STrust: A Trust Model for Social Networks

Transcription

1 20 Internatonal Jont Conference of IEEE TrustCom-/IEEE ICESS-/FCST- STrust: A Trust Model for Socal Networks Surya Nepal, Wanta Sherchan, Cecle Pars Informaton Engneerng Lab CSIRO ICT Centre, Australa Frstname.Lastname@csro.au Abstract We propose a trust model for socal networks wth the am of buldng trust communtes that nspre members to share ther experences, feelngs and opnons n an open and honest way wthout the fear of beng judged. The unque feature of our model s that the trust value s derved from the socal captal bult n the socal networks over a perod of tme. Frst, we ntroduce a framework for buldng trust communtes usng socal trust. We then defne an underlyng socal trust model, called STrust. Fnally, we report the current state of the development and the analyss of the proposed trust model. Keywords- trust model, socal trust, socal captal, socal networks. I. INTRODUCTION The concept of socal networks s not new. Ths concept was frst ntroduced by J.A. Barnes n 954 [6]. Barnes descrbes socal networks as connected graphs where nodes represent enttes, and edges represent the nterdependences between two nodes. Enttes could be ndvduals, groups, organsatons or government agences. Smlarly, the edges could be nteractons, nvtatons, trades, values, etc. In recent tmes, the emergence of Web-based socal networks (e.g., Myspace, Facebook, etc.) has gven dfferent meanngs to socal networks n terms of ther szes []. The publc accessblty of Web-based socal networks usng moble phones makes such platform ubqutous [7]. Recent statstcs by HtWse [8] suggest that the user retenton rates of socal networks are as good as onlne bankng at hgh nnetes. Ths clearly ndcates the lkely longevty of socal networks. Therefore, they are most lkely to reman as the manstream platform for the foreseeable future. The use of Web-based socal networks was ntally lmted to ndvdual users for connectng them to ther frends and famles [9]. The phenomenal growth of users n recent tmes has not gone unnotced. Governments and enterprses have already started explotng the potental use of socal networks as platforms for delverng and mprovng ther servces. However, there have been reports n meda of many ncdents of breachng prvacy of ndvduals through socal networks [0], [2]. Gven the open nature of Webbased socal networks and ther current level of popularty, users have started to queston whether socal networks are brngng an end to prvacy. Prvacy s a very mportant consderaton for users. In order to balance the open nature of socal networks and safeguard the prvacy concerns of users, t s mportant to buld trust communtes. A trust communty s a communty that creates an envronment where ts members can share ther thoughts, opnons and experences n an open and honest way wthout concerns about ther prvacy and fear of beng judged. These communtes should be bult on the foundaton of authentcty, open sharng, lke-mndedness and mutual respect. We contend that socal trust provdes an deal foundaton for buldng trust communtes. Therefore, trust becomes an mportant aspect of socal networks. The concept of trust has been studed n many dscplnes ncludng socology [], psychology [2], economcs [3], and computer scence [6]. Each of these dscplnes has consdered trust from dfferent perspectves. There s no sngle consensus defnton of trust n the lterature. In general, trust s a measure of confdence that an entty or enttes wll behave n an expected manner despte the lack of ablty to montor or control the envronment on whch they operate [3]. Many of these defntons are lmted to the dscplne of ts orgn and may not be drectly applcable to socal networks. Therefore, t s mportant to also look at trust from the pont of vew of socal networks. The most mportant asset of any socety or a socal network s ts socal captal [5]. Socal captal refers to the collectve value assocated wth a socal network. The collectve value of socal networks can be measured usng dfferent approaches. Brune [4] studes socal captal through three dfferent approaches: relatonal, collectve and generalzed. All these approaches are applcable to socal networks. The more relevant to socal networks s Brune s defnton of socal captal from the pont of vew of collectve approach as follows:... socal captal s vewed as a collectve resource that facltates cooperaton at the small group level... no longer resdes wth an ndvdual but exsts through relatonshps between actors... s based on the densty of nteractons..." It s clear from the defnton that the nteractons play an mportant role n defnng and measurng the socal captal n socal networks. In socal networks, socal captal refers to the value of socal networks derved from the nteractons that provdes a support for bondng between lke-mnded people. We defne socal captal for socal networks as the densty of nteractons among actors that yeld benefts to them. Our defnton s n lne wth Putnam's vew [5] that refers to socal captal as a resource that can be drawn on by ndvduals through buldng up connectons wth others n socal and communty groups. Our defnton has two mportant factors: (a) the number of nteractons and (b) the / $ IEEE DOI 0.09/TrustCom

