merge.docx

merge.docx

《merge.docx》由会员分享，可在线阅读，更多相关《merge.docx（12页珍藏版）》请在冰点文库上搜索。

merge

Animprovedexplicittrust-basedmethod

ABSTRACT

Collaborativefiltering（CF）iswidelyusedtechniquetogenerate

recommendations,butitsuffersfromtwoissues:

datasparsity（itreferstotheproblemthattheuserscanrateonlyalimitednumbersofitems）andcoldstart（itreferstothedifficultytobootstrappingtheRecommender

Systemsforthenewusersandnewitems）.Trust-awareRecommender

Systemsmakecontributiontoaddressingthetwoissueseffectively,butit’salsoabigchallengetomergethetrustinformationintoCFtoachievebetterperformance.Inthispaper,weprovideanimprovedtrustvaluecomputationmethodtoincorporatetheimplicittrustwithexplicittrust.Ourmainideaistoutilizethetargetuser’sreviewsratingtoinferitsimplicittrustinthesystem,whichenhancetheexplicittrust-basedtechniquesbymergingusers’interactionexperiencewiththetargetuser’scontributiontoalloftheuserswhorateitsreviews.Wededicatetoresolvethepreviouslymentionedproblems.Experimentalresultsshowourmethodiseffectiveintermsofaccuracywhilepreservingagoodcoverage.

1Introduction

Inrecentyears,trusthasbeenstudiedinnumbersofresearchand

ithasbeenprovedtobeeffectiveinreducinginformationoverload.Severaltrust-basedmethodshavebeenproposedtoaddressthedatasparsityandrecommendationaccuracyproblem[9,].Trust-awarerecommendersystemsallowausertorateotherusers,thesystemcanquicklymakerecommendationsusingexplicitneighborset,andatrustmetric[8]reliesonaweboftrustfordefiningavalueforhowmuchausercantrustotherusersinthesystem.Thereareseveralapproachesarepresentedusingexplicittruststatementsin[10,3,11].TidalTrustisanalgorithmforinferringtrustrelationships,whichisusedin[10].Theactiveuserasksitstrustedneighborsforatrustratingforthetargetuser,andthencalculatesaweightedaverageoftrustratingfromtheneighborstothetargetuser.TidalTrust[1]isusedasthebasisforgeneratingpredictiveratingspersonalizedforeachuser.Theaccuracyoftherecommendedratingsisshowntooutperformbothasimpleaverageratingandtheratingsproducedbyacommonrecommendersystemalgorithm.RayandMahantiarguethattrustedneighborsmayhavedifferentpreferencesandbyremovingthetrustlinkswithlowsimilaritycanfurtherimprovetheperformance[2].Anothermethodistomergetheratingsoftrustedneighborstorepresentthepreferenceoftheactiveuser[3],andaccordingtothemergedratingsetfindsimilarusers,thenincorporatethesimilarusersandtrustedneighbors’ratingstopredicttheratingoftheactiveusers.Experimentalresultsonreal-worlddatasetsshowthatexistingtrustmetricscannotprovidesatisfyingperformance[6],andindicatethatfuturemetricsshouldbedesignedmorecarefully.

Ontheotherhand,implicittrustisbasedonthepastratingbehaviorofindividualprofilesratherthantheusers’directexperience.

TherearemanyotherapproachesusingImplicittrustinferredfromuserpastbehaviors,JO’Donovan&BSmythproposeanumbersofcomputationalmodelsoperatingattheprofile-level（averagetrustfortheprofileoverall）andattheprofile-item-level[4]（averagetrustforaparticularprofilewhenitcomestomakingrecommendationsforaspecificitem）.Theydescribehowtrustinformationcanbeincorporatedintotherecommendationprocessanddemonstratethatithasapositiveimpactonrecommendationquality.NLathia,SHailes&LCaprapresentatrustedk-nearestrecommendersalgorithm,itallowstheuserexploittheratinginformationoftheotherusers,fromwhichtheycanlearnwhoandhowmuchtotrustoneanother.GuibingGuo1,JieZhang1,DanielThalmann1,AnirbanBasu2,NeilYorke-Smithconductanempiricalstudytoexploretheabilityoftrustmetrics[6]todistinguishexplicittrustfromimplicittrustandtogenerateaccuratepredictions.

Explicittrustreflectsthedegreetowhichtheusers’satisfactionswithotherstheyhavedirectexperience.Butinlarge-scaleenvironments,directexperienceisoftennotsufficientorevennon-existent.Althoughspecifictrustvaluesarepossibleinrealsystems,theamountoftrustinformationisrelativelylittlecomparedtothenumberofratings.Ontheotherhand,Insuchcases,predictionisbasedonuser’s“indirectexperience”–opinionsobtainedfromotheragents.Thesemethodsusingexplicittrustbasedonthetransitiveintermofthetopicoftrustwhichisinherentlydoubtful,andwhetheritiseffectiveornot,trustpropagationdoesnotprovidesignificantbenefits.Previouslymentionedapproachescannotimprovetheaccuracyaswellascoverage.

Inthispaper,ourmainideaistofindaneffectivemethodtoimprovetheperformanceoftheexplicittrust-basedapproachbymergingtheimplicittrustinformationaccordingtothereviewsratingofthetargetuser.Thisistosay,wecomputetheimplicittrustvalueofthetargetuserifitisinthetrustedneighborsetoftheactiveuserotherthansimplycalculatefromthetrustedfriends.Moreover,wesearchfortheoptimalportiondistributingtheexplicitandimplicittrust.

2TheMergeMethod

Animportantclassificationoftrustmetricsisinglobalandlocalones.Localtrustmetricstakeintoaccounttheverypersonalandsubjectiveviewsoftheusersandpredictdifferentvaluesoftrustmetricsinotherusersforeverysingleuser.Insteadglobaltrustmetricspredictaglobal“reputation”valuethatapproximateshowmuchthecommunityasawholeconsidersacertainuser[7].Andinourpaper,wetaketheexplicittrustasthelocaltrustconsideringpersonalbias.Animportantfactwemusttakeintoaccountistheratingranges,duetothediversityofusers’habits,moodandcontexts,therangeofratingsgivenbyauserisprobablydifferentforanother.InadditionthattheExtendedEpinionsdatasetweuseonlycontaintwovalues-1fordistrustand1fortrust,whichcannotfullyexpresstherangeoftrustworthyoftheusers’towardsanother.Herewetaketheimplicittrustastheglobaltrusttocopewiththecold-startproblemfortrustedneighbors,weproposeanapproachtoinferimplicittrustfromusers’ratingprofiles,asimplicitissymmetric,whileoneaspectoftrustisasymmetric,mergingexplicitandimplicittrustcanovercometheweaknessaswellascoldstaruser