On social delay tolerant networking aggregation, tie detection,and routing

On Social Delay-Tolerant Networking:Aggregation, Tie Detection,

and RoutingKaimin Wei, Deze Zeng, Member, IEEE, Song Guo, Senior Member, IEEE, and Ke Xu

Abstract—Social-based routing protocols have shown their promising capability to improve the message delivery efficiency inDelay Tolerant Networks (DTNs). The efficiency greatly relies on the quality of the aggregated social graph that is determined by themetrics used to measure the strength of social connections. In this paper, we propose an improved metrics that leads to high-qualitysocial graph by taking both frequency and duration of contacts into consideration. Furthermore, to improve the performance ofsocial-based message transmission, we systematically study the community evolution problem that has been little investigated in theliteration. Distributed algorithms based on the obtained social graph are developed such that the overlapping communities andbridge nodes (i.e., connecting nodes between communities) can be dynamically detected in an evolutionary social network.Finally, we take all the results above into our social-based routing design. Extensive trace-driven simulation results show thatour routing algorithm outperforms existing social-based forwarding strategies significantly.

Index Terms—Social-aware routing, connection strength metric, social graph, community, bridge node

Ç

1 INTRODUCTION

DELAY-TOLERANT NETWORKs (DTNs) [1] emerge as anindispensable complement to the traditional wireless

network to support a variety of delay-tolerant applications.They are characterized by the lack of contemporaneousand continuous end-to-end connections, but only opportunisti-cally intermittent connections. In such networks, opportunisticrouting [2], [3], [4], [5], [6], [7], [8], [9] has attracted muchattention because its ‘‘carry-and-forward’’ data disseminationscheme well exploits the unexpected transmission opportuni-ties. As its simplest form, epidemic routing [3] allows messagesto be aggressively replicated and forwarded to each encounterednode, incurring high resource (e.g, energy, buffer, etc.)consumption. To address this issue, various improved oppor-tunistic routing schemes [4], [5] have been proposed byexploring a critical issue on how to select appropriate interme-diate nodes, i.e., relay nodes, such that messages are forwardedonly to them, other than blindly to all encountered ones.

It has been discovered that some DTNs like mobile socialnetwork (MSN) [10] exhibit human behaviors, whosebenefits for relay node selection have been validated byseveral social-based routing protocols such as BubbleRap [4],PeopleRank [5] and SimBet [11]. To exploit the socialnetwork property in these DTNs, we take a systematicalapproach by developing the following techniques that are ofgreat importance: 1) high-quality social graph aggregation,

2) dynamic detection of social ties, and 3) efficient social-aware routing, in which, furthermore, one serves as anunderlying basis for the next.

Existing routing protocols for social DTNs overlook theconstruction of high-quality social graph [12]. In particular,without accurate measurement of the connections betweennodes, some of them even simplify treat simple socialgraph as contact graph. The limitation of existing metricsfor connection strength motivates us to re-examine how todefine an improved one for high-quality social graphaggregation. We then invent a hybrid connection strengthmetrics and incorporate it into a distributed density-basedaggregation approach to build social graph. Because of hisstrong descriptive capability in distinguishing connectionstrength, even existing social-based routing protocols (e.g.,BubbleRap [4] and PeopleRank [5]) can benefit by adoptingour proposed metrics with a better performance than theiroriginal versions.

The quality of social ties also plays an essential role tothe performance of social-aware routing protocols. Existingcommunity detection methods, however, can hardly uncoveran accurate community structure of a DTN since they eitherrely on empirical threshold values [13] or ignore communityevolutions (e.g., nodes leaving from a community) [14]. Evenworse, how to identify and exploit the bridge nodes for social-based routing, especially for inter-community communica-tions, has been little studied in the literature. In contrast, wedevelop a distributed community detection approach touncover overlapping communities and capture their evolu-tions, without any predetermined empirical value. Mean-while, a self-recommending approach is proposed for bridgenode identification.

The final step is to exploit the extracted social tiesfor social-aware DTN routing. In particular, forwardingcandidate set (i.e., the messages that shall be forwarded)

. K. Wei and K. Xu are with the State Key Lab of Software DevelopmentEnvironment, Beihang University, China. E-mail: {weikaimin, kexu}@nlsde.buaa.edu.cn.

. D. Zeng and S. Guo are with the Department of Computer Science andEngineering, The University of Aizu, Japan. E-mail: [email protected].

Manuscript received 2 Apr. 2013; revised 14 Sept. 2013; accepted 2 Oct. 2013.Date of publication 17 Oct. 2013; date of current version 16 May 2014.Recommended for acceptance by Y. Xiang.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference the Digital Object Identifier below.Digital Object Identifier no. 10.1109/TPDS.2013.264

1045-9219 � 2013 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.

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 25, NO. 6, JUNE 2014 1563

and their forwarding order should be carefully selected suchthat the message delivery probability is maximized. However,this issue is always overlooked in existing social-basedrouting protocols [4], [5], [15], in which the buffered messagesare randomly chosen and then greedily forwarded node oneby one until the duration ends. To deal with this issue, we wepropose a connection-strength aware routing (CSAR) proto-col, in which the message utility is defined to determine theforwarding set and order. We further explore the benefit ofthe detected community structure and apply the inter-community multi-copy and intra-community single-copy policyin the consideration of both delivery performance andresource consumption. Extensive real-trace driven simula-tions show that CSAR achieves the delivery performanceclosing to epidemic routing and significantly outperformsexisting social-based routing protocols. For example, CSARcan achieve a higher message delivery ratio by 50 percent and65 percent, but a lower overhead by 18 percent and 29 percentthan BubbleRap and PeopleRank, respectively.

