Collaborative Reputation Mechanisms in Electronic Marketplaces

Transcription

1 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences 999 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences Collaboraive epuaion Mechanisms in Elecronic Markeplaces Giorgos Zacharia, Alexandros Moukas and Paie Maes MIT Media Laboraory 20 Ames Sree, oom 305, Cambridge MA 0239 USA lysi,moux,paie@media.mi.edu Absrac The members of elecronic communiies are ofen unrelaed o each oher, hey may have never me and have no informaion on each oher's repuaion. This kind of informaion is vial in Elecronic Commerce ineracions, where he poenial counerpar's repuaion can be a significan facor in he negoiaion sraegy. This paper proposes wo complemenary repuaion mechanisms ha rely on collaboraive raing and personalized evaluaion of he various raings assigned o each user. While hese repuaion mechanisms are developed in he conex of elecronic commerce, we believe ha hey may have applicabiliy in oher ypes of elecronic communiies such as charooms, newsgroups, mailing liss ec.. Inroducion Consumer o consumer elecronic ransacion sysems like Kasbah [], ebay [2] and "ONSALE Exchange Aucion Classifieds" [3] creae online marke places ha bring ogeher users unknown o each oher. Kasbah is an ongoing research projec o help realize a fundamenal ransformaion in he way people ransac goods -- from requiring consan monioring and effor, o a sysem where sofware agens do much of he bidding and negoiaing on a user's behalf. A user waning o buy or sell a good creaes an agen, gives i some sraegic direcion, and sends i off ino he agen markeplace. Kasbah agens pro-acively seek ou poenial buyers or sellers and negoiae wih hem on heir creaor's behalf. Each agen's goal is o make he "bes deal" possible, subjec o a se of user-specified consrains, such as a desired price, a highes (or lowes) accepable price, and a dae o complee he ransacion []. OnSale Exchange is an online aucion specializing in compuer producs, consumer elecronics, sporing goods, and aucion classifieds, where sellers can lis heir iems for sale and buyers compee in he aucion-like bidding sysem o buy he posed iems. These kinds of online markeplaces inroduce wo major issues of rus among he users of he sysem:. The poenial buyer has no physical access o he produc of ineres while he/she bids or negoiaes. Therefore he seller could misrepresen he condiion or he qualiy of his/her produc in order o ge more money. 2. The seller or buyer may decide no o abide by he agreemen reached a he elecronic markeplace asking a some laer ime o renegoiae he price, or even refusing o commi he ransacion. An approach, which could solve he above menioned problems, would be o incorporae in he sysem a repuaion brokering mechanism, so ha each user can acually cusomize his/her pricing sraegies according o he risk implied by he repuaion values of he poenial counerpars. epuaion is usually defined as he amoun of rus inspired by he paricular person in a specific seing or domain of ineres [4]. In "Trus in a Crypographic Economy" [5] repuaion is regarded as asse creaion and i is evaluaed according o is expeced economic reurns. epuaion is conceived as a mulidimensional value. An individual may enjoy a very high repuaion in one domain, while he/she has a low repuaion in anoher. For example, a Unix guru will naurally have a high rank regarding Linux quesions, while he may no enjoy ha high a repuaion for quesions regarding Microsof s operaing sysems. These individual repuaion sandings are developed hrough social ineracion among a loosely conneced group ha shares he same ineres. We are developing mehods hrough which we can auomae he social mechanisms of repuaion for he purposes of an elecronic markeplace. These repuaion mechanisms are implemened and esed in he Kasbah elecronic markeplace. In Kasbah, he repuaion values of he individuals rying o buy or sell books or CDs are a major parameer of he behavior of he buying, selling or finding agens of he sysem. In his paper we describe wo repuaion mechanisms: /99 $0.00 (c) 999 IEEE

2 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences 999 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences a) Sporas is a simple repuaion mechanism which can be implemened irrespecively of he number of raed ineracions, and b) Hisos is a more complex repuaion mechanism ha assumes ha he sysem has been somehow boosrapped (probably by using Sporas) so ha here is an abundance of raed ineracions o creae a dense web of pairwise raings. In he firs secion of he paper we ouline he problem we are rying o solve and he problems we faced during he iniial implemenaion of he sysem. The second secion describes relaed work and he hird secion oulines specific problems inheren o online markeplaces like Kasbah. In he fourh and he fifh secion we describe he soluion adoped in he case of Kasbah, and in he sixh secion we presen resuls from simulaions. Finally he las secion oulines fuure goals for research in repuaion mechanisms for online communiies. 