the telecom operators’ perspective, there are many useful cases and motivations to adopt the D2D technology as evidenced by considerable recent standardization efforts in the 3rd Generation Partnership Project (3GPP) [7]–[10]. The use of D2D for cellular operators has many advantages. From a business-case point of view, it will provide novel proximity services that are typically handled via short-range technologies such as Bluetooth or Zigbee. Such proximity services will allow operators to compete with WiFi in densely populated areas as previously discussed. D2D provides the operators with a means for offloading data traffic from their backhaul, which has become a major bottleneck with the deployment of heterogeneous networks. Moreover, another important business case is the use of D2D in public safety communications and emergency situations where infrastructure is damaged [30]. Clearly, D2D communication over cellular has more advantages over WiFi in specific locations and situations. In practice, it is well known that cellular systems can provide higher QoS guarantees during mobility, as opposed to WiFi systems. In fact, the main motivation for using D2D over licensed bands is to provide seamless coverage during highspeed mobility. Indeed, mobility becomes a major hindrance for WiFi connectivity in areas where no hotpots exist. In contrast, cellular base stations provide a more seamless and ubiquitous coverage. D2D communication over cellular uses the licensed spectrum which is owned by the network operator. Thus, D2D

180

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 1, JANUARY 2015

uncertainty in the contact length [29]. Thus, we measure the variance Ii,j of the contact period distribution to reflect the fluctuation:  (Xn − Mi,j )2 Ii,j = n . (2) Ni,j Fig. 3. Encounter history between UE i and UE j.

over cellular offers better QoS guarantees and more bandwidth than WiFi. However, we note that the model proposed by our paper can apply to D2D communication over both cellular and WiFi networks. In particular, our model is not technology specific and thus, users can still decide to use D2D over WiFi bands via protocols such as WiFi direct [51]. Our model also applies to case in which WiFi does not available at all and only cellular access is possible. The OffSN can reflect local users’ social ties. Proper metrics need to be adopted to depict the degree of the connections among users in the OffSN. Due to mobility, grouping users by using only their last known location, can be unsuitable and can lead to intermittent or dropped connectivity. Indeed, in public areas such as airports and train stations, most of the users have high mobility patterns and it is difficult to predict their future locations. In [26], the authors identify that human mobility shows a very high degree of temporal and spatial regularity, and that each individual returns to a few highly frequented locations with a significant probability. Thus, if we define the users’ connection degree in OffSN according to their encounter history or daily routes, stable D2D connections can be formed. Thus, such social ties lead to higher probabilities to successfully transmit data among users. If any two UEs are within the D2D communication distance, the BS can detect and mark them as encountered. The encounter duration between individuals follows a continuous distribution with positive value for all real values greater than zero. In the existing literatures, we see that call holding time can also exhibit this same property. There are many works that attempt to model the distribution of call holding time, in which the gamma distribution Γ(k, θ) has been shown to be an accurate model [27], [28]. k and θ are two parameters that define the shape of the distribution, which are related to the mean and variance of the variables. So we can adopt a Γ(k, θ) distribution to model the encounter duration between two users which is different from other existing literatures. To find the value for the two parameter k and θ, we need to derive the mean and variance of the contact duration first. As shown in Fig. 3, given the contact duration Xn and the number of encounters Ni,j between UE i and UE j in an OffSN (which correspond to two users i and j in the OnSN), an estimate of the expected contact duration length Mi,j can be given by:  Xn Mi,j = n . (1) Ni,j The variance represents the fluctuation in the contact period. If two UEs have the same average contact period, the one with larger fluctuations would be less preferable due to the additional

Given the mean and variance of the encounter period, we can derive the encounter duration distribution: X ∼ Γ(k, θ) = 2 Γ(Mi,j /Ii,j , Ii,j /Mi,j ). Thus, the probability density function (PDF) of encounter duration can be given by: f (x; k, θ) =

1 1 k−1 − x x e θ θk Γ(k)

(3)

∞ where Γ(k) = 0 tk−1 e−t dt. If the contact duration is not sufficient to complete a certain transmission, the communication session cannot be carried out successfully. We adopt a closeness metric, wi,j to represent the probability of establishing a successful communication period between UE i and UE j. The probability of qualified contact duration is the complementary of the probability of disqualified communication duration, so wi,j can be given by:   γ k, Xmin θ f (u; k, θ)du = 1 − Γ(k)

X min

wi,j = 1 −

(4)

0

which ranges from 0 to 1, where Xmin is the minimal contact duration required to successfully transmit one content data packet, γ(k, Xmin /θ) is the lower incomplete Gamma function which is given by 

Xmin γ k, θ

Xmin θ





=

tk−1 e−t dt.

(5)

0

Xmin depends on many factors such as the channel conditions between the two UEs (e.g., the higher the signal strength between the UEs, the smaller Xmin ), or the content size, among others. A larger closeness wi,j indicates a higher future contact opportunity between UE i and UE j. B. Online Social Network Model OnSN is the platform for content links to disseminate. The distribution of contents’ popularity in OnSN follows a certain PDF. Deriving this distribution allows to predict current users’ content selections. However, finding a closed-form expression of this distribution can be challenging. In addition, the distribution is highly varying over time as users continue to access. So if we simply assume that data are drawn from a given probability distribution. It is not suitable our case. While the nonparametric method gives us another way of estimating the distribution. The use of such a model can automatically infer an adequate distribution model from a limited data set with little complexity, thus learning the network structure [32] instead of fitting any parameterised distributions. We define the number of users in the OnSN as N which, in turn, corresponds to N UEs in the OffSN. The total number

ZHANG et al.: SOCIAL NETWORK AWARE DEVICE-TO-DEVICE COMMUNICATION IN WIRELESS NETWORKS

of available contents in the OnSN is denoted by K. Given the large volume of content available online, we can assume that K = Kh + K0 , K → ∞. Kh represents the set of contents that have viewing histories and K0 is the set of contents that do not have any such a history. We begin by analyzing the content selection problem for the case in which only one type of content is selected by each user. For content k, let πk be the probability that content k is selected by a user. We place a Beta distribution β(α/K, 1) prior on πk . Here, α/K and 1 are the parameters which determine the probability of whether the user chooses to select a content or not. Then, zn donates the selection result of user n, which follows the conjugate probability of the Beta distribution, the Bernoulli distribution [33], [34]. Thus, we have α