2 nature of nteractons. Socal captal becomes meanngless f the number of nteractons among members n the socal networks s very hgh, but the nature of those nteractons s not benefcal to ndvduals. Therefore, socal captal represents the densty of postve nteractons among actors n the onlne communty. These postve nteractons lay the foundaton for buldng trust communtes. In ths paper, we propose a socal trust model, called STrust, usng socal captal wth an am of buldng trust communtes. In our trust model, we separate the nteractons derved from socal captal nto two groups: popularty and engagement. The popularty based nteractons are n general based on the trustworthness of a member n the communty. Ths means f a member s trusted by other members n the communty, popularty based nteractons of the member wll ncrease. Smlarty, the engagement based nteractons are n general based on how much a member trusts other members n the communty. Ths means f a member trusts other members n the communty, engagement based nteractons of the member wll ncrease. Our model STrust separates trust values derved from these two types of nteractons as Popularty Trust (PopTrust) and Engagement Trust (EngTrust). The benefts of separaton of these two types of trust are that they can be used to recommend dfferent thngs. For example, the popularty trust could be used to recommend the leaders n the communty. Smlarly, recommendaton to be frends or buddes can be made usng the engagement trust. In ths paper, we frst defne the STrust model and then present ts evaluaton. The major contrbutons of the paper can be summarzed as follows: We propose a novel framework for buldng trust communtes n socal networks usng socal captal and recommendaton system. We propose a novel socal trust model, called STrust, for socal networks wth the am of buldng trust communtes. Unlke other trust models, the proposed socal trust model s based on the prncples of socal scence and derved usng socal captal. The proposed socal trust model separates the nteractons n a socal network nto two groups (popularty based nteractons and engagement based nteractons), whch enables the model to capture passve nteractons such as readng comments wthout leavng any feedback. The rest of the artcle s structured as follows. Secton II provdes the motvaton for the work by descrbng the need for socal trust n buldng trust communtes and proposes a fve steps framework for buldng trust communtes. Secton III presents our proposed socal trust model, called STrust. We also brefly present the analyss of the socal trust model and report our plan to mplement and use t n Secton IV. Fnally, Secton V presents the concludng remarks and dscusses our future work. II. A FRAMEWORK FOR BUILDING TRUST COMMUNITIES Socal networks have been wdely used n applcatons rangng from ndvduals to enterprses to governments. The am of these applcatons s to brng people together and buld communtes for them, where they can share ther experences and express ther feelngs. Such communtes can provde to ther members a wealth of experence and knowledge from other members, and a venue for communcatng and mprovng servces to servce provders. As reported n the lterature [3], trust s a crtcal factor n the success of these communtes. Wthout trust n place, members may not wsh to share ther knowledge or experence wth other communty members due to the fear of ther nformaton and denttes beng msused, or fear of beng judged. Therefore, we need to buld a communty where members trust each other as well as the communty and servce provders. We call such communtes trust communtes. The queston s then how do we buld a trust communty from a socal network. In ths paper, we propose a framework to buld trust communtes from socal networks. Our framework conssts of the followng fve steps as shown n Fgure 2. Personal Profle: A user frst regsters n a socal network and creates an account. The user then provdes personal detals that he/she would lke to share wth other members n the communty such as emal address, date of brth, hobbes, etc. Personal Identty: Ths step nvolves gvng an dentty to the user. For example, selecton of an mage or avatar, selecton of events of nterest, and selecton of frends are a few example of establshng an dentty of the user n the communty. We also refer the frst two steps as a user model. Fgure. Our framework for buldng trust communtes n socal networks Socal Captal: Users buld ther socal captal n ths step by nteractng wth each other. Example actvtes n ths step ncludes nvtng someone to be a frend, acceptng nvtatons to be frends, provdng comments on content put by others, provdng a comment on a comment, etc. The 842