The remainder of this paper is structured as follows.Section 2 presents related work. Section 3 introducesthe system model. Section 4 proposes a new connectionmetric for social graph aggregation. Section 5 elaboratesour distributed algorithms for community and bridgedetection. Section 6 presents our CSAR protocol. Section 7evaluates and analyzes the simulation results. Finally,Section 8 concludes this paper. The specifications ofalgorithms proposed in the paper are given in thesupplementary file which is available in the ComputerSociety Digital Library at http://doi.ieeecomputersociety.org/10.1109/264.

2 RELATED WORK

2.1 Social Graph AggregationTo obtain high-quality social graph, several social graphaggregation algorithms have been developed and can begenerally categorized into time-based and density-basedclasses. BubbleRap [4] is a time-based approach whichaggregates contacts appeared in a sliding window. How-ever, it is difficult to determine the window length. Thedensity-based approach [5] adopts a connection metric thatchooses the appropriate connections to build a social graph.Compared to the time-based class, the density-based oneaggregates the desired contacts, other than simply anycontacts, into the social graph. While density-based algo-rithms achieve better performance as revealed in [12], theiradopted metrics (e.g., contact frequency), which charac-terizes the connection strength for desired contact selection,has limitation in describing the factor of contact durationthat would be critical to the performance as well.

2.2 Social Tie DetectionSocial graph describes social ties in terms of, for example,community structure [4], [16] and bridge (i.e., connectingnodes between communities) [11], [17]. On the communitydetection, Santo et al. [18] give a survey on variouscommunity detection approaches. It can be seen thatmost existing community detection approaches operatein a centralized manner, making them hard to be directlyapplied to DTNs. To address this issue, distributed

community detection algorithms were proposed recently.Hui et al. [4] use k-clique and weighted networkapproaches to detect overlapping communities. In [13], theyfurther introduce three distributed community detectionapproaches. All these approaches require a predeterminedfamiliar set threshold and fail to discover communityevolutions, and even some of them can not uncoveroverlapping communities. Later on, Li et al. [14] propose adistributed algorithm to detect overlapping communitieswithout any predetermined threshold. However, it onlyfocuses on community creation and therefore still can notdiscover any community evolutions.

2.3 Social-Aware RoutingThe potential of social ties on the guidance of DTN routing(e.g., relay node selection) has been widely explored by adiversity of social-based routing protocols. Daly et al. [11]propose SimBet which takes advantage of betweennesscentrality and similarity to make forwarding decisions.Hui et al. [4] present BubbleRap which forwards messagesto the nodes with increasingly large betweenness centralityuntil they reach destination communities. Later on,PeopleRank [5] makes use of past encounter informationto rank nodes and delivers messages to the nodes with ahigher utility value. User-Centric [19] takes the social contactpatterns and mobile user interests into relay node selection.Furthermore, existing work focuses on how to guide therouting of one message from a single unicast session.When multiple messages to different destinations coexist,due to the limitation and unexpectation of contact duration,it is impossible to guarantee that every message proceedsnormally as the guidance in routing. Therefore, it is inevitableto investigate the management of multiple messages. How-ever, this is usually overlooked in existing work.

3 SYSTEM MODEL

We consider a category of DTNs exhibiting certain socialfeatures like MSN [10], where the communication devicesare attached and moved with human beings. The trans-mission between any two mobile nodes happens only whenthey move into the reciprocal communication range of eachother. Mobile nodes evolve into communities according totheir social properties because their mobility is of long-term regularities. Such type of DTN is represented by asocial graph G ¼ ðV; EÞ, where vertex set V and undirectededge setE denote all mobile nodes and their pairwise socialconnections, respectively. A social graph is constructed byaggregating all historical contact information. To imple-ment in a distributed manner, each mobile node imaintains a local social graph Gt

i ¼ ðV ti ; E

tiÞ (V t

i � V andEti � E) over time t, in which the neighbor set of i is

denoted as Nti ðNt

i � V ti Þ.

Based on the social graph, social ties like communitystructure and bridge nodes could be detected. A commu-nity is a group of nodes with distinguishable features (e.g.,common interest) from others outside. Let Ct

i be the set ofall communities detected by node i at time t. For anycommunity C 2 Ct

i, it is denoted as C ¼ ðV tC; E

tCÞ, where

V tC � V and Et

C � E.

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 25, NO. 6, JUNE 20141564

In general, communities are not isolated but inter-connected by edges or common nodes, in which anyinvolving node is called a bridge node and can facilitateinter-community communications. Let BðC;C0Þ indicatethe set of bridge nodes connecting communities C and C0.Any node i may serve as bridge for multiple communitypairs, denoted as Bi.

It is well accepted that bridging centrality [20] is anefficient metric to measure how well a node is locatedbetween connected communities. In [21], bridging central-ity of a node i is defined as

BCi ¼ FðiÞ �YðiÞ; (1)

where FðiÞ and YðiÞ are betweenness centrality andbridging coefficient, respectively. Betweenness centralityFðiÞ is defined as

FðiÞ ¼X

s6¼i6¼t2V

�stðiÞ�st

; (2)

where �st is the number of shortest paths between nodes sand t, and �stðvÞ is the number of shortest paths passingthrough a node i out of �st. Bridging coefficient YðiÞ can becalculated as

YðiÞ ¼ 1

dðiÞX

v2Ni

�ðvÞdðvÞ � 1

; (3)

where dðiÞ is the degree of node i and �ðvÞ is the number ofedges leaving the neighbors of node v among the edgesincident to each direct neighbor v of node i. We notice thatYðiÞ is determined by the connection degree (i.e., numberof neighbors) of node i in the social graph.

4 SOCIAL GRAPH AGGREGATION

In this section, we propose a new connection strengthmetric that can accurately measure the potential transmis-sion capabilities between nodes in DTNs. Based on thismetric, a density-based distributed social graph aggrega-tion approach is applied to construct a social graph on eachmobile node.