,1 , πk ∼ β K zn |πk ∼ Bernoulli(πk ). (6) Second, when the users are allowed to select multiple contents, the distribution over the number of selected contents follows zn |πk ∼ Binomial(πk ). Taking the limit of K → ∞, the distribution approaches P oisson(α). As the number of users increases from 1 to N , we adopt the Indian Buffet Process (IBP) [35] model which serves as a powerful analytical tool for learning the content popularity distribution and predicting users’ selections. The IBP is a stochastic process which models a restaurant problem in which each diner samples from some subset of an infinite selection of dishes on offer at a buffet. The first customer will select its preferred dishes according to a Poisson distribution with parameter α. Since all dishes are new to this customer, no reference or external information exists so as to influence the selection. However, once the first customer completes the selection, the following customers will have prior information about those dishes based on the first customer’s feedback. Therefore, the decisions of subsequent customers are influenced by the previous customers’ feedbacks. Customers learn from the previous selections to update their beliefs on the dishes and the probabilities with which they will choose the dishes. The behavior of content selection in OnSN is analogous to the dish selection in an IBP. If we view our OnSN as an Indian buffet, the online content as the infinite number of dishes, and the users as customers, we can interpret the contents spreading process online by an IBP. Users enter OnSN sequentially to request their desired content. When a user downloads its content, the viewing times of contents are changed. This action will affect the probability that this content to be requested by others. Popular contents will be requested more frequently. While those contents that are only favored by a few number of people, or those new produced content will be requested less frequently. So the probability distribution can be implemented from the IBP directly. In Fig. 4, we show one realization of an IBP. Customers are labeled sequentially in an ascending order of their labels. The shaded block represents the event that the nth user selected dish k. In IBP, the first customer selects each dish with equal

181

Fig. 4. One realization of Indian Buffet Process.

probability of α/K, and ends up with the number of dishes following P oisson(α) distribution. For the subsequent customers n = 2, . . . , N , the probability of also having dish k already /n, where mn−1 is belonging to previous customers is mn−1 k k the number of customers prior to n who select dish k [35]. Repeating the same argument as the first customer, customer n will also have m0n new dishes not tasted by the previous customers following a P oisson(α/n) distribution, which is proved in the Appendix A. The probabilities of selecting certain dishes can be used as ). For “old” dishes which have the prior information πk (mn−1 k =  0. For “new” dishes which have been tasted before, mn−1 k = 0. After user n completes its not been sampled before, mn−1 k selection, πk will be updated to πk (mnk ). This learning process is also illustrated in Fig. 4. K0n is the number of dishes that have not been sampled before user n’s selection session. We can see the selection probability for dish k updated every time after each customer’s selection. Let Z be a random binary N × K matrix indicating which contents are selected by each user, with znk = 1 if user n has content k, and 0 otherwise. The probability of any particular random binary N × K matrix Z occurs is [35]: Kh (N − mk )!(mk − 1)! α Kh · exp(−αHN ) · P (Z) = N n N! k=1 K1 ! n=1

(7) where K1n is the number of dishes being sampled by customer  n, HN is the harmonic number, HN = N j=1 (1/j), and mk is the number of times the kth dish has been selected. IV. P ROPOSED T RAFFIC O FFLOADING A LGORITHM In the previous section, we have introduced the basic model to formulate the social connection of UEs in D2D communication and predict users’ selection. In this section, we can integrate the two layer networks together and propose a new traffic offloading algorithm based on the OffSN and OnSN models. A. Data Rate in OffSN First, we model the intra-OffSN interference due to resource sharing between D2D and cellular communication. In OffSN, the D2D transmissions and the BS transmissions will interfere.

182

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 1, JANUARY 2015

During the downlink period of D2D communication, UEs will experience interference from other cellular and D2D communications as they share the same frequency band. In this respect, the received power of the link between UE i and UE j can be expressed as [5]: 2 Pi = Pr,ji = Pj · |hji |2 = Pj · d−η ji · |h0 |

(8)

where Pj is the transmit power, η is path loss exponent which ranges from 2 ≤ η ≤ 5, dji is the distance between UE j and UE i, hji is the channel response of the link between UE j and UE i, and h0 is the complex Gaussian channel coefficient that follows the complex normal distribution CN (0, 1). When a certain UE i is serviced via the cellular network, its data rate can be given by:

 PB |hBi |2 c . (9) Ri = W log2 1 +  2 j  βj  i Pj  |hj  i | + N0 Alternatively, when a UE i is being served by UE j via D2D communication with co-channel interference, the data rate will then be given by:

 Pj |hji |2 d  Ri = W log2 1 + . PB |hBi |2 + j  =j βj  i Pj  |hj  i |2 + N0 (10) Here PB , Pj , and Pj  are the transmit powers of the BS and D2D transmitter j and j  , respectively, N0 is the additive white Gaussian noise (AWGN) at the UE receivers, W is the channel bandwidth. Hereinafter, without loss of generality, we assume that W = 1. βj  i in (9) indicates the presence of interference from D2D communication to cellular communication. If there exist an interference, βj  i = 1;otherwise βj  i = 0. For βj  i in (10), j  = j and j  = i, so j  βj  i Pj  h2j  i represents the interference from the other D2D pairs that share spectrum resources with link UE j and UE i. When a D2D communication is setup, due to the uncertainty in the UEs’ movement, the transmission may fail. Thus, the expected data rate at a UE i is given by: Vid = wi,j Rid + (1 − wi,j )Ric .

(11)

When the D2D communication fail before finishing transmission, the BS will revert back to continue the service. The transmission rate of the UE i that are only served by the BS without underlying D2D is given by:   PB |hBi |2 c Vi = log2 1 + . (12) N0 B. Content Selection in OnSN In the studied model, we use IBP to model the contents distribution of each OnSN. When the UE i, whose sequence number is n in OnSN starts to surf online, this user will sample ). To this the content based on the prior information πk (mn−1 k end, this user will access old content with probability mn−1 /n, k and access new content with a P oisson(α/n) distribution. ) is updated to the posterior After content selection, πk (mn−1 k

probability πk (mnk ). The total amount of content each user selected can be drawn from a Poisson distribution: mn =

K 

znk ∼ P oisson(α).

(13)

k=1

C. System Utility From the BS perspective, the goal is to maximize the system data rate and maintaining a desirable QoS. Even though the BS can offload traffic to the D2D links, controlling the switching over cellular and D2D communication leads to additional overhead. Thus, there exists a certain cost such as control signals transmission and information feedback during the access process [5]. Therefore, for the BS that is serving a certain UE i with sequence number n, we propose the following utility function:  K h  mn−1 k Vid + m0n Ric − mn Cc (14) UB (i) = n k=1

where Cc is the overhead cost for controlling the resource allocation process. In (14), we capture the expected data rate of a BS having underlaid D2D communication links. The data rate is the sum of both D2D and cellular communication. Since the BS needs to send a control signal and receive feedback of the status D2D communication, the effective data rate must account for the rate that is incurred during the control phase. Thus, (14) quantifies the expected data rate of D2D communications when transmitting old contents, plus the cellular communication data rate when transmitting new content, minus the data rate incurred for transmitting control signals. The total traffic that the BS offloaded by D2D communication is mn Vic − mn Cc − m0n Ric . D. Proposed Algorithm In order to offload as much traffic as we can, so that we could increase the data rate of the system. Meanwhile, due to the uncertainty of user movement, we should find ways to back up the intermittent D2D communication, thus to guarantee users satisfaction. We propose a novel and robust algorithm that can offload the traffic of BS without any sacrifice on users’ satisfactions. To build up the system, we need to collect user encounter history for OffSN and trace requested contents for OnSN. The two network layers reflect the users’ activities and relationships from different aspects. The algorithm consists of multiple stages. In the first stage, The users’ closeness to one another is obtained by processing the collected encounter history, then forming the OffSN. the BS focuses on high user density areas, and collects the encounter history between users. For websites that provide a portal to access content, such as Facebook and YouTube, the BS will assign a special tag. Once a user visits such tagged websites, the BS will inspect whether the user is located in an OffSN or a “white” area. If the user is in a “white” area, its requests. If the user is located in an OffSN, the BS will wait until the user requests contents. By tracing contents that are requested by users, the system will obtain the information that constitutes the users’ activities in the OnSN