2. elaed Work The sudy of repuaion brokering for online communiies has generally focused on conen filering. We can divide he relaed work ino wo major caegories: he non-compuaional ones like he local Beer Business Bureaus [6] and he compuaional ones. The compuaional mehods cover a broad domain of applicaions, from raing of newsgroup posings and webpages, o raing people and heir experise in specific areas. In his secion we focus on he relaed compuaional mehods and compare heir major feaures in Table. One approach o building a repuaion mechanism is o have a cenral agency ha keeps records of he recen aciviy of he users on he sysem, very much like he scoring sysems of credi hisory agencies [7]. This cenral agency also keeps records of complains by users in exual formas and even publishes warnings agains possibly malicious users, prey much like he local Beer Business Bureaus in he US [6]. Adoping such a soluion requires a lo of overhead on he behalf of he providers of he online communiy. Oher proposed approaches are more disribued. For example, approaches such as Yena [8], Weaving a web of Trus [9], or he Plaform for Inerne Conen Selecion (PICS) [0], (like he ecreaional Sofware Advisory Council []) would require ha users give a raing for hemselves and eiher have a cenral agency or oher rused users verify heir rusworhiness. We can make he reasonable assumpion ha no user would ever label him/herself as a non-rusworhy person. Thus all new members would have o awai he validaion by oher rusworhy users of he sysem. A user would end up evaluaing heir counerpars' repuaion as a consequen of he number and he rusworhiness of he recommendaions for each user. Friend of a Friend Finder (FFF) [2], Yena and Weaving Web of Trus inroduce compuaional mehods for creaing personal recommendaion sysems, he former wo for people and he laer for webpages. FFF and he Weaving a Web of Trus rely on he exisence of a conneced pah beween wo users, while Yena clusers people wih shared ineress according o he recommendaions of users ha know each oher and can verify he asserions hey make abou hemselves. All hree sysems require he a priori exisence of social relaionships among he users of heir online communiy, while in he online markeplaces, deals are brokered among people who probably have never me each oher. Collaboraive filering is a echnique used o deec paerns among he opinions of differen users and o make recommendaions o people, based on ohers who have shown similar ase. I essenially auomaes he process of "word of mouh" o produce an advanced, personalized markeing scheme. Examples of collaboraive filering sysems are HOM [3], Firefly [3] and GroupLens [4]. GroupLens is a collaboraive filering soluion for raing he conen of Usene aricles and presening hem o he user in a personalized manner. Users are clusered ogeher according o he raings hey give o he same aricles. The user sees he aricles wih a value equal o he average of he raings given o he aricle by users in he same cluser. The mos relevan compuaional mehods o our knowledge, are he repuaion mechanism of he OnSale Exchange [3] and he ebay [2]. OnSale allows is users o rae and submi exual commens abou sellers and overall repuaion value of a seller is he average of his/her raings hrough his usage of he OnSale sysem. In ebay, sellers receive +, 0 or as feedback for heir reliabiliy in each aucion and heir repuaion value is calculaed as he sum of hose raings over he las six monhs. In OnSale, he newcomers have no repuaion unil someone evenually raes hem, while on ebay hey sar wih zero feedback poins. However, bidders in he OnSale Exchange aucion sysem are no raed a all. OnSale ries o ensure he bidders' inegriy hrough a raher psychological measure: bidders are required o regiser wih he sysem by submiing a credi card. OnSale believes ha his requiremen helps o ensure ha all bids placed are legiimae, which proecs he ineress of all bidders and sellers. In boh sies he repuaion value of a seller is available wih any exual commens ha may exis o he poenial bidders /99 $0.00 (c) 999 IEEE 2

3 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences 999 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences Table Comparison of repuaion sysems Sysem Compuaional Pair-wise raing Personalized Texual commens GroupLens Yes raing of aricles Yes Elo & Glicko Yes resul of game OnSale Yes buyers rae sellers FairIsaac Yes Yes Local BBB s Yes Web of Trus Yes Self raing of cos Yes Kasbah Yes Yes Yes Firefly Yes aing of recommendaions Yes Ebay Yes buyers rae sellers Yes The laes release of Kasbah [] feaures a Beer Business Bureau service ha implemens he repuaion mechanisms we describe below. 