4.1 Contact Frequency vs. Contact DurationIn density-based aggregation approach, not all but only theselected connections will be included into a social graphsuch that this social graph shall be maintained at a certaindensity to prevent the forwarding decision from degen-erating to random [12]. As a result, connection strength isused as a metric to determine whether a connection shall beadded into the social graph. It can be derived fromhistorical contact information, e.g., contact frequency andaverage contact time [12], [14]. The former, denoting thenumber that two nodes encounter in a period of time, does

well in evaluating frequent contacts of mobile nodes butwith short duration. The latter, representing the averageduration of each encounter for two nodes, is good atassessing occasional contacts with long duration. However,these metrics cannot measure contacts well when theencounter behaviors are complex.

Fig. 1 shows four contact cases observed during time 0and T , in which a box denotes the period that two nodes arewithin the communication range of each other. Supposeeach message has the same size and can be transmittedwithin t. We notice that contact frequency fails todistinguish cases (a) and (c), and average contact durationper contact cannot distinguish cases (a) and (d). Anotherpossible metric is average contact duration per time unitthat does well in differentiating case (a) from case (d), butnot from (b). Our obervation shows that single-criterionbased metric cannot accurately measure the forwardingcapacity between two nodes.

4.2 A Hybrid MetricMotivated by the facts observed from Fig. 1, we design ahybrid metric that measures connection strength by takingboth contact frequency and contact duration into consid-eration. The connection strength HMi, between nodes i andj is defined as

HMij ¼ � � ð1� e�nðT ÞÞ þ ð1� �Þ �R T

0 fðtÞdtT

; (4)

where nðT Þ is the contact number between nodes i and j in½0; T �, fðtÞ is a binary value indicating whether i and j are incontact at time t ðfðtÞ ¼ 1Þ or not ðfðtÞ ¼ 0Þ, and � 2 ½0; 1� isa tunable scaling parameter according to network feature.

Note that terms 1� e�nðT Þ andR T

0 fðtÞdt=T actuallyindicate the contact frequency and the average contactduration per time unit, respectively. By tuning parameter �according to network features, the weight of contactfrequency and average contact duration can be adjusted.We set high weight on the contact frequency for someDTNs, where the mobile nodes frequently encountereach other but with short duration. For example, datasetRollernet [22] shows that participants skate together and theyfrequently encounter and separate. However, for DTNs likeReality [23], where mobile nodes keep a long contact eachencounter, the contact duration should be weighted. In someother cases with complicated and irregular encounterpatterns, it is difficult to determine the relative importanceof contact frequency or average contact time. How the valueof � affects the forwarding performance will be furtherdiscussed by real-trace driven experiments in Section 7.

Finally, to obtain the social graph on each node, weapply our hybrid connection strength metric HM to thedistributed density-based algorithm proposed in [12].The superiority of the social graph using Eq. (4) over theones using existing metrics to the routing performanceshall be validated by experiments later.

5 DISTRIBUTED SOCIAL-TIE DETECTION

As we have known, social ties like community structureand bridge node are very helpful to relay node selection in

Fig. 1. Four different contact cases in time interval [0, T].

WEI ET AL.: DELAY TOLERANT NETWORKING: AGGREGATION, TIE DETECTION, AND ROUTING 1565

social-based routing protocols. To extract these social ties,we propose a distributed community detection algorithm(DCDA) and a bridge node self-recommending approachbased on the obtained social graph.

5.1 Dynamic Community Detection

5.1.1 Community Formation ConditionA community is usually regarded as a group of nodes withmore internal connections than external connections in asocial graph [18]. To formally quantify such relationship,we introduce intra-connection density dþC and inter-connection density d�C of a community C, which are definedas the ratios of the number of internal edges and externaledges to all possible internal edges, respectively, i.e.,

dþC ¼EþC��

jVC j jVC j � 1ð Þ=2 (5)

and

d�C ¼E�C��

jVC j jVC j � 1ð Þ=2; (6)

whereEþC andE�C denote a set of edges with both endpointsand only one endpoint in C, respectively.

In addition, the number of nodes in a community shallexceed a certain number �. For example, we usually do nottreat two nodes as a community. A recent study [18], [24]shows that the size of communities is usually not less than4 in mobile social networks. Therefore, in this paper, weassume that all communities shall be with size at least 4, i.e.,� ¼ 4. In summary, a group of nodes can be regarded as acommunity C if and only if

dþC 9 d�C and jCj � �: (7)

From (7), we can see that the community structure wouldbe affected once edge appearance or disappearance hap-pens. Therefore, the community detection algorithm shallbe invoked right after the aggregation process, by which anedge may be added into or removed from the social graph.

5.1.2 Community FormationA local social graph is always maintained and updated byeach node i during a social graph aggregation process. If anew edge, e.g., eij, is added in the social graph, the community

structure on node i shall be updated by 1) creating a newcommunity or 2) merging two existing communities, assummarized in Algorithm 1 in the supplementary fileavailable online. Illustrative examples are given in Fig. 2,where communities are shown by shaded regions.

Basic community creation: Whenever a new edge eijappears, node i always tries to build a new communityCtemp based on the common neighbors between i and j,no matter whether j is within one or multiple communities(e.g., Fig. 2b) or not (e.g., Fig. 2a). If it satisfies thecommunity formation condition (7) but is not yet formed,Ctemp will be regarded as a new community and added intothe current community set. Otherwise, the appearance ofeij does not affect the existing communities.

Community merging: Sometimes, the new communityCtemp may share substructures (i.e., overlap) with otherexisting communities, as shown in Fig. 2c. In this case, thetwo communities shall be merged into a bigger one. To thisend, after building a new community Ctemp, Algorithm 1tries to combine it with its overlapped communities if themerged one still satisfies (7). Finally, the original commu-nities should be removed from the community set.