3 purpose of ths step s to create an envronment for nteractons. Socal Trust: In ths step, we evaluate socal trust of an ndvdual user or the communty as a whole based on the socal captal. The socal trust may also nclude some nformaton about personal dentty and profle dependng on the trust model used for evaluaton. We refer the latter two steps as socal model. Recommendaton: The last step n our approach s the recommendaton. Our framework supports two types of recommendatons: (a) content-based recommendaton and (b) collaboratve flterng based recommendaton. The content-based recommendaton s based on the smlarty between two users although they may not have any drect nteracton. We plan to use the user model (personal dentty and personal profle) to support such recommendatons. The collaboratve flterng based recommendaton s based on the smlarty between two users based on the nteractons. We plan to use the socal model (socal captal and socal trust) to support such recommendatons. The am of the recommender s to make the onlne communty relevant to the members so that we can ncrease the socal captal and socal trust, whch n turn s used by the recommendaton system to recommend new actvtes or content. Ths cycle wll contnue untl we succeed n buldng the trust communty. Furthermore, t s necessary to contnue ths process beyond the buldng phase to mantan the trust communtes, due to the temporal decay of trust (also known as forgettng). In ths paper, our focus wll be on the socal model. We next develop a socal trust model, called STrust, usng socal captal and then evaluate t. III. PROPOSED SOCIAL TRUST MODEL (STRUST) In ths secton, we frst ntroduce the concept of socal trust and socal captal. We then defne a socal trust model, called STrust, for socal networks A. Socal Trust The noton of trust s used n dfferent dscplnes to represent dfferent concepts. Here, we frst explan what we mean by trust n socal networks. We defne socal trust as the frm belef n the competence of an entty to act as expected, such that ths frm belef s not a fxed value assocated wth the entty, but rather t s subject to entty s behavour and apples only wthn a specfc context at any gven tme. There are three mportant aspects of trust: (a) trust depends on the user behavour, (b) trust s dynamc and (c) trust s context dependent. We explan each of these brefly before defnng our socal trust model. User behavor: Socal trust depends on the behavour of an ndvdual n the socal networks. In the context of an onlne communty, the behavour of an ndvdual s derved from hs/her socal captal (nteractons) n the communty. There are two types of nteractons: actve and passve. An example of actve nteracton ncludes havng a large number of frends, postng nformaton regularly, replyng to other members posts, etc. However, not all members n the communty are actve partcpants. There are sgnfcant numbers of members who are passve partcpants n the communty. Interactons of the passve members n the communty nclude readng posts, readng artcles, regular vsts to the communty, etc. These members may not contrbute or share ther experences or feelngs, but they are consumers of nformaton, and ths s also very valuable. These types of nteractons have receved least attenton n the trust models n the current lterature. These two types of nteractons collectvely buld the socal captal of the communty and are used to evaluate socal trust. Temporal factor: Another mportant aspect of trust s that t s tme dependent. The decay of socal captal wth tme s the fact of socal lfe n socal networks. An nteracton that has happened more recently may have more value than those that have happened some tme back. Therefore, tme s an mportant factor to capture the change n the behavour of an ndvdual. For example, a member X may have a good relatonshp wth another member Y at tme t, but ths relatonshp may weaken f there are not recent nteractons between them. Ths means a communty member mantans nteractons wth other members wth whom they have establshed a trust relatonshp. Context: The