3. Desideraa for online repuaion sysems While he above discussed repuaion mechanisms have some ineresing qualiies, we believe hey are no perfec for mainaining repuaions in online communiies and especially in online markeplaces. This secion describes some of he problems of online communiies and heir implicaions for repuaion mechanisms. In online communiies, i is relaively easy o adop a new or change one's ideniy. Thus, if a user ends up having a repuaion value lower han he repuaion of a beginner, he/she would have an incenive o discard his/her iniial ideniy and sar from he beginning. Hence, i is desirable ha while a user's repuaion value may decrease afer a ransacion, i will never fall below a beginner's value. We herefore decided for he repuaion mechanisms described in he following secion ha a beginner canno sar wih an average repuaion. We also wan o make sure ha even if a user sars receiving very low repuaion raings, he/she can improve his/her saus laer a almos he same rae as a beginner. If he repuaion value is evaluaed as he arihmeic average of he raings received since he user joined he sysem, users who perform relaively poorly in he beginning, have an incenive o adop a new ideniy so ha hey ge rid of heir bad repuaion hisory. Anoher problem is ha he overhead of performing fake ransacions in boh Kasbah and OnSale Exchange is relaively low (OnSale does no charge any commission on is Exchange service ye). Therefore wo friends migh decide o perform some dozens of fake ransacions, raing each oher wih perfec scores so as o boh increase heir repuaion value. Even if we allow each user o rae anoher only once, anoher way o falsely increase one's repuaion would be o creae fake ideniies and have each one of hose rae he user's real ideniy wih perfec scores. A good repuaion sysem would avoid boh hese problems. We have o ensure ha hose raings given by users wih an esablished high repuaion in he sysem are weighed more han he raings given by beginners or users wih low repuaions. In addiion he repuaion values of he users should no be allowed o increase a infinium like he case of ebay, where a seller may chea 20% of he ime bu he/she can sill mainain a monoonically increasing repuaion value. Finally we have o consider he effec of he memory of our sysem [4]. The larger he number of raings used in he evaluaion of repuaion values he highes he predicabiliy of he mechanism i ges. However, since he repuaion values are associaed wih human individuals and humans change heir behavior over ime i is desirable o disregard very old raings. Thus we ensure ha we he prediced repuaion values are closer o he curren behavior of he individuals raher heir overall performance. 4. Sporas: A repuaion mechanism for loosely conneced online communiies Sporas provides a repuaion service based on he following principles:. New users sar wih a minimum repuaion value, and hey build up repuaion hroughou heir aciviy on he sysem /99 $0.00 (c) 999 IEEE 3

4 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences 999 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences The repuaion value of a user should no fall below he repuaion of a new user no maer how unreliable he user is. 3. Afer each raing he repuaion value of he user is updaed based on he feedback provided by he oher pary o reflec his/her rusworhiness in he laes ransacion. 4. Two users may rae each oher only once. If wo users happen o inerac more han once, he sysem keeps he mos recenly submied raing. 5. Users wih very high repuaion values experience much smaller raing changes afer each updae. This approach is similar o he mehod used in he Elo [5] and he Glicko [6] sysem for pairwise raings. Each user has one repuaion value, which is updaed as follows: As we can see from Equaion, he change in he repuaion value of he user receiving a raing of W i from user i oher, is proporional o he repuaion value i oher of he raer himself. The expeced raing of a user is his/her curren repuaion value over he maximum repuaion value allowed in he sysem. Thus if he submied raing is less han he expeced one he raed user loses some of his repuaion value. The value of θ deermines how fas he repuaion value of he user changes afer each raing. The larger he value of θ, he longer he memory of he sysem. Thus, jus like credi card hisory schemes [7], even if a user eners he sysem being really unreliable in he beginning, if he/she improves laer, his/her repuaion value will no suffer forever from he early poor behavior. Φ E + = θ Φ ( ) = ( D ) + e ( ) + = D oher ( ) ( W E( )) i σ Equaion Sporas formulae Where, is he number of raings he user has received so far, θ is a consan ineger greaer han, W i represens he raing given by he user i, oher is he repuaion value of he user giving he raing D is he range of he repuaion values, σ is he acceleraion facor of he dumping funcion Φ. The smaller he value of σ, he seeper he dumping facor Φ(). New users sar wih repuaion equal o 0 and can advance up he maximum of The repuaion raings vary from 0. for errible o for perfec. Since he repuaion of a user in he communiy is he weighed average of non-negaive values, i is guaraneed ha no user can ever have a negaive repuaion value, hus no user can ever have lower