5.1.3 Community PartitionDuring the aggregation process, an edge may also disap-pear from the social graph since a certain density shouldbe kept such that the social graph avoids becoming toodense and homogeneous [12]. The edge removal may incurthe partition of certain existing communities. However,existing distributed community detection algorithms [4],[13] simply overlook these events and fail to capture thepartition of communities, resulting in inaccuracy in thecommunity structure description. To deal with suchevents, we propose an algorithm specified in Algorithm 2as given in the supplementary file available online.

Not all edge disappearance events will result incommunity partition. For example, if edge eij does notbelong to any community as shown in Figs. 3a, 3b, 3c, itsremoval will not affect any existing community structure.The existing community structure shall be retained. On theother hand, for the cases (d) and (e) in Fig. 3, where eij iswithin a community, the remaining nodes after its deletionshould be restructured as follows. If condition (7) still holds,the current community structure shall keep unchanged

Fig. 2. Examples of community formation.

Fig. 3. Examples of community partition.


(e.g., Fig. 3d). Otherwise, it will be divided into smallercommunities or even disbanded (e.g., Fig. 3e). In the lattercase, Algorithm 1 shall be applied again to reconstruct thenew community structure on node i and all its neighbors inthe affected community.

5.2 Bridge Node IdentificationExploring bridge nodes locating between different com-munities as relay nodes not only helps quick disseminationof messages far away from the original community, butalso shortens the transmission hops between source anddestination. A prerequisite is to identify the bridge nodesbetween communities.

5.2.1 Weighted Bridging CentralityOn the measurement of bridge node, it is assumed in (1) thatall edges are of the same strength. Under our heterogeneousmodel, the edges in the social graph are with differentweights. The strong edge implies that it could transmit moremessages in a limited time, but the un-weighted bridgingcoefficient defined in (3) ignores such important feature.

To address this problem, we define weighted bridgingcoefficient Y0ðiÞ of node i based on our hybrid connectionstrength metric (4) as follow:

Y0ðiÞ ¼ dðiÞ�1

Pj2Ni

dðjÞ�1; (8)

where dðiÞ is the weighted degree of node i calculated as

dðiÞ ¼X

i6¼j2V ti

aði; jÞ �HMij: (9)

Note that aði; jÞ in (9) is a binary value indicating whetheri connects j in the social graph (i.e, aði; jÞ ¼ 1) or not (i.e.,aði; jÞ ¼ 0).

In addition, since FðiÞ in (1) is based on the wholenetwork, we utilize the method presented in [25] to getthe approximate betweenness, which correlates to thebetweenness based on the entire network. By substituting(8) and the approximate betweenness into (1), we obtainthe weighted bridging centrality.

The centralized computation of bridging centralitytakes time at order of Qðn3Þ, where n is the number ofnodes in a network. Our distributed computation achievesmuch lower time complexity of Qðd2

maxÞ, where dmax is themaximum node degree.

5.2.2 Bridge Node Self-RecommendingWhile the weighted bridging centrality can measure theimportance of a node located between communities, itcannot determine whether a node is a bridge node or not.Therefore, we propose a bridge node self-recommendingalgorithm in this section.

An edge update in a social graph may change thebridging centrality of the involving nodes. In other words,a normal node may become a bridge node or vice versaafter the update. Without loss of generality, we considersuch updates in a two-hop neighbor network. When twonodes i and j encounter, they shall update their bridgingcentrality by exchanging information. Based on the

updated bridging centrality, Algorithm 3 and Algorithm 4,as given in the supplementary file available online, will beinvoked to deal with bridge node identification andremoval, respectively.

Bridge node identification: First of all, if node i is theonly node shared by two communities C 2 Ct

i and C0 2 Ctj,

and has not yet been detected as a bridge, it will identifyitself as a bridge node as shown in lines 3-5 of Algorithm 3.Otherwise, node i will recommend itself as a candidate ofbridge node between C and C0 by sending a recommendationpacket to each encountered neighbor in C, if either conditionbelow holds as shown in lines 6-12 of Algorithm 3, in whichRiðC;C0Þ represents all known bridge candidates acrossC and C0.

1. Node i has not received any such recommendationpackets (i.e., RiðC;C0Þ ¼ ;) and its bridging centralitybecomes larger.

2. Node i has not been recognized as the bridge nodeof C and C0, i.e., i 62 BðC;C0Þ, and the weightedbridging centrality of node i is the highest in allknown candidates in RiðC;C0Þ.

Moreover, in the second case, node i considers itself asa bridge node by setting BiðC;C0Þ ¼ BiðC;C0Þ [ fig andBi Bi [ fðC;C0Þg.

Bridge node removal: The weighted bridge centrality ofa bridge node varies during the social graph aggregationprocess. In parallel to the process of bridge node identification,Algorithm 4 is also invoked to determine whether bridge nodei would degrade to a normal one when its bridging centralitybecomes smaller caused by edge updates in the social graph.For any community pairC 2 Ct

i andC0 2 Ctj, between which i

serves as a bridge node, if its weighted bridging centrality isnot the highest any more among the candidates in RiðC;C0Þ,node iwill cancel its role of bridge node betweenC andC0, andnotify each encountered neighbor in C by a retirement packet.

Meanwhile, upon the reception of a retirement orrecommendation packet, node i shall update Ri by remov-ing the corresponding node or including the recommendedone with its associated bridging centrality, respectively.

6 CONNECTION STRENGTH AWARE ROUTING

In this section, we present our distributed CSAR algorithm,which is specified in Algorithm 5 in the supplementary fileavailable online. It consists of two phases at node i: social-tiedetection and routing, when it meets node j. The first phasebuilds community structure and identifies the bridgesbetween communities by exchanging the social-tie informa-tion, as shown in line 1 of Algorithm 5. Then, the updatedsocial ties are utilized to address the problems of forwardingordering, message copy control, and buffer management in thesecond phase as follows.

6.1 Forwarding OrderingWhen two nodes encounter, since their contact duration islimited and unexpected, it is essential to determine whichmessages shall be forwarded (i.e., forwarding candidate set)and in what order these messages shall be transmitted (i.e.,forwarding order).