other mportant element n our defnton of socal trust s context. We llustrate ths wth an example. A member X n the communty trusts another member Y for hs/her recommendaton of moves. X may not trust Y for a book or restaurant recommendaton. Here, moves represent the context of trust between X and Y. Ths reflects the realty n our real lfe. For example, a wfe may trust her husband to fx a problem wth the TV, but may not trust hm to fx the car. There are many dfferent ways of representng contexts, rangng from professonal jobs such as mechancs to observatons, nformaton and experences. In trust and reputaton systems, t s assumed that X trusts Y means that X smply follows Y or agrees wth Y. For example, when a member X trusts another member Y n the onlne communty, and Y recommends a move Z, X would lke to see the move Z. Ths s not always true n real lfe. It may be the case that Y lkes romantc moves and X lkes acton moves. X trusts Y, but may never follow Y s recommendaton of moves because X knows Y s taste n moves s dfferent. Therefore, t s mportant to dstngush between the dfferent contexts of trust and determne whether t eventuates to certan actons. B. Socal Captal Model In the above, we defned three core elements n our defnton of socal trust. Our socal trust model STrust endeavors to capture all these three essental elements. We next descrbe the proposed socal model to represent and capture socal captal. Our socal trust model s based on socal captal,.e., postve nteractons among the members n the communty. Ths s n contrast to exstng trust lterature where all nteractons are treated equally and the passve nteractons do not get much attenton. We separate socal captal nto 843

4 two types: popularty and engagement. We further explan the popularty and engagement trust usng a graphcal representaton of nteractons n an onlne communty as shown n Fgure 3. propertes Descrpton (short descrpton of the context) and Tme (tme of creatng the context). Fgure 2. Network model of socal captal The nodes n the graph represent communty members and the edges represent the nteractons between them. The nodes could represent other enttes n the communty such as actvtes or contents, whch we refer to as passve nodes. If we consder node B, there are four outgong lnes and three ncomng lnes. The outgong lnes support the engagement trust, and the ncomng lnes support the popularty trust. Each arrowed lne provdes nformaton towards popularty trust for one sde (snk or recevng end) and engagement trust on the other sde (source or ntatng end). For example, there s an arrowed lne comng from node A towards node B n Fgure 3. Ths lne provdes nformaton towards the engagement trust of node A and the popularty trust of node B. The sold lne represents actve nteracton, and dotted lne represents passve nteracton. We next descrbe our proposed data model to capture the socal captal as shown n Fgure 4. The nodes n the graph are the major enttes n the model resource (A), ratng (R), user (U), comment (C) and context (X). The leaves of the graph represent the propertes of each of the enttes. The arrows represent the relatonshps between the enttes. The relatonshps between the enttes are as follows: User provdes ratngs (usually for a resource or a comment). Therefore, resource has ratngs and comment has ratngs. A User consumes resources and wrtes comments. A resource has a context. A comment has a context. A resource may have other resources. Smlarly, a comment may have other comments. Each entty has a large number of propertes. We have shown some mportant propertes n the fgure. The Ratng entty has two propertes- Score and Tme. Score represents the ratng value and tme ndcates the tme of ratng. The entty User has two propertes UsrId (dentty of the user) and Name (name of the user). The Comment entty has two propertes Tme (tme of creaton of comment) and Content (content of comment). The Resource entty has three propertes Tme (tme of creaton of the resource), ResId (dentty of the resource) and Content (content of the resource). The Context entty has two Fgure 3. A data model for capturng socal captal C. Socal Trust Model (STrust) We next descrbe our socal trust model STrust that captures the three essental elements of socal trust usng the data captured by socal captal model descrbed earler. Our trust model contans two types of trust derved from the two types of nteractons defned earler. Popularty Trust: The popularty trust refers to the popularty of an ndvdual member n the communty. Popularty of a member refers to the trustworthness of the member from the perspectve of other members n the communty. We frst need