ZHANG et al.: SOCIAL NETWORK AWARE DEVICE-TO-DEVICE COMMUNICATION IN WIRELESS NETWORKS

which is captured by the IBP model. The system have the prior ) of the content distribution in the OnSN information π(mn−1 k based on previous users’ requests. As soon as an user requests data, the BS detects if there are any resources in the OffSN, and then can choose to set up D2D communication or not based on the result. For old content, the BS will send control signal to the UE j that has the highest closeness wi,j with user i. Then, UE j and UE i establish a D2D communication link. Even if the D2D communication is setup successfully, the BS still waits until the data transmission process finishes. If the D2D communication fails, the BS will revert back to serve the user directly. For new contents, the BS serves the user directly. After the selection is complete, the prior information updates to the posterior probability π(mnk ). The proposed D2D communication algorithm is summarized in Algorithm 1.

183

links and the data over cellular bands is encrypted using secure cryptographic algorithms that are much more secure than those used in other technologies such as WiFi. In practice, however, encryption does not prevent the receiving node from knowing the source node since they are forming a D2D link between one another. In this case, the receiving node will know that the source node has downloaded this content before which can raise privacy concerns. This can be dealt with by providing incentives for users to accept to share and by providing mechanisms that enable the users to selectively control what content they are willing to share data. This can be combined with state-of-the-art encryption mechanisms to avoid any malicious attacks (man-inthe-middle attacks). In addition, in our work, we assume that the operator has global information about encounter patterns and content download activities. This information can be easily protected by the operators, the same way the operator respects the privacy of data and voice logs and communications over its network nodes. The issues of privacy are important for any D2D mechanism, however, such analysis is beyond the scope of this current paper whose focus is to quantify the gains from knowing the social ties, assuming that such knowledge is handled properly via privacy-ensuring mechanisms. V. P ERFORMANCE E VALUATION To evaluate the traffic offloading performance of our algorithm, we derive a bound on the amount of traffic that can be offloaded by using Chernoff bound. In addition, we will try to find the approximated cdf of the expected traffic offloading amount by using the Skellam distribution. The definitions of Chernoff bound and Skellam distribution are listed in Appendix B. We will use the two theorems directly in the following sections. A. Chernoff Bound

Our OnSN reflects the users’ influence on one another based on their likelihood to request a similar content, as captured by the IBP, rather than based on friendships in social networking sites such as Facebook. In our model, the OnSN is established by the users’ own online activities. This implies that our approach does not require one to establish a new platform or to get data from social networking sites such as Facebook. E. Privacy Issues For any social-aware mechanism, issues related to privacy will arise. First, we note that, in wireless cellular D2D, the base station controls the activation and deactivation of D2D

Finding the amount of traffic offloaded is equivalent to finding the amount of contents that have been downloaded, and, thus, are locally accessible. Those locally accessible contents is related to both the number of total contents and new contents selected by the users. Before finding a closed-form expression on the distribution of the number of old content, we first derive a useful bound. Here, we adopt the Chernoff bound for analysis. For simplicity, we will only solve the finite-content case in which K < ∞. In our model, the expectation of total number of contents that a given user selects is mn ∼ P oisson(α). Therefore, the Chernoff bound of mn is: ⎤α ⎡ (M α −1) e (15) P {mn < M } < ⎣   M ⎦ , mn ∈ [0, α], M α

α

⎤α M −1) ( α e P {mn > M } < ⎣   M ⎦ , mn ∈ [α, ∞). M α ⎡

(16)

α

As already mentioned, the number of new contents m0n ∼ P oisson(α/n). Thus, the Chernoff bound on the number of new contents can be derived in the same way. Then, the number

184

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 1, JANUARY 2015

 Δ 0 of old contents mhn = K k=1 znk = mn − mn . Hence, define Δ the expected number of old contents μ = E{mhn } = E{mn } − E{m0n } = ((n − 1)/n)α, ∀δ > 0, with a Chernoff bound of     n−1 n α e−δ n−1 h α < P mn < (1 − δ) , (17) n (1 − δ)(1−δ) when mn ∈ [0, μ], and     n−1 n α eδ n−1 h α < P mn > (1 + δ) , (18) n (1 + δ)(1+δ) when mn ∈ [μ, ∞). As we can see, the number of old contents is the difference of two Possion distribution. The difference of two Possion distribution follows the Skellam distribution. Fig. 5.

Characterizing encounter duration by Gamma distribution.

B. Skellam Distribution By applying the pmf and cdf of the Skellam distribution into our case, we can write the pmf of the number of old contents with viewing history mhn = mn − m0n in our model as:   mh α(n+1) 2α α

n h − n 2 =e f mn ; α, n I|mhn | √ . (19) n n The cdf of the number of old content is,   α

F (M ) = P mhn  M = F M ; α, n   M  mh α(n+1) 2α n − n 2 e n I|mhn | √ = . n h

team [41]. We calculate the mean and variance of the users’ encounter durations. Then we use the Gamma distribution X ∼ 2 /Ii,j , Ii,j /Mi,j ) to model the real data. In Γ(k, θ) = Γ(Mi,j Fig. 5, we plot the PDF obtained by Gamma distribution and the real data of encounter duration together. The two curves are consistent with each other, which demonstrates the correctness and effectiveness of our method. B. Validation of IBP

(20)

mn =0

As we can see, it is impossible to find the explicit form and compute the result of (19) and (20). But we can approximate (19) and (20) using the Saddlepoint approximation. Then, we have the approximated pmf and cdf for the Skellam distribution as follows:  k D 1 e−(μ1 +μ2 )+C+D , (21) fˆ(k) =  μ2 2π(C + D)   Fˆ (M ) = P mhn  M   k  M  D 1  e−(μ1 +μ2 )+C+D = , (22) μ 2π(C +D) 2 k=0

 Δ Δ where C = (k + k 2 + 4μ1 μ2 )/2 and D = 2μ1 μ2 /(k +  k 2 + 4μ1 μ2 ). The specific procedure to get this approximation can be found in Appendix C. With the cdf function we can get the approximate number of old contents that each user select. Thus, we can estimate the traffic that can be offloaded. VI. S IMULATION R ESULTS AND A NALYSIS A. Validation of Beta Distribution In Section III-A, we use the Gamma distribution to model the user’s encounter duration. To validate the effectiveness of the Gamma distribution, we exploit a data set of sensor mote encounter records that occurred between a group of participants at the University of St. Andrews as reported by the CRAWDAD