Let Fi denote the forwarding candidate set at node i,which is constructed as specified in lines 2-8 of Algorithm 5.The basic idea is that for each message m carried by nodei with a destination dðmÞ, it will be forwarded to node jonly if the delivery probability is improved in terms ofconnection strength, i.e.,

DUmji � HMjdðmÞ �HMidðmÞ 9 0: (10)

Specifically, only the following two cases will be considered:

1. If the encountered node j is in a community ofdestination dðmÞ (line 3 of Algorithm 5), we shall putm in Fi due to the fact that node j shall frequentlyencounter its community members.

2. If neither i nor j belongs to any community of thedestination dðmÞ but j is a bridge node (line 5 ofAlgorithm 5), m shall be also added into Fi. This isbecause the bridge node would be beneficial todeliver messages to its destination even far wayfrom the current community.

Notice that not all messages in Fi could be finallyforwarded to the encountered node in a short contactduration. In this case, the forwarding order of the messagesin Fi should be carefully considered. Intuitively, messageswhose delivery probability can be mostly increased shall beput in the front of the forwarding queue. Therefore,messages in Fi are sorted and forwarded in a descendingorder of DUm

ji as shown in line 9 of Algorithm 5.

6.2 Message Copy ControlIn general, the more copies generated for a message, thehigher delivery performance could be achieved. However,this may incur an unaffordable delivery overhead. On theother hand, while limiting the number of message copiesconserves consumption of resources (e.g., buffer, energy),it may degrade the delivery performance. Therefore, themessage copy control becomes a critical design issue whendelivery efficiency is considered.

To balance the tradeoff between the delivery performanceand overhead, we propose an inter-community multi-copy andintra-community single-copy mechanism. It is applied to delivereach message m in sorted Fi as shown in lines 10-16 ofAlgorithm 5. Right after forwarding m to the encounterednode j, node iwill deletem from its buffer if j shares a commoncommunity with message m’s destination dðmÞ (line 12), orstill keep the message in its local buffer otherwise (line 14). Therationale behind such design is that when messagem’s currentcarrier i is far away from the destination node (i.e., node i doesnot belong to any community of dðmÞ), spreading multiplecopies in a network can increase the chance of reaching itsdestination. On the other hand, once a message reaches its

destination’s community already, the message is still with ahigh probability to be finally delivered to its destination eventhere is only a single copy within the community.

6.3 Buffer ManagementDuring a contact, node imay also receive multiple messagesfrom its encountered node j. Owing to the buffer capacitylimitation, we shall consider which messages should bediscarded when the buffer overflows. To address this issue,we adopt a greedy approach that the message with thelowest value of HMidðmÞ is dropped as shown in line 17 ofAlgorithm 5.

7 PERFORMANCE EVALUATION

7.1 Simulation SetupIn order to evaluate the effectiveness of our proposedhybrid connection strength metrics (i.e., HM) based socialgraph aggregation, community detection approach (i.e.,DCDA) and routing protocol (i.e., CSAR), we conductintensive simulations on the widely-used DTN simulatorONE [29] using real traces, as summarized in Table 1.

In our simulations, each message is created with arandom interval from 30 to 40 seconds, a source-destinationpair randomly chosen from all nodes in the network, apredetermined time-to-live (TTL) representing the delayrequirement, and a uniform packet size of 25 KB. More-over, each node has a buffer size of 3 MB and bandwidthbetween any two encountered nodes is set as 250 Kbps.For each setting, 40 simulation instances with differentrandom seeds are conducted for statistical confidence.

We employ the following metrics to valuate the efficiencyof routing protocols.

/ Delivery Ratio: the ratio of the number of deliveredmessages to the number of generated messages.

/ Delivery Overhead: the average number of forward-ings of the same message before it reaches the destination.

/ Delivery Delay: the average time for delivering amessage to its destination over the network.

7.2 Performance of Connection Strength MetricsWe use MF and MD to denote the metrics purely based oncontact-frequency and contact-duration, respectively. Tovalidate the improved quality of aggregated social graphsby our proposed hybrid metrics, we compare the perfor-mance of existing social-based routing protocols Bubble-Rap and PeopleRank under various connection strengthmetrics HM, MF and MD.

We normalize the delivery ratio of BubbleRap andPeopleRank to the one obtained by direct delivery, whichserves as a lower bound on delivery ratio. As shown in

TABLE 1Characteristics of Five Mobility Datasets


Fig. 4, the performance under HM always achieves the bestfor both BubbleRap and PeopleRank. Our experimentalresults indicate that even existing unmodified social-basedrouting protocols can benefit from our hybrid connectionstrength metric HM. This is attributed to the fact that ingeneral HM provides a more accurate measurement onconnection strength that leads to high-quality social graphand superior routing performance.

7.3 Performance of CommunityDetection Algorithms

In order to evaluate the efficiency of our proposed DCDAalgorithm, we compare it with two state-of-the-art distrib-uted community detection algorithms: distributed k-clique[4] and modularity [13], in terms of community similarity,which is defined according to Jaccard index [30] as follows:

SimðCd; CcÞ ¼VCd

TVCcj j

VCdSVCcj j ; (11)

whereCd andCc are the communities detected by a distributedalgorithm and a centralized algorithm, respectively. We usethe centralized k-clique algorithm [12] to generate Cc in ourexperiments. A larger community similarity indicates that thecommunities detected by the distributed algorithm are moresimilar to the ones obtained by a centralized algorithm. Wecheck the average community similarities on datasets Realityand Infocom06 at several sampling points during runtime soas to assess the community evolution.

We find that the similarity by DCDA is always higher thanthe other two denoted as D k-clique and Modularity in Fig. 5.

For example, our DCDA algorithm improves similarity by17 percent and 13 percent compared to distributed k-cliqueand modularity, respectively, for dataset Infocom06. Weattribute this advantage to the unique capability of DCDAin capturing community partition, while both distributedk-clique and modularity only consider community creation.