to dentfy and derve metrcs that can be used to model popularty trust from the socal captal (or nteractons). Example metrcs for popularty trust nclude the number of postve feedback receved on the member s posts, the number of nvtaton requests for frendshp receved by the member, etc. Detaled descrpton of the metrcs s out of the scope of ths paper. We model the popularty trust usng beta famly of probablty dstrbuton functon [4]. Let U be the set representng the number of members n + the communty and PT represent the total number of postve nteractons a member u U has wth the member u j U wth respect to popularty trust and represented as +. Smlarly, the negatve nteracton s represented as PT. A member n the communty may be nvolved n a number of actvtes related to a sngle context. A member may post a large number of messages n dfferent contexts or wthn a sngle context. For example, a member may have a post related to a move. The move represents the context. There may be a large number posts on the move. We need to 844

5 consder ths whle evaluatng total postve nteracton for a context (x) PT. It s defned as follows. PT = + x PT = where X represents the number of contexts, and A represents the number of actvtes n each context. The popularty trust (PopTrust) of a member u U for a partcular context (x) s then defned as: M PT + x = PT + PT + 2 x j u ) = M The aggregaton over all contexts gves the popularty trust of the member n the communty as follows: u ) = x= u x Engagement Trust: The engagement trust refers to the nvolvement of an ndvdual member n the communty. Engagement of a member represents the trust the member has towards the communty. Smlar to the popularty trust, we need to derve metrcs that can be used to model engagement trust from the socal captal (or nteractons). Example metrcs for engagement trust nclude the number of postve feedback provded on the posts, the number of nvtaton requests sent to other members, etc. Engagements n a communty can be of two types and hence the correspondng metrcs: actve and passve. Actve engagement nvolves sendng requests or postng comments or provdng ratngs anythng where there s an nput from the member. Passve engagement nvolves someone vstng the communty regularly, readng the posts, and observng the conversatons. In our model, we have consdered both actve engagement and passve engagement. We defne the engagement trust model n a smlar way to the popularty trust model as follows. M + ET x = ET + ET + 2 x j EngTrust( u ) = M EngTrust( u ) = where x= ) EngTrust( u ET = +, and x ET = The socal trust (STrust) of an ndvdual member u U n the communty s then gven by: STrust u ) = α. u ) + ( α ). EngTrust( u ) x ) ( where α represents the value of weght n the range of 0 to. If α =, the socal trust of an ndvdual ndcates how much other members n the communty trust hm/her. Ths s equvalent to reputaton of a member n the communty. If α =0, the socal trust represents how much a member trusts others n the communty. In deal trust communtes, all members n the socety have almost the same socal trust. We can defne the socal trust of a communty as: STrust( c) = M = SocalTrust( u ) M An mportant queston n ths context s: what s an deal value of socal trust for a communty so that we can call the communty a trust communty? Answerng ths queston s challengng and nterestng. As we are bootstrappng the trust value for a communty at 0.5, the socal trust of the communty must be greater than 0.75 to be even consdered as a trust communty. Ths depends on a number of factors, and the analyss of these factors s outsde the scope of ths paper. Our socal trust model so far has covered the behavor (nteractons) and context. We next descrbe the temporal factor. The behavour of the members may change over tme. Assume that the socal trust of an ndvdual member s ndexed on tme. If there are T numbers of ntervals, then the socal trust of a member m at tme t s gven by. T ( T ) STrust( u ) λ = STrust( u, t) = T where 0 λ. IV. ANALYSIS The proposed trust model and correspondng recommendaton systems are beng mplemented for an onlne communty we are developng for delverng human servces (for an Australan Government agency). We have chosen Lferay as the mplementaton platform. As the communty portal s under development, we have analyzed the proposed trust model n a smulated envronment. The purpose of the smulaton s to explan some features of the model further. In our smulated communty, we have 000 users and each user has 0000 nteractons. We generated a random number to select the member wth whom a user s nteractng. We also used the Gaussan dstrbuton to generate the postve and negatve nteractons wth mean value 0 and the devaton. Snce the Gaussan dstrbuton s unform, we forced the frst half of the users to have nteractons wth only users from 0 to 500. Fgure 5 shows the overall results of our smulaton. As shown n the result, the bootstrap trust value for all users s 0.5 as we have used the beta famly probablty dstrbuton functon [4]. The engagement trust for the frst half s greater than the second half due to our forceful nteractons. We then changed the value of and observe the change n socal trust. (accessed on March, 20) 845