han ha of a beginner. Also he weighed average schema guaranees ha no user exceeds he maximum repuaion value of If a user has a persisen real repuaion value, he ieraion of Equaion over a large number of raings will give as an esimae very close o ha value [Figure ]. Figure Change of repuaion for 0 differen users over 00 raings wih θ=0 5. Hisos: A repuaion mechanism for highly conneced online communiies "Alhough an applicaion designer's firs insinc is o reduce a noble human being o a mere accoun number for he compuer's convenience, a he roo of ha accoun number is always a human ideniy." Weaving a Web of Trus [9]. The repuaion mechanism described in he previous secion provides a global repuaion value for each member of he online communiy, which is associaed wih hem as par of heir ideniy. Besides he online agen mediaed ineracion, our users will evenually have o mee each oher physically in order o commi he agreed ransacion, or hey may even know each oher hrough oher social relaionships. The exising social relaionships as well as he acual physical ransacion process creae personalized biases on he rus relaionships beween hose users. FFF [2] and he PGP /99 $0.00 (c) 999 IEEE 4

5 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences 999 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences web of Trus [7] use he idea ha as social beings we end o rus a friend of a friend more han a oal sranger. Following a similar approach, we decided o build a more personalized sysem. In Weaving a Web of Trus [9], wha maers is ha here is a conneced pah of PGP signed webpages beween wo users. In our case we have o ake ino consideraion he differen repuaion raings connecing he users of our sysem. We can represen he pairwise raings in he sysem as a direced graph, where nodes represen users and weighed edges represen he mos recen repuaion raing given by one user o anoher, wih direcion poining owards he raed user. If here exiss a conneced pah beween wo users, say from A o A L, we can compue a more personalized repuaion value for A L. A.6 A 5 A A A A 4.6 A 3.7 A.2 6 A3.3 A 7.9 A 8 A 0.7 A 2 Figure 2 A direced graph represening he raing pahs beween user A and A 3 When he user A submis a query for he Hisos repuaion value of a user A L we perform he following compuaion: a) The sysem uses a Breadh Firs Search algorihm o find all direced pahs connecing A o A L ha are of lengh less han or equal o N. As described above we only care abou he chronologically θ mos recen raings given o each user. Therefore, if we find more han θ conneced pahs aking us o user A L, we are ineresed only in he mos recen θ pahs wih respec o he chronological order of he raing evens represened by he las edge of he pah. b) We can evaluae he personalized repuaion value of A L if we know all he personalized repuaion raings of he users a he las node of he pah before A L. Thus, we creae a recursive sep wih a mos θ pahs wih lengh a mos N c) If he lengh of he pah is only, i means ha he paricular user, say C, was raed by A direcly. The direc raing given o user C is used as he personalized repuaion value for user A. Thus, he recursion erminaes a he base case of lengh θ and has an order of growh bounded by: O ( θ N ) Equaion 2 Order of growh of Hisos Noe ha for any lengh θ user A may have even been among he las θ users ha have raed A L direcly. However, user A has he opion of geing oher peoples' opinions abou A L by evaluaing his personalized value for A L in a more collaboraive fashion. Also for he purpose of calculaing he personalized repuaion values, we use a slighly modified version of he repuaion funcion described above. For each user A L, wih m conneced pahs coming owards A L from A, we calculae he repuaion of A L as follows: oher = Φ( ) ( W ) + ' θ ' = min ( θ, m) m = deg( A θ θ ' θ ' Equaion 3 Hisos formulae L ) Where deg (A L ) is he number of conneced pahs from A o A L wih lengh less han or equal o he curren value of L. In he base case where L=, since we have a conneced pah i means ha A has raed A him/herself, he personalized value for A is naurally he raing given by A. In order o be able o apply he Hisos mechanism we need a highly conneced graph. If here does no exis a pah from A o A L wih lengh less han or equal o N, we fall back o he simplified Sporas repuaion mechanism /99 $0.00 (c) 999 IEEE 5