Another interesting phenomenon observed from Fig. 5is that all these approaches have a relatively low similarityin the early stage. This is because each node takes time to collecttopology information in a distributed way. In addition, thesimilarity detected by distributed k-clique and modularitydecreases faster than that of DCDA after a certain time. Asshown in the results from dataset Infocom06, for example,this happens after 24 hours in the DCDA algorithm, but only6 hours in distributed k-clique and modularity algorithms.The reason behind is that only DCDA adopts a partitiondetection mechanism to remove those nodes which havealready left their previous communities such it achieveshigher similarity to the real one.

7.4 Performance of Social-AwareRouting Algorithms

In this section, we evaluate a set of social-aware routingprotocols: Epidemic [3], PeopleRank [5], BubbleRap [4], andour proposed CSAR, in which parameter� is set empiricallyon each dataset. The evaluation on the effect of parameter� can be found in the supplementary file available online.We present the experimental results based on datasetsInfocom06, Reality and Pmtr only since others are similarand thus omitted.

Fig. 4. Delivery performance of routing protocols under various connection strength metrics. (a) PeopleRank. (b) BubbleRap.

Fig. 5. Similarity of communities detected at different sampling time points. (a) Infocom06. (b) Reality.


7.4.1 Delivery RatioFig. 6 shows the delivery ratio as a function of TTL. It isclear to see that Epidemic has the highest delivery ratioamong all forwarding algorithms since it greedily forwardsmessages to each encountered node. In addition, we also noticethat CSAR outperforms both BubbleRap and PeopleRank.Taking Reality as an example, CSAR outperformsBubbleRap by 27 percent and PeopleRank by 40 percentwhen TTL is three hours. This phenomenon comes from thefollowing reasons.

1. First, recall that the performance of social-basedforwarding algorithms greatly rely on the social-tieinformation for relay node selection. The social tiesdetected in CSAR are based on the high-qualitysocial graph aggregated using our hybrid connec-tion strength metrics defined in (4).

2. We attribute the second reason to the message copycontrol strategies. Our CSAR algorithm adopts theinter-community multi-copy mechanism that spawnscopies of a message before it reaches its destination’scommunities. It significantly increases the messagedelivery probability. BubbleRap and PeopleRanksimply make use of a single replication strategyduring the entire routing process.

3. Finally, CSAR considers messages scheduling in adescending order of improved connection strengthdefined in (10) when multiple ones need to be

forwarded at each node, while BubbleRap andPeopleRank simply adopt a random selection strat-egy each each transmission opportunity.

We also notice from Fig. 6 that the delivery ratio showsas an increasing function of TTL. Furthermore, themaximum delivery ratios vary, i.e., around 50 percent,70 percent and 30 percent obtained by CSAR in datasetsReality, Infocom06 and Pmtr, respectively, due to theheterogeneity of inter-community contact rates on differentdatasets. For example, messages destined to the nodes in adifferent community from its carrier are difficult to bedelivered within a given TTL in dataset Pmtr, where inter-community contact rates are low. However, they are morelikely to be delivered within the same TTL in datasetInfocom06, where participants with the same interests in aconference session would meet again in other relatedsessions afterwards.

7.4.2 Delivery OverheadFig. 7 gives the evaluation results on the overhead ratio of thefour routing protocols on different datasets. With no doubt,Epidemic has the highest overhead ratio due to its floodingnature, while CSAR always achieves the lowest overhead ratioin all traces. In Reality, for example, the overhead ratio ofCSAR is only 71 percent and 81 percent of PeopleRank andBubbleRap, respectively, when TTL is set to a week. Specifi-cally, we have the following observations and conclusions.

Fig. 6. Comparison of all forwarding algorithms on several datasets in terms of delivery ratio. (a) Reality. (b) Infocom06. (c) Pmtr.

Fig. 7. Comparisons of all forwarding algorithms on several datasets in terms of overhead ratio. (a) Reality. (b) Infocom06. (c) Pmtr.


1. From the experimental results, we find that mes-sages are forwarded few times by bridge nodesbefore reaching the destination community, causingthat the multi-copy approach utilized in inter-community communication only brings a smallamount of overhead.

2. Both CSAR and BubbleRap utilize community struc-ture to select relay nodes, while PeopleRank overlooksthe importance of community. Nodes are more likelyto interact with the ones in the same community thanothers outside the community. To exploit such featureis essential to reduce the number of transmissions.

3. Moreover, community-aware forwarding can pre-vent a message from flowing outside the destina-tion’s community and unnecessary transmissionscould be avoided.

4. Finally, CSAR carefully determines the relay candi-date set and their forwarding order such that theutility of each transmission opportunity is maxi-mized. In other words, CSAR is able to avoidforwarding messages to those with relatively lowerforwarding capabilities. This can improve the trans-mission efficiency in each transmission opportunityand thus can reduce the overhead significantly. Incontrast, both PeopleRank and BubbleRap onlymake random forwarding for multiple messages.

We also observe that the overhead ratio increases withTTL. This is because larger TTL allows messages to staylonger in a network and thus to obtain more opportunitiesto be forwarded. Under any value of TTL, the advantage ofCSAR can be always observed.

7.4.3 Delivery DelayInstead of showing the absolute values of delivery delaydue to their large span on different TTLs, we present thenormalized results in Fig. 8. It defined as the ratio of delayobtained by a social-based routing protocol to the shortestdelay achieved by Epidemic because of its flooding nature.

As shown in Fig. 8, CSAR performs better than bothBubbleRap and PeopleRank in all cases. Taking Pmtr asan example, the normalized delivery delay of CSAR is10 percent and 18 percent lower than BubbleRap andPeopleRank, respectively, when TTL is set as 12 hours.We attribute its superior performance to the followingdirect and indirect reasons.