To validate our IBP model, we adopted a data set from [48] which aims at characterizing the popularity growth patterns of online videos on YouTube. In the data set, there are three data sets that describe the popularity growth of top videos, random selected videos, and copyrighted videos. By comparing with this real data, we could validate if IBP could model user’s online behavior in Fig. 6. In the simulation, we first simulate the IBP for 10000 rounds results with 100 dishes with selection history. Then, we randomly select 100 videos that have no more than 10000 views from the data set. We sort the videos by their views in a descending order. Thus, videos with more views will have higher rank, vice versa. In Fig. 6(a) we plot the popularity pattern of videos by loglog function and in Fig. 6(b) we plot the videos selection probability. Fig. 6(a) and (b) clearly show that videos from different data sets have different popularity patterns. From subfigure (a) which is a loglog plot, we can see that the simulation data fits quite well with videos having a large viewing history videos. The section of the curve that exhibit some discrepancy pertains to videos which have a small number of views ranging from 1 to 10. Since it is a loglog figure, the differences of the four traces are accentuated. In subfigure (b), the IBP simulated trace accurately fits the real data. We note that, videos whose viewing history is small, will not have a major impact in the traffic offloading performance, as opposed to videos with a large viewing history. In general, the four different lines in both figures follows the same pattern. Under the same rank, a video simulated by IBP has close views and selecting probability as the three real data sets. Thus, we can conclude that by using IBP to model user’s selection probability on contents is correct and effective.

ZHANG et al.: SOCIAL NETWORK AWARE DEVICE-TO-DEVICE COMMUNICATION IN WIRELESS NETWORKS

Fig. 6. Characterizing online selection by IBP. (a) Video rank and views. (b) Video rank and selection probability.

C. Validation of the Proposed Traffic Offloading Algorithm To evaluate the performance of our algorithm, we need the online social ties and offline encounter history of a group of users. Here the data set by the CRAWDAD team still satisfies our needs. In the first data set, they deployed 27 T-mote invent devices over a period of 79 days among 22 undergraduate students, 3 postgraduate students, and 2 members of staff in the Department of Computer Science building. This data set helps us to establish our physical layer OffSN. In the second data set, they collected the participants’ Facebook friend lists to generate a social network topology. With those information we can generate the corresponding OnSN. Then, we adopt the IBP to generate users’ selections online under the assumption that the size of content library is unbounded. We assume that the content selection process has already been performed for a number of times. Thus, the BS can obtain the prior information of the content distribution. The physical wireless network parameters are set as follows. The radius of an cellular network is set up as 500 m. The noise spectral density is −174 dBm/Hz, which is the environmental noise power spectral density for LTE simulation when the

185

temperature T = 290 ◦ K [4], [44]. Noise figure at each device is 9 dB. The antenna gains and transmit power of BS is 14 dBi and 46 dBm. For each device, the antenna gain and transmit power are set to 0 dBi and 23 dBm, respectively. For simplicity, we fixed the content size for all UEs. The content size is captured by the length of time it needs to be transmitted. In this respect, for the encounter duration, the main parameter of interest is Xmin , the length of encounter duration required to setup a D2D communication. In practice, the content transmission time can vary from a few seconds to a few minutes. If we set a large threshold for Xmin , D2D communication will be successful with a high probability. Such a strong connection implies that only a few users would be qualified to act as content providers. Thus, the probability to find a content provider UE and setup a D2D communication will be small. In this case, the base station will have to serve users directly which will reduce the effectiveness of traffic offload. In contrast, if we choose a small value for Xmin , more users will be eligible to act as content providers, however, the transmission process becomes more susceptible to the users’ mobility. Thus, it is quite important to find the tradeoff the between number of users serving as content providers and the probability of successful D2D communication. Based on the experimental studies on device encounter duration in [25], we observe that when one user is fixed while another is passing by, the connection time of a Bluetooth transmission is 14 seconds. For modern-day devices, the encounter duration in such a scenario will naturally be longer due to the improved performance of contemporary smartphones. When content size is fixed, within a given D2D communication distance, a longer connection time will imply a higher successful transmission probability. Thus, in order to guarantee a high probability of successful D2D communication, we propose to setup D2D communication only between two relatively static users (both static or both moving in parallel) or two users such that one is fixed while the other is moving. We will not account the case in which two users are moving in different directions since such a case is not suitable for sustaining D2D transmission (due to the very short encounter duration). Thus, we choose Xmin = 20 seconds. However, we note that the proposed model and algorithm can accommodate any other values for Xmin . We present four algorithms to set up the D2D communication, called Maximal Closeness (proposed in Algorithm 1), Minimal Distance, Maximal RSSI, and Random. Minimal Distance, Maximal RSSI, and Random are extensions of Algorithm 1 with some minor changes. For the Maximal Closeness, when the BS starts to set up the D2D communication, it will choose the UE j that is within the communication range of content requester i that has the maximal closeness wi,j . In Minimal Distance, the BS will choose the UE j that is the nearest (geographically) to the content requester. In the Maximal RSSI algorithm, the BS will choose UE j which has the maximal received signal strength indicator (RSSI) value than other neighboring UEs, regardless of its closeness wi,j . This algorithm is in line with significant research works that use the RSSI as a parameter in their localization algorithms [45], [46]. This allows us to investigate how the performance relates to

186

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 1, JANUARY 2015

Fig. 7. The impact of the parameter α on data rate of the network. Fig. 9. varies.

Fig. 8. The relationship between data rate of the network and maximum D2D communication distance.

RSSI or the UE’s location. Lastly, in the Random Algorithm, BS will just randomly choose an UE that has the content and within D2D communication distance. In Fig. 7, we investigate whether the user’s online activity degree will affect the data rate of the network. As we can see from Fig. 7, with the increase of parameter α, the data rate of the network increases. Indeed, when the users tend to make more selections, more data must be transmitted. Moreover, when α increases, users are more tempt to select new contents each round. Thus will cause more traffic to the BS. The traffic includes not only the useful data, but also the control signals the BS needs to send for the D2D communication arrangement. Fig. 7 shows that the proposed approach yields a significant improvement in terms of the sum-rate. For all α, compared to the no D2D underlaid cellular network, the D2D underlaid network could double the system’s data rate. Fig. 8 shows the variations of the data rate as the maximum D2D communication distance varies. Clearly, as the D2D communication distance increases, system will have more possibilities to detect available content providers in the larger area. As a result, more traffic of the cellular network can be offloaded, as more D2D transmissions can be set up. However, under this assumption, the data rate by D2D communication will decrease due to the increase of distance. While the data rate when serving by BS will keep the same as the network