6 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences 999 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences Figure 3 Change of epuaion wih respec o value of he oher user and he weigh received. Figure 4 Change of epuaion wih respec o he value of he wo users. Figure 5 Change of epuaion wih respec o he value of he user raed and he weigh received. Figure 6 Change of epuaion wih respec o he value of he oher user and he weigh received. 6. esuls While we are sill gahering real daa from our experimen wih Kasbah, we ran some simulaions in order o es our sysem. The four figures above represen he resuls from some preliminary simulaions we have run in order o evaluae our proposed soluion for Sporas. Figure 6 shows he resuls of an older version of Sporas, where he repuaion value is calculaed in he same way as in Hisos, where only he las θ raings coun owards he oal curren repuaion value of he user. This is how we acually calculae he repuaion values in each sep in Hisos, wihou showing he effec of ransiive perspecives. In he firs hree graphs we calculaed he repuaion as follows: + = Φ θ D oher ( ) i i Wi Equaion 4 Simulaion formula Figure 3 and Figure 6 give he change of he epuaion of a user A, wih average repuaion (500), raed by 20 users wih repuaions varying from 50 o The graph shows how much he repuaion of he user would change if he/she received any raing beween 0. and. Figure 4 gives he change of he repuaion of a user A, who receives an average raing wih respec o his/her own repuaion and he user B who raes A /99 $0.00 (c) 999 IEEE 6

7 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences 999 Proceedings of he 32nd Hawaii Inernaional Conference on Sysem Sciences The graph shows how he change in he repuaion of A varies if he repuaions of users A and B vary from 50 o Figure 5 gives he change in he repuaion of a user A, if A is raed by a user B wih an average repuaion (500), wih respec o he previous repuaion of user A and he raing which user B gives o user A. Like he wo previous cases he ranking of user B varies from 50 o 3000, and he weigh B gives o A varies from 0. and. These graphs demonsrae he desired behavior by saisfying all he desideraa in a pairwise repuaion sysem for an online communiy. As we can see from Figure 4 and Figure 5, even if he user giving he feedback has a very high repuaion, he/she canno affec significanly he repuaion of a user wih an already very high repuaion. However if he user raed has a low repuaion raing, he/she occurs much more significan updaes whenever he/she receives a new feedback. In Figure 4 we can also see ha when he user giving he feedback has a very low repuaion, he effec of he raing is very small unless he user being raed has a very low repuaion value him/herself. In his case he effec is acually negaive for he user being raed. In Figure 5 we observe exacly he same phenomena wih respec o he weigh given as feedback. 7. Conclusion Collaboraive filering mehods have been around for some years now, bu hey have focused on conen raing and selecion. We have developed wo collaboraive repuaion mechanisms ha esablish repuaion raings for he users hemselves. Incorporaing repuaion mechanisms in online communiies may induce social changes in he way users paricipae in he communiy. We have discussed desideraa for repuaion mechanisms for online communiies and presened 2 sysems ha were implemened in Kasbah, an elecronic markeplace. However, furher esing is required in order o evaluae he effecs of our mechanisms in boh he domains of Elecronic Commerce and online communiies. We are currenly developing a web-based gaeway ha will ac as he repuaion server of a lisserv lis. [3] OnSale hp:// [4] Sephen Paul Marsh, Formalising Trus as a Compuaional Concep, PhD Thesis, Universiy of Sirling, April 994. [5] Joseph M. eagle Jr., Trus in a Crypographic Economy and Digial Securiy Deposis: Proocols and Policies, Maser Thesis, Massachuses Insiue of Technology, May 996. [6] Beer Business Bureau, hp:// [7] Fair Isaak Co, hp:// [8] Lenny Foner, Yena: A Muli-Agen, eferral Based Machmaking Sysem, The Firs Inernaional Conference on Auonomous Agens (Agens '97), Marina del ey, California, February 997. [9] ohi Khare and Adam ifkin, "Weaving a Web of Trus", summer 997 issue of he World Wide Web Journal (Volume 2, Number 3, Pages 77-2). [0] Paul esnick and Jim Miller, PICS: Inerne Access Conrols Wihou Censorship, Communicaions of he ACM, 996, vol. 39(0), pp [] ecreaional Sofware Advisory Council hp:// [2] Nelson Minar and Alexandros Moukas, Friend of a Friend Finder, Sofware Agens Group, MIT Media Laboraory, Working paper 998 [Firefly] hp:// [3] Upendra Shardanand and Paie Maes, "Social Informaion Filering: Algorihms for Auomaing 'Word of Mouh, " Proceedings of he CHI-95 Conference, Denver, CO, ACM Press, 5/995. [4] Paul esnick, Neophyos Iacovou, Miesh Suchak, Peer Bergsrom, John iedl, GroupLens: An Open Archiecure for Collaboraive Filering of Nenews, from Proceedings of ACM 994 Conference on Compuer Suppored Cooperaive Work, Chapel Hill, NC: Pages [5] Elo, A. E., "The aing of Chessplayers, Pas and Presen", Arco Publishing, Inc. (New York) 978. [6] Mark E. Glickman, Paired Comparison Models wih Time- Varying Parameers, PhD Thesis, Harvard Universiy, May 993. [7] Simson Garfinkel. PGP: Prey Good Privacy, O'eilly and Associaes, eferences [] Anhony Chavez and Paie Maes. "Kasbah: An Agen Markeplace for Buying and Selling Goods". Proceedings of he Firs Inernaional Conference on he Pracical Applicaion of Inelligen Agens and Muli-Agen Technology (PAAM'96). London, UK, April 996 [2] ebay hp:// /99 $0.00 (c) 999 IEEE 7