1. The direction reason is that using a multi-copyapproach in inter-community transmission shortensdelivery delay.

2. Our proposed message scheduling mechasim basedon (4) is an indirect reason as explained earlier.

3. Another important indirect reason is that our buffermanagement scheme drops the messages with thelowest probability to be delivered when the buffersize is insufficient.

Furthermore, we find out from Fig. 8 that the social-based routing protocols perform similarly to Epidemicwhen the TTL is small but shall deviate when the TTLincreases. This is because when the TTL is small, only themessages whose destinations are close to their sources canbe delivered successfully before they expire and hence allrouting protocols achieve low delivery delay. However, asTTL becomes larger, the mechanism of various routingprotocols starts functioning because many transmissionopportunities are available to each message before itsexpiration. After TTL reaches some value, the averagedelay converges to a certain value, indicating that theavailable transmission opportunities become limited underthose TTL settings. In summary, the substantial advantageof CSAR is exhibited under any values of TTL.

8 CONCLUSION

In this paper, we first introduce a new hybrid connectionstrength metrics in order to aggregate high-quality socialgraph that can even improve the performance of existingunmodified social-based routing protocols. We then pro-pose the distributed community detection algorithm andthe bridge node identification algorithm to discover theessential social ties for routing protocol design. It achievesmuch higher detection accuracy than existing algorithmsthanks to its capability of community partition capturing.To fully explore the benefits of our fundamental studies onsocial graph aggregation and social-tie detection, wepropose the routing algorithm CSAR, in which the issueson forwarding ordering, message copy control are buffermanagement are particularly addressed. Through exten-sive real-trace driven experiments, the superior efficiencyof CSAR is validated as it significantly outperformsexisting social-based routing protocols in terms of deliveryratio, overhead ratio and average delay.

Fig. 8. Comparison of all forwarding algorithms on several datasets in terms of average delay. (a) Reality. (b) Infocom06. (c) Pmtr.


ACKNOWLEDGMENT

This research was supported by National 863 Program ofChina (Grant No. 2012AA011005) and the fund of the StateKey Lab of Software Development Environment (GrantNo. SKLSDE-2013ZX-06).

REFERENCES

[1] K. Fall, ‘‘A Delay-Tolerant Network Architecture for ChallengedInternets,’’ in Proc. ACM SIGCOMM, 2003, pp. 27-34.

[2] K. Wei, X. Liang, and K. Xu, ‘‘A Survey of Social-Aware RoutingProtocols in Delay Tolerant Networks: Applications, Taxonomyand Design-Related Issues,’’ IEEE Commun. Surveys Tuts., vol. 16,no. 1, pp. 556-578, May 2013.

[3] A. Vahdat and D. Becker, ‘‘Epidemic Routing for PartiallyConnected Ad Hoc Networks,’’ Duke Univ., Durham, NC, USA,Tech. Rep. CS-200006, 2000.

[4] P. Hui, J. Crowcroft, and E. Yoneki, ‘‘Bubble Rap: Social-BasedForwarding in Delay Tolerant Networks,’’ in Proc. ACM MobiHoc,2008, pp. 241-250.

[5] A. Mtibaa, M. May, C. Diot, and M. Ammar, ‘‘PeopleRank: SocialOpportunistic Forwarding,’’ in Proc. IEEE INFOCOM, 2010, pp. 1-5.

[6] J. Wu and Y. Wang, ‘‘Social Feature-Based Multi-Path Routingin Delay Tolerant Networks,’’ in Proc. IEEE INFOCOM, 2012,pp. 1368-1376.

[7] M. Xiao, J. Wu, C. Liu, and L. Huang, ‘‘TOUR: Time-SensitiveOpportunistic Utility-Based Routing in Delay Tolerant Networks,’’in Proc. IEEE INFOCOM, 2013, pp. 2085-2091.

[8] C. Liu and J. Wu, ‘‘On Multicopy Opportunistic ForwardingProtocols in Nondeterministic Delay Tolerant Networks,’’ IEEETrans. Parallel Distrib. Syst., vol. 23, no. 6, pp. 1121-1128, June 2012.

[9] Q. Yuan, I. Cardei, and J. Wu, ‘‘An Efficient Prediction-BasedRouting in Disruption-Tolerant Networks,’’ IEEE Trans. ParallelDistrib. Syst., vol. 23, no. 1, pp. 19-31, Jan. 2012.

[10] Y. Li, L. Qiu, D. Jin, L. Su, P. Hui, and L. Zeng, ‘‘Contact DurationAware Evaluation for Content Dissemination Delay in Mobile SocialNetwork,’’ Wireless Commun. Mobile Comput. [Online]. Available:http://onlinelibrary.wiley.com/enhanced/doi/10.1002/wcm.2355.

[11] E.M. Daly and M. Haahr, ‘‘Social Network Analysis for Routingin Disconnected Delay-Tolerant MANETs,’’ in Proc. ACMMobiHoc, 2007, pp. 32-40.

[12] T. Hossmann, T. Spyropoulos, and F. Legendre, ‘‘Know ThyNeighbor: Towards Optimal Mapping of Contacts to Social Graphsfor DTN Routing,’’ in Proc. IEEE INFOCOM, 2010, pp. 1-9.

[13] P. Hui, E. Yoneki, S.Y. Chan, and J. Crowcroft, ‘‘DistributedCommunity Detection in Delay Tolerant Networks,’’ in Proc. ACM/IEEE Int’l Workshop Mobility Evol. Internet Architect., 2007, pp. 7:1-7:8.

[14] F. Li and J. Wu, ‘‘LocalCom: A Community-Based EpidemicForwarding Scheme in Disruption-Tolerant Networks,’’ in Proc.IEEE SECON, 2009, pp. 1-9.

