A Model of a Trust-based Recommendation System on

arXiv:nlin/0611054v3[nlin.AO]18Sep2007FrankE.Walter,StefanoBattiston,andFrankShweitzer:AModelofaTrust-basedReommendationSystemonaSoialNetworkRevisedManusriptforJAAMAS(September07,2007).AModelofaTrust-basedReommendationSystemonaSoialNetworkFrankE.Walter,StefanoBattiston,andFrankShweitzerChairofSystemsDesign,ETHZurih,Kreuzplatz5,8032Zurih,Switzerlandfewalterethz.h,sbattistonethz.h,fshweitzerethz.hAbstratInthispaper,wepresentamodelofatrust-basedreommendationsystemonasoialnet-work.Theideaofthemodelisthatagentsusetheirsoialnetworktoreahinformationandtheirtrustrelationshipstolterit.Weinvestigatehowthedynamisoftrustamongagentsaettheperformaneofthesystembyomparingittoafrequeny-basedreommendationsystem.Furthermore,weidentifytheimpatofnetworkdensity,prefereneheterogeneityamongagents,andknowledgesparsenesstoberuialfatorsfortheperformaneofthesystem.Thesystemself-organisesinastatewithperformaneneartotheoptimum;theperformaneonthegloballevelisanemergentpropertyofthesystem,ahievedwithoutexpliitoordinationfromtheloalinterationsofagents.Keywords:ReommenderSystem,Trust,SoialNetwork1IntrodutionandMotivationInreentyears,theInternethasbeomeofgreaterandgreaterimportaneineveryone’slife.Peopleusetheiromputersforommuniationwithothers,tobuyandsellprodutson-line,tosearhforinformation,andtoarryoutmanymoretasks.TheInternethasbeomeasoialnetwork,linkingpeople,organisations,andknowledge[33℄andithastakentheroleofaplatformonwhihpeoplepursueaninreasingamountofativitiesthattheyhaveusuallyonlydoneinthereal-world.Thisdevelopmentonfrontspeoplewithaninformationoverload:theyarefaingtoomuhdatatobeabletoeetivelylteroutthepieesofinformationthataremostappropriateforthem.TheexponentialgrowthoftheInternet[21℄impliesthattheamountofinformationaessibletopeoplegrowsatatremendousrate.Historially,peoplehaveinvarioussituationsalreadyhadtoopewithinformationoverloadandtheyhaveintuitivelyappliedanumberofsoialmehanismsthathelpthemdealwithsuhsituations.However,manyofthese,inludingthenotionoftrust,donotyethaveanappropriatedigitalmapping[23℄.Findingsuitablerepresentationsforsuhoneptsisatopiofon-goingresearh[1;10;17;23;27;29℄.Theproblemofinformationoverloadhasbeeninthefousofreentresearhinomputersieneandanumberofsolutionshavebeensuggested.Theuseofsearhengines[9℄isoneapproah,butsofar,theylakpersonalisationandusuallyreturnthesameresultforeveryone,eventhough1/22FrankE.Walter,StefanoBattiston,andFrankShweitzer:AModelofaTrust-basedReommendationSystemonaSoialNetworkRevisedManusriptforJAAMAS(September07,2007).anytwopeoplemayhavevastlydierentprolesandthusbeinterestedindierentaspetsofthesearhresults.Adierentproposedapproaharereommendationsystems[24;25;26;30℄.Inthefollowing,wepresentamodelofatrust-basedreommendationsystemwhih,inanautomatedanddistributedfashion,ltersinformationforagentsbasedontheagents’soialnetworkandtrustrelationships[15;19;25℄.Trustisatopiwhihhasreentlybeenattratingresearhfrommanyelds,inluding,butnotlimitedto,omputersiene,ognitivesienes,soiology,eonomis,andpsyhology.Asaresultofthis,thereexistsaplethoraofdenitionsoftrust,somesimilartoeahother,somedierentfromeahother.Intheontextofourmodel,trustanbedenedastheexpetanyofanagenttobeabletorelyonsomeotheragent’sreommendations.Therearemanyareasofappliationinwhihsuhsystems,orsimilarones,areappliable:someobviousexampleswouldbethefailitiestoshareopinionsand/orratingsthatmanyshoppingorautioningwebsitesoer,butthesamepriniplesofombiningsoialnetworksandtrustrelationshipsanbeappliedinotherdomainsaswell:forexample,inthesientiommunity,informofareommendationsystemforjournal,onferene,andworkshopontributions.ThemodelthatwearegoingtopresentenablesaquantitativestudyoftheproblemandalsoprovidesaskethforasolutionintermsofarealInternetappliation/webservie.Theideaattheoreofthemodelisthatagents•leveragetheirsoialnetworktoreahinformation;and•makeuseoftrustrelationshipstolterinformation.Inthefollowing,wedesribethemodelandtheresultsobtainedbysimulatingthemodelwithmulti-agentsimulations.Tosomeextent,itisalsopossibletomakeanalytialpreditionsoftheperformaneofthesystemasafuntionofthepreferenesoftheagentsandthestrutureofthesoialnetwork.Theremainderofthepaperisorganisedasfollows:inthefollowingsetion,weputourworkintotheontextoftherelatedwork.Then,wepresentourmodelofatrust-basedreommen-dationsystemonasoialnetwork.Thisisfollowedbyananalysisoftheresultsfromomputersimulationsandanalytialapproximationsofthemodel.Subsequently,weillustrateanumberofpossibleextensionsandonludewithasummaryofthework.2RelatedWorkReentresearhinomputersienehasdealtwithreommendationsystems[26;30℄.Suhsys-temsmostlyfallintotwolasses:ontent-basedmethodssuggestitemsbymathingagentproleswithharateristisofprodutsandservies,whileollaborativelteringmethodsmeasurethe2/22FrankE.Walter,StefanoBattiston,andFrankShweitzer:AModelofaTrust-basedReommendationSystemonaSoialNetworkRevisedManusriptforJAAMAS(September07,2007).similarityofpreferenesbetweenagentsandreommendwhatsimilaragentshavealreadyhosen[30℄.Often,reommendationsystemsareentralisedand,moreover,theyareoeredbyentit