6 Fgure 6 shows the resultant socal trust values for dfferent users for dfferent values of. Trust Value Member Bootstrap PopTrust EngTrust STrust Fgure 4. Dfferent trust values for 000 users n our communty STrust Member Fgure 5. Socal trust values for dfferent weghts alph0.5 alpho.25 alph0 alph0.75 alph.0 Fnally, we studed the temporal aspect of socal trust n our model. We have chosen the fve temporal wndows (0, 2, 4, 6, 8, 0), where the frst temporal wndow represents the current tme perod. Assumng that there wll be no further nteractons n the communty n future, we studed the effect of tme on the decay of socal trust value. Fgure 7 shows the results for two users chosen randomly from our communty. Trust Value Temporal Wndows (T) Fgure 6. Temporal aspect of socal trust V. CONCLUSIONS u=2 u=545 In ths paper, we have presented a framework to buld trust communtes through a model based on socal trust derved from socal captal. A model for socal trust, called STrust, s defned based on ndvdual members popularty and engagement n the communty. We have also presented the potental benefts of the proposed model. We have dentfed a number of avenues for potental future work. Frst, we need to develop an approach for bootstrappng the onlne communty so that the causal cycle between socal trust and socal captal can be solved. Second, the ssue of determnng when an onlne communty becomes a trust communty needs to be resolved. Ths means we need to determne a threshold socal trust value of the communty that would make an onlne communty a trust communty. Thrd, the temporal decay on socal captal may or may not cause the temporal decay on trust. How do we dentfy t n an onlne communty? Fnally, the measurement of the effectveness of our approach s not easy. We plan to mplement a communty portal for real users of an e- government applcaton (human servces delvery) and test and evaluate the model wth real users. REFERENCES [] Golbeck, J. The dynamcs of Web-based socal networks: Membershp, relatonshps, and change. Frst Monday 2(): [2] Young, A. L. and Quan-Haase, A. Informaton revelaton and nternet prvacy concerns on socal network stes: a case study of facebook. Internatonal conference on Communtes and technologes, 2009, pp [3] Sngh, S. and Bawa, S. Prvacy, Trust and Polcy based Authorzaton Framework for Servces n Dstrbuted Envronments. Internatonal Journal of Computer Scence 2(2):85-92, [4] Jøsang, A. The Beta Reputaton System. 5th Bled Electronc Commerce Conference e-realty: Constructng the e-economy Bled, [5] Paldam, Martn. Socal Captal: One or Many? Defnton and Measurement. Journal of Economc Surveys 4(5): [6] Barnes, J. A. Class and commttees n a norwegan sland parsh. Human Relatons 7 (): 39-54, 954 [7] Humphreys, L. Moble socal networks and socal practce: A case study of dodgeball. Journal of Computer-Medated Communcaton 3 (): , [8] Hanchard, S. Measurng trust and socal networks - where do we put our trust onlne? [9] Haythornthwate, C. Socal networks and nternet connectvty effects. Informaton, Communcaton and Socety 8 (2): [0] Gross, R. and Acqust, A Informaton revelaton and prvacy n onlne socal networks ACM workshop on Prvacy n the electronc socety. NY, USA, pp [] Molm, L. D., Takahash, N., and Peterson, G. Rsk and trust n socal exchange: An expermental test of a classcal proposton. Amercan Journal of Socology 5 (05): [2] Cook, K. S., Yamagsh, T., Cheshre, C., Cooper, R., Matsuda, M., and Mashma, R. Trust buldng va rsk takng: A cross-socetal experment. Socal Psychology Quarterly 2(68): [3] Huang, F. Buldng socal trust: A human-captal approach. Journal of Insttutonal and heoretcal Economcs (JITE) 63 (4): [4] Auréle Brune. Meanngful dstnctons wthn a concept: Relatonal, collectve and generalzed socal captal. Socal Scence Research, 38(2): [5] Putnam, R. (2000). Bowlng alone: the collapse and revval of Amercan communty. New York: Smon & Schuster. [6] Maheswaran, M., Tang, H. C., and Ghunam, A. Towards a gravtybased trust model for socal networkng systems. Internatonal Conference on Dstrbuted Computng Systems Workshops