[15] L. Pelusi, A. Passarella, and M. Conti, ‘‘Opportunistic Networking:Data Forwarding in Disconnected Mobile Ad Hoc Networks,’’ IEEECommun. Mag., vol. 44, no. 11, pp. 134-141, Nov. 2006.

[16] P. Hui and J. Crowcroft, ‘‘How Small Labels Create BigImprovements,’’ in Proc. IEEE PerCom Workshops, 2007, pp. 65-70.

[17] T. Nepusz, A. Petroczi, L. Negyessy, and F. Bazso, ‘‘FuzzyCommunities and the Concept of Bridgeness in Complex Net-works,’’ Phys. Rev. E, Stat. Nonlin. Soft Matter Phys., vol. 77, no. 1,pp. 016107-1-016107-12, Jan. 2008.

[18] S. Fortunato, ‘‘Community Detection in Graphs,’’ Phys. Rep., vol. 486,no. 3-5, pp. 75-174, Feb. 2010.

[19] W. Gao and G. Cao, ‘‘User-Centric Data Dissemination inDisruption Tolerant Networks,’’ in Proc. IEEE INFOCOM, 2011,pp. 3119-3127.

[20] W. Hwang, T. Kim, M. Ramanathan, and A. Zhang, ‘‘BridgingCentrality: Graph Mining from Element Level to Group Level,’’in Proc. ACM SIGKDD, 2008, pp. 336-344.

[21] W. Hwang, Y. Cho, A. Zhang, and M. Ramanathan, ‘‘BridgingCentrality: Identifying Bridging Nodes in Scale-Free Networks,’’University at Buffalo, Buffalo, NY, USA, Tech. Rep., 2006.

[22] J. Leguay and F. Benbadis, CRAWDAD Data Set UPMC/Rollernet,Feb. 2009. [Online]. Available: http://crawdad.cs.dartmouth.edu/upmc/rollernet.

[23] N. Eagle and A. Pentland, ‘‘Reality Mining: Sensing Complex SocialSystems,’’ Pers. Ubiquitous Comput., vol. 10, no. 4, pp. 255-268,May 2006.

[24] J. Leskovec, K.J. Lang, A. Dasgupta, and M.W. Mahoney, ‘‘StatisticalProperties of Community Structure in Large Social and InformationNetworks,’’ in Proc. ACM WWW, 2008, pp. 695-704.

[25] P. Pantazopoulos, M. Karaliopoulos, and I. Stavrakakis, ‘‘On theLocal Approximations of Node Centrality in Internet Router-Level Topologies,’’ in Proc. 7th Int’l Workshop Self-Org. Syst., 2013,pp. 1-12.

[26] J. Scott, R. Gass, J. Crowcroft, P. Hui, C. Diot, and A. Chaintreau,CRAWDAD Data Set Cambridge/Haggle/IMOTE/INFOCOM2006,May 2009. [Online]. Available: http://crawdad.cs.dartmouth.edu.

[27] G. Bigwood, D. Rehunathan, M. Bateman, T. Henderson, andS. Bhatti, Data Set ST andrews/sassy, June 2011. [Online]. Available:http://crawdad.cs.dartmouth.edu.

[28] P. Meroni, S. Gaito, E. Pagani, and G.P. Rossi, Data Set UNIMI/PMTR2008. [Online]. Available: http://crawdad.cs.dartmouth.edu.

[29] A. Keranen, J. Ott, and T. Karkkainen, ‘‘The ONE Simulator forDTN Protocol Evaluation,’’ in Proc. 2nd Int’l Conf. Simul. ToolsTech., 2009, pp. 1-10.

[30] P. Jaccard, ‘‘Etude Comparative de la Distribution Florale DansUne Portion des Alpes et des Jura,’’ Bull. Soc. Vaud. Sci. Nat., vol. 37,pp. 547-579, 1901.

Kaimin Wei joined the State Key Lab ofSoftware Development Environment (SKLSDE),in 2010, and is currently a PhD Student at theBeihang University in Beijing, majoring in com-puter science. His current research interests arein the areas of wireless mobile networks, delay/disruption tolerant networks, mobile social net-work and algorithm design.

Deze Zeng received the BS degree from Schoolof Computer Science and Technology, HuazhongUniversity of Science and Technology, China in2007, and the MS and PhD degrees in computerscience from University of Aizu, Aizu-Wakamatsu,Japan, in 2009 and 2013, respectively. He iscurrently a Research Assistant in University ofAizu, Japan. His current research interests includecloud computing, networking protocol design andanalysis, with a special emphasis on delay-tolerantnetworks and wireless sensor networks. He is a

member of the IEEE.

Song Guo received the PhD degree in computerscience from the University of Ottawa, Canada,in 2005. He is currently a Full Professor atSchool of Computer Science and Engineering,the University of Aizu, Japan. His researchinterests are mainly in the areas of protocoldesign and performance analysis for reliable,energy-efficient, and cost effective communica-tion in wireless networks. He received the BestPaper Awards at ACM Conference on UbiquitousInformation Management and Communication

2014, IEEE Conference on Computational Science and Engineering2011, and IEEE Conference on High Performance Computing andCommunications 2008. Dr. Guo currently serves as Associate Editor ofthe IEEE Transactions on Parallel and Distributed Systems. He is in theeditorial boards of ACM/Springer Wireless Networks, Wireless Commu-nications and Mobile Computing, International Journal of DistributedSensor Networks, International Journal of Communication Networks andInformation Security, and others. He has also been in organizingcommittee of many international conferences, including serving as aGeneral Chair of MobiQuitous 2013. He is a Senior Member of theIEEE.


Ke Xu received theBE,ME, andPhDdegrees fromBeihang University, in 1993, 1996, and 2000,respectively. He is a professor at Beihang Univer-sity, China.His research interests include algorithmand complexity, data mining, and network.

. For more information on this or any other computing topic,please visit our Digital Library at www.computer.org/publications/dlib.


Education

On social delay tolerant networking aggregation, tie detection,and routing