Average data rate of the network, as the cost for control signaling

size is fixed. This assumption is proved in Fig. 8. In addition, we could also see that, when the maximal D2D transmission distance increases, the performance of Minimal Distance could outperform the Maximal Closeness scheme. This is due to the fact that, if the maximal wi,j UE j is located far away, Rid will decrease to a small value. Compare to the closest UE with large Rid , the expected data rate Vid in Maximal Closeness will be smaller than that in Minimal Distance. Fig. 8, within a certain distance, D2D underlaid algorithms have higher data rate than the traditional cellular network. However, if we keep increasing the maximal distance for D2D communication, it will results in worse performance than the cellular network. This results from that UE cannot have the same transmission power and antenna gain as a BS. We should also notice that, with the increasing of the D2D transmission distance, the associated UE costs (e.g., power consumption) will also increase. Thus, the increase of D2D communication distance will provide additional benefits to the system by offloading more traffic, but not for users. In Fig. 9, we show the variation of the data rate of the network control signaling overhead cost varies. This cost captures the signaling and overhead needed for a BS to switch between different modes and to control the D2D links. Naturally, this cost impacts the overall traffic offloading performance of the system. In our simulation, we define the cost as the deduction to the gain in data rate from 5% to 50%. As we can see that, if we continue to increase the cost on managing D2D communication, the D2D communication will lose its competition with traditional cellular network. Thus, it is very important to control the management cost on D2D communication before it can be widely adopted. Fig. 10 shows the variation of the data rate of the network when the network’s size varies, while the transmit powers and maximal D2D distances remain fixed. As we can see from Fig. 10, a larger network size leads to a lower data rate. In particular, Fig. 10 shows that, the proposed algorithm is outperformed by the other schemes. However, as the network size increases, the Maximal Closeness gradually outperforms all the other approaches. This result is due to the fact that, for small networks, the proximity of the users to the BS implies

ZHANG et al.: SOCIAL NETWORK AWARE DEVICE-TO-DEVICE COMMUNICATION IN WIRELESS NETWORKS

Fig. 10. Average data rate of the network, as the size of cellular network size varies.

187

at the number of 15. Then we simulate the user’s selection and plot the empirical cdf. The simulated empirical cdf also lies between the upper and lower Chernoff bound as we expected. In addition, The approximated cdf line is quite close to the simulated empirical cdf line, which proves our analysis in the previous section. The Chernoff bound is a widely adopted method to find upper and lower bounds of sums of independent random variables. In particular, it provides us with exact minimum and maximum values for the amount of traffic that will be offloaded. The Saddlepoint approximation has higher accuracy than Chernoff bound. However, the Saddlepoint approximation does not provide an error estimation as in the case of the Chernoff bound. In practice, we not only want to have a closed-form expression of the amount of traffic that can be offloaded, but we would also want to know the specific range of our value. In this respect, the Chernoff bound and the Saddlepoint approximation nice complement one another for the studied problem. VII. C ONCLUSION

Fig. 11. The Chernoff bound, approximated cdf and empirical cdf.

that exploiting D2D communication may not bring in any significant benefits. Indeed, as seen from (11), the expected data rate will be dominated by Ric when the network size is small. Rid will decrease due to the increasing interference from BS. Thus, the maximal closeness wi,j will lead to lower expected data rate. But as the size of cellular network increases, the transmission rate when serving by BS decreases due to the increasing distance. Thus, the Rid will dominate the expected data rate and the Maximal Closeness becomes better than the other three algorithms. In most cases, the Maximal Closeness outperforms the other three algorithms. The Minimal Distance comes by the second, followed by the Maximal RSSI and Random. But their performance also depends on many other factor, such as physical layer conditions. Like the size of cellular network, the transmission power of BS and UEs, and etc. Fig. 11 shows the Chernoff bound and Saddlepoint approximation of the cdf of the number of old content user selected in the previous section. Here, we set α = 20, then plot the Chernoff bound and Saddlepoint approximation of the cdf of the 4th user’s number of old contents. Fig. 11 clearly shows that the approximated cdf lies between the Chernoff bound. By what we have mentioned in the previous section, the mean value of the number of old contents is E{mhn } = ((n − 1)/n)α = 15. Thus, there is a gap between the upper bound and lower bound

In this paper, we have proposed a novel approach for improving the performance of D2D communication underlaid over a cellular system, by exploiting the social ties and influence among individuals. We have established the OnSN to analyse the OffSN users’ online activities. By modeling the influence among users on contents selection online using the Indian Buffet Process, we have obtained the distribution of contents requests, and thus can get the probabilities of each contents to be requested. We have shown that, using the proposed algorithm, the data rate of the system has been increased. Simulation results based on real traces have proved the effectiveness of our model and have shown that the overall performance depends on a variety of network and user parameters. A PPENDIX A In Section III, we have mentioned that the nth customer will have m0n new dishes not tasted by the previous customers follows a P oisson(α/n) distribution. To prove the statement, we state a theorem that will be used later. Theorem 1: If n · p = α is fixed as n approaches infinity, then Binomial(k; n, p) is approximated by P oisson(k; α). With Theorem 3, we can prove the following statement in Section III. Proof: In the finite case of IBP where K < ∞, we assume the previous n − 1 customers have tasted m dishes, and the probability P for the nth customer having m0n new dishes is:   K − m m0n 0 p (1 − p)K−m−mn P = m0n   ∼ Binomial m0n ; K − m, p , (23) where p is the probability that each dish is firstly selected by nth customer. Therefore, 1 (πk )1 (1 − πk )n−1 P (πk )dπk

p= 0

(24)

188

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 1, JANUARY 2015

(α/K)−1

where P (πk ) = (α/K)πk is the prior distribution conforms to Beta(α/K, 1). Then, we have 1 (1 − πk )n−1

p=

α α (πk ) K dπk . K

(25)

0

For the infinity case when K → ∞, K − m can be regarded as K. Thus,     Binomial m0k ; K − m, p ≈ Binomial m0k ; K, p . (26)

where k = n1 − n2 , μ1 and μ2 are the mean value for the two Poisson distributions. I|k| (·) is the modified Bessel function of the first kind. The cumulative distribution function (cdf) of the Skellam distribution is, F (K) = P {k  K} = F (K; μ1 , μ2 )   k2 K  μ1 √ e−(μ1 +μ2 ) I|k| (2 μ1 μ2 ). = μ2

Then the expected value of the selected number of new dished will be 1

1 ≈

α

(1 − πk )n−1 α(πk ) K dπk

μ =p · K = 0

α(1 − πk )n−1 dπk =

α . n

(27)

0

Since K → ∞, p · K is fixed to α/n, by Theorem 3 we have recalled previously,   α

Binomial m0k ; K, p ≈ P oisson m0k ; . (28) n 

Hence, the statement has been proved. A PPENDIX B

Theorem 2: Let X1 , . . . , X n be a sequence of independent trials with P (Xi ) = pi , X = ni=1 Xi , and μ = E[X]. Then: For any δ > 0, there is a bound when X ∈ [0, μ] μ  e−δ , (29) P {X < (1 − δ)μ} < (1 − δ)(1−δ) and a bound when X ∈ [μ, ∞)



eδ P {X > (1 + δ)μ} < (1 + δ)(1+δ)

μ

⎤μ k −1 ) ( ⎥ ⎢e μ P {X > k} < ⎣ k ⎦ , X ∈ [μ, ∞). μ ⎡

To approximate (19) and (20) using the Saddlepoint approximation [43], there are two functions that we need to use as follows: Definition  ∞1: Moment Generating Function (MGF): M (s) = E(esx ) = −∞ esx f (x)dx, for some s that makes the integral converge. Definition 2: Cumulant Generating Function (CGF): K(s) = ln M (s). For discrete integer-valued random variable X, the Saddlepoint approximation of its pmf f (k), based on its CGF, K(s) can be written as: fˆ(k) = 

Theorem 3: The difference n1 − n2 of two statistically independent random variables N1 and N2 each having Poisson distributions with different expected values μ1 and μ2 has a discrete probability distribution which follows the Skellam distribution which the pmf is written as:   k2 μ1 √ I|k| (2 μ1 μ2 ), (33) f (k; μ1 , μ2 ) = e−(μ1 +μ2 ) μ2

exp {K(ˆ s) − sˆk} ,

(35)

(36) (37)

Substituting es into K  (s) and K  (s), we have  2μ μ k + k 2 + 4μ1 μ2   1 2 K (ˆ − s) = , 2 k + k 2 + 4μ1 μ2 Δ

= C − D, (38) 2 2μ μ k + k + 4μ1 μ2  1 2 + s) = , K  (ˆ 2 k + k 2 + 4μ1 μ2 Δ

k μ

2πK  (ˆ s)

μ1 esˆ − μ2 e−ˆs = k,  k + k 2 + 4μ1 μ2 esˆ = . 2μ1

= C + D, (32)

1

s) = k. where K  (ˆ s −s For the Skellam distribution, M (s) = e−(μ1 +μ2 )+μ1 e +μ2 e and K(s) = −(μ1 + μ2 ) + μ1 es + μ2 e−s . Hence, K  (s) = s) = k μ1 es − μ2 e−s and K  (s) = μ1 es + μ2 e−s . Apply K  (ˆ we have,

(30)

Substituting δ = 1 − (k/μ) into (15) and δ = (k/μ) − 1 into (16) we get ⎡ ⎤μ k −1) ( μ ⎢e ⎥ P {X < k} < ⎣ k ⎦ , X ∈ [0, μ], (31) k μ

A PPENDIX C

μ .

(34)

k=0

(39)

 Δ Δ where C = (k + k 2 + 4μ1 μ2 )/2 and D = 2μ1 μ2 /(k +  k 2 + 4μ1 μ2 ). Therefore, we have the approximated pmf and cdf for the Skellam distribution as follows:  k D 1 −(μ1 +μ2 )+C+D ˆ e , (40) f (k) =  μ 2π(C + D) 2 Fˆ (a) = P {X  a}  k  a  D 1 −(μ1 +μ2 )+C+D  = e . (41) μ2 2π(C + D) k=0

ZHANG et al.: SOCIAL NETWORK AWARE DEVICE-TO-DEVICE COMMUNICATION IN WIRELESS NETWORKS

R EFERENCES [1] L. Song, D. Niyato, Z. Han, and E. Hossain, Wireless Device-to-Device Communications and Networks. Cambridge, U.K.: Cambridge Univ. Press. [2] “Cisco visual networking index: Global mobile data traffic forecast update,” Whitepaper, vol. 10, no. 8, pp. 2752–2763, Feb. 2011. [3] T. Q. S. Quek, G. de la Roche, I. Guvenc, and M. Kountouris, Small Cell Networks: Deployment, PHY Techniques,and Resource Management. Cambridge: Cambridge Univ. Press, Apr. 2013. [4] C. H. Yu, K. Doppler, C. B. Ribeiro, and O. Tirkkonen, “Resource sharing optimization for device-to-device communication underlaying cellular networks,” IEEE Trans. Wireless Commun., vol. 10, no. 8, pp. 2752–2763, Aug. 2011. [5] C. Xu, L. Song, Z. Han, D. Li, and B. Jiao, “Resource allocation using a reverse iterative combinatorial auction for device-to-device underlay cellular networks,” in Proc. IEEE GLOBECOM, Anaheim, CA, USA, Dec. 3–7, 2012, pp. 4542–4547. [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 content video system,” in Proc. 7th ACM SIGCOMM IMC, San Diego, CA, USA, Oct. 2007, pp. 1–14. [7] B. Xing, K. Seada, and N. Venkatasubramanian, “An experimental study on Wi-Fi Ad-Hoc mode for mobile device-to-device video delivery,” in Proc. IEEE INFOCOM, Rio de Janeiro, Brazil, Apr. 20–25, 2009, pp. 1–6. [8] M. Zulhasnine, C. Huang, and A. Srinivasan, “Efficient resource allocation for device-to-device communication underlaying LTE network,” in Proc. 6th IEEE Int. Conf. WiMob, Niagara Falls, ON, Canada, Oct. 11–13, 2010, pp. 368–375. [9] T. Koskela, S. Hakola, T. Chen, and J. Lehtomäki, “Clustering concept using device-to-device communication in cellular system,” in Proc. IEEE WCNC, Sydney, NSW, Australia, Apr. 18–21, 2010, pp. 1–6. [10] N. Golrezaei, A. F. Molisch, and A. G. Dimakis, “Base-station assisted device-to-device communication for high-throughput wireless video networks,” in Proc. IEEE ICC, Ottawa, ON, Canada, Jun. 10–15, 2012, pp. 7077–7081. [11] S. Wu, J. M. Hofman, W. A. Mason, and D. J. Watts, “Who says what to whom on twitter,” in Proc. 20th Int. Conf. WWW, Hyderabad, India, 2011, pp. 705–714. [12] H. Kwak, C. Lee, H. Park, and S. Moon, “What is twitter, a social network or a news media?” in Proc. 19th Int. Conf. WWW, Raleigh, NC, USA, Apr. 26–30, 2010, pp. 591–600. [13] A. Goyal, F. Bonchi, and L. Lakshmanan, “Learning inuence probabilities in social networks,” in Proc. 3rd ACM Int. Conf. WSDM, New York, NY, USA, Feb. 3–6, 2010, pp. 241–250. [14] W. Chen, Y. Wang, and S. Yang, “Efficient inuence maximization in social networks,” in Proc. 15th ACM SIGKDD Int. Conf. KDD, Paris, France, Jun. 28–Jul. 1, 2009, pp. 199–208. [15] R. Xiang, J. Neville, and M. Rogati, “Modeling relationship strengthen online social networks,” in Proc. 19th Int. Conf. WWW, Raleigh, NC, USA, Apr. 26–30, 2010, pp. 981–990. [16] M. D. Choudhurry, W. A. Mason, J. M. Hofman, and D. J. Watts, “Inferring relevant social networks from interpersonal communication,” in Proc. 19th Int. Conf. WWW, Raleigh, NC, USA, Apr. 26–30, 2010, pp. 301–310. [17] F. Malandrino, M. Kurant, A. Markopoulou, C. Westphal, and U. C. Kozat, “Proactive seeding for information cascades in cellular networks,” in Proc. IEEE INFOCOM, Orlando, FL, USA, Mar. 25–30, 2012, pp. 1719–1727. [18] M. P. Wittie, V. Pejovic, L. Deek, K. C. Almeroth, and B. Y. Zhao, “Exploiting locality of interest in online social networks,” in Proc. 6th Int. CoNEXT, Philadelphia, PA, USA, Nov. 30–Dec. 3, 2010, p. 25. [19] T. Rodrigues, F. Benvenuto, M. Cha, K. Gummadi, and V. Almeida, “On word-of-mouth based discovery of the web,” in Proc. ACM SIGCOMM Conf. IMC, Berlin, Germany, Nov. 2–4, 2011, pp. 381–396. [20] E. Jaho and I. Stavrakakis, “Joint interest- and locality-aware content dissemination in social networks,” in Proc. IEEE 6th Int. Conf. WONS, Snowbird, UT, USA, Feb. 2–4, 2009, pp. 173–180. [21] S. Scellato, C. Mascolo, M. Musolesi, and J. Crowcroft, “Track Globally, Deliver Locally: Improving Content Delivery Networks by Tracking Geographic Social Cascades,” in Proc. 20th Int. Conf. WWW, Hyderabad, India, Mar. 28–Apr. 1, 2011, pp. 457–466. [22] X. Wang, M. Chen, Z. Han, T. Kwon, and Y. Choi, “Content dissemination by pushing and sharing in mobile cellular networks: An analytical study,” in Proc. IEEE 9th Int. Conf. MASS, Las Vegas, NV, USA, Oct. 2012, pp. 353–361.

189

[23] Y. Li et al., “Energy-efficient optimal opportunistic forwarding for delaytolerant networks,” IEEE Trans. Veh. Technol., vol. 59, no. 9, pp. 4500– 4512, Nov. 2010. [24] S. Ioannidis, A. Chaintreau, and L. Massoulie, “Optimal and scalable distribution of content updates over a mobile social network,” in Proc. 28th IEEE INFOCOM, Rio de Janeiro, Brazil, Apr. 19–25, 2009, pp. 1422–1430. [25] B. Han et al., “Mobile data offloading through opportunistic communications and social participation,” IEEE Trans. Mobile Comput., vol. 11, no. 5, pp. 821–834, May 2012. [26] M. C. González, C. A. Hidalgo, and A. Barabási, “Understanding individual human mobility patterns,” Nature, vol. 453, no. 7196, pp. 779–782, Jun. 2008. [27] J. Guo, F. Liu, and Z. Zhu, “Estimate the call duration distribution parameters in GSM system based on K-L divergence method,” in Proc. IEEE Int. Conf. WiCom, Shanghai, China, Sep. 21–25, 2007, pp. 2988–2991. [28] M. M. Alwakeel and V. A. Aalo, “A teletraffic performance study of mobile LEO-Satellite cellular networks with gamma distributed call duration,” IEEE Trans. Veh. Technol., vol. 55, no. 2, pp. 583–596, Mar. 2006. [29] F. Li and J. Wu, “LocalCom: A community-based epidemic forwarding scheme in disruption-tolerant networks,” in Proc. IEEE SECON, Rome, Italy, Jun. 22–26, 2009, pp. 574–582. [30] E. Yaacoub, “On the use of device-to-device communications for QoS and data rate enhancement in LTE public safety networks,” presented at the IEEE WCNC—Workshop Device-to-Device Public Safety Communications, Istanbul, Turkey, Apr. 6–9, 2014. [31] Qualcomm LTE Direct, LTE Direct: Research and Use Cases 2012. [Online]. Available: URL=http://www.qualcomm.com/about/research/ projects/lte-direct [32] F. Wood, T. L. Griffiths, and Z. Ghahramani, “A non-parametric bayesian method for inferring hidden causes,” in Proc. 22th Conf. Uncertainty Artif. Intell., Cambridge, MA, USA, Jul. 13–16, 2006, vol. 22, pp. 536–543. [33] Z. Ghahramani, T. L. Griffiths, and P. Sollich, “Bayesian nonparametric latent feature models,” in Proc. Valencia/ISBA 8th World Meet. Bayesian Statist. Benidorm, Alicante, Spain, Jun. 1–6, 2006, pp. 1–19. [34] Y. Teh, D. Görür, and Z. Ghahramani, “Stick-breaking construction for the Indian buffet process,” in Proc. 11th Int. Conf. AISTAT, San Juan, Puerto Rico, Mar. 21–24, 2007, pp. 1–8. [35] T. L. Griffiths and Z. Ghahramani, “The Indian buffet process: An introduction and review,” J. Mach. Learn. Res., vol. 12, no. 4, pp. 1185–1224, Apr. 2011. [36] Y. W. Teh, M. I. Jordan, M. J. Beal, and D. M. Blei, “Hierarchical dirichlet processes,” J. Amer. Statist. Assoc., vol. 101, no. 476, pp. 1566–1581, Dec. 2006. [37] T. Karagiannis, J. Y. Le Boudec, and M. Vojnovic, “Power law and exponential decay of inter contact times between mobile devices,” in Proc. 13th Annu. ACM Int. Conf. MobiCom, Montreal, QC, Canada, Sep. 9–14, 2007, pp. 183–194. [38] W. J. Hsu, T. Spyropoulos, K. Psounis, and A. Helmy, “Modeling timevariant user mobility in wireless mobile networks,” in Proc. 26th IEEE INFOCOM, Anchorage, AK, USA, May 6–12, 2007, pp. 758–766. [39] X. Zhang, G. Neglia, J. Kurose, and D. Towsley, “Performance modeling of epidemic routing,” Int. J. Comput. Telecommun. Netw., vol. 51, no. 10, pp. 2867–2891, Jul. 2007. [40] E. Bulut and B. K. Szymanski, “Friendship based routing in delay tolerant mobile social networks,” in Proc. IEEE GLOBECOM, Miami, FL, USA, Dec. 6–10, 2010, pp. 1–5. [41] G. Bigwood, D. Rehunathan, M. Bateman, T. Henderson, and S. Bhatti, “Exploiting self-reported social networks for routing in ubiquitous computing environments,” in Proc. IEEE Int. Conf. WiMob, Avignon, France, Oct. 12–14, 2008, pp. 484–489. [42] M. T. Irfan and L. E. Ortiz, “On influence, stable behavior, and the most influential individuals in networks: A game-theoretic approach,” Artif. Intel., vol. 215, pp. 79–119, Oct. 2014. [43] R. W. Butler, Saddlepoint Approximations With Applications. Cambridge, U.K.: Cambridge Univ. Press, 2007. [44] S. Sesia, I. Toufik, and M. Baker, LTE: The UMTS Long Term Evolution. New York, NY, USA: Wiley, 2009. [45] M. Saxena, P. Gupta, and B. Nath Jain, “Experimental analysis of RSSIbased location estimation in wireless sensor networks,” in Proc. 3rd Int. Conf. COMSWARE, Bangalore, India, Jan. 6–10, 2008, pp. 503–510. [46] H. S. AbdelSalam and S. Olariu, “Towards enhanced RSSI-based distance measurements and localization in WSNs,” in Proc. IEEE INFOCOM Workshops, Rio de Janeiro, Brazil, Apr. 19–25, 2009, pp. 1–2. [47] A. T. Parameswaran, M. I. Husain, and S. Upadhyaya, “Is RSSI a reliable parameter in sensor localization algorithms—An experimental study,” in Proc. 28th Int. SRDS, Niagara Falls, NY, USA, Sep. 27–30, 2009, pp. 1–5.

190

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 1, JANUARY 2015

[48] F. Figueiredo, F. Benevenuto, and J. Almeida, “The tube over time: Characterizing popularity growth of YouTube videos,” in Proc. 4th ACM Int. Conf. WSDM, Kowloon, Hong Kong, Feb. 9–12, 2011, pp. 1–5. [49] M. Cha, H. Kwak, P. Rodriguez, Y. Ahn, and S. Moon, “Analyzing the video popularity characteristics of large-scale user generated content systems,” IEEE/ACM Trans. Netw., vol. 17, no. 5, pp. 1357–1370, Oct. 2009. [50] M. A. Sumlinski, “International reserves-too much of a Zipf’s thing,” in IMF Working Papers, Nov. 2008, pp. 1–121. [51] D. Camps-Mur, A. Garcia-Saavedra, and P. Serrano, “Device-to-device communications with Wi-Fi direct: Overview and experimentation,” IEEE Wireless Commun., vol. 20, no. 3, pp. 96–104, Jun. 2013. [52] L. Sankar, S. R. Rajagopalan, and H. V. Poor, “A theory of utility and privacy of data sources,” in Proc. IEEE Int. Symp. Inf. Theory, Austin, TX, USA, Jun. 10–14, 2010.

Yanru Zhang (S’13) received the B.S. degree in electronic engineering from the University of Electronic Science and Technology of China, in 2012. She is currently working toward the Ph.D. degree with the Department of Electrical and Computer Engineering, University of Houston, Houston, TX, USA. Her current research interests include deviceto-device communications, social networks, contract theory, and matching theory. Ms. Yanru has been a Reviewer for IEEE TWC, GLOBECOM, ICC, WCNC, etc.

Erte Pan (S’14) received the B.E. degree in electrical and computer engineering from Wuhan University, China, in 2010. He is currently working toward the Ph.D. degree in the Department of Electrical and Computer Engineering, University of Houston, Houston, TX, USA. Since June 2013, he has been a Research Assistant with the Wireless Networking, Signal Processing, and Security Laboratory, University of Houston. His research interests include nonparametric inference, deep learning networks, big data analysis, and sublinear methods on smart grid networks.

Lingyang Song (S’03–M’06–SM’12) worked as a Postdoctoral Research Fellow with the University of Oslo, Oslo, Norway and Harvard University, until rejoining Philips Research U.K. in March 2008. Since May 2009, he has been a Full Professor with the School of Electronics Engineering and Computer Science, Peking University, Beijing, China. His main research interests include multiple-input multipleoutput systems, orthogonal frequency-division multiplexing, cooperative communications, cognitive radio, physical-layer security, game theory, and wireless ad hoc/sensor networks. Dr. Song is currently on the Editorial Board of IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS and of IET Communications and Journal of Network and Computer Applications. He received K. M. Stott Prize for excellent research during his doctoral years. He also received the Best Paper Awards at the IEEE International Conference on Wireless Communications, Networking and Mobile Computing (WiCOM 2007), at the First IEEE International Conference on Communications in China (ICCC 2012), at the IEEE Wireless Communication and Networking Conference (WCNC2012), at the International Conference on Wireless Communications and Signal Processing (WCSP 2012), and at the International Conference on Communications (ICC 2014); the Best Student Paper Award at the Seventh International Conference on Communications and Networking in China (ChinaCom2012); and the 2012 IEEE Asia Pacific Young Researcher Award.

Walid Saad (S’07–M’10) received the B.E. degree in computer and communications engineering from the Lebanese University, Beirut, Lebanon, in 2004; the M.E. degree in computer and communications engineering from the American University of Beirut, Beirut, in 2007; and the Ph.D. degree from the University of Oslo, Oslo, Norway, in 2010. Since August 2014, he has been an Assistant Professor with the Department of Electrical and Computer Engineering, Virginia Polytechnic Institute and State University (Virginia Tech), Blacksburg, VA, USA. Prior to joining Virginia Tech, he was an Assistant Professor with the Department of Electrical and Computer Engineering, University of Miami, Coral Gables, FL, USA, and has held several research positions at institutions such as Princeton University, Princeton, NJ, USA, and the University of Illinois at Urbana-Champaign, Urbana, IL, USA. He is the author or coauthor of one book and over 95 international conference and journal publications in these areas. His research interests include wireless and small cell networks, game theory, smart grid, network science, cognitive radio, wireless security, and self-organizing networks. Dr. Saad received the U.S. National Science Foundation CAREER Award in 2013. He was the author/coauthor of the papers that received the Best Paper Award at the Seventh International Symposium on Modeling and Optimization in Mobile, Ad Hoc and Wireless Networks (WiOpt), in June 2009, at the Fifth International Conference on Internet Monitoring and Protection (ICIMP) in May 2010, and at the IEEE Wireless Communications and Networking Conference (WCNC) in 2012.

Zaher Dawy (S’97–M’04–SM’09) received the B.E. degree in computer and communications engineering from the American University of Beirut (AUB), Beirut, Lebanon, in 1998 and the M.E. and Dr.Ing. degrees in communications engineering from Munich University of Technology (TUM), Munich, Germany, in 2000 and 2004, respectively. Since September 2004, he has been with the Department of Electrical and Computer Engineering, AUB, where he is currently an Associate Professor. His research and teaching interests include wireless communications, cellular technologies, context-aware mobile computing, mobile solutions for smart cities, computational genomics, bioinformatics, and mobile health. Dr. Dawy serves as an Associate Editor for the IEEE C OMMUNICATIONS S UR VEYS AND T UTORIALS , as the Editor for Elsevier Physical Communications and Elsevier Pervasive and Mobile Computing, and as the Executive Editor for Wiley Transactions on Emerging Telecommunications Technologies. He received the Abdul Hameed Shoman Award for Young Arab Researchers in 2012, the IEEE Communications Society 2011 Outstanding Young Researcher Award in Europe, Middle East, and Africa Region, the AUB Teaching Excellence Award in 2008, the Best Graduate Award from TUM in 2000, the Youth and Knowledge Siemens Scholarship for Distinguished Students in 1999, and the Distinguished Graduate Medal of Excellence from Hariri Foundation in 1998.

Zhu Han (S’01–M’04–SM’09–F’14) received the B.S. degree in electronic engineering from Tsinghua University, Beijing, China, in 1997 and the M.S. and Ph.D. degrees in electrical engineering from the University of Maryland, College Park, MD, USA,in 1999 and 2003, respectively. From 2000 to 2002, he was an R&D Engineer with JDSU, Germantown, MD. From 2003 to 2006, he was a Research Associate with the University of Maryland. From 2006 to 2008, he was an Assistant Professor with Boise State University, Idaho. He is currently an Associate Professor with the Department of Electrical and Computer Engineering, University of Houston, Houston, TX, USA. His research interests include wireless resource allocation and management, wireless communications and networking, game theory, wireless multimedia, security, and smart grid communications. Dr. Han has been an Associate Editor of IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS since 2010. Dr. Han is the winner of IEEE Fred W. Ellersick Prize 2011. Dr. Han is an NSF CAREER award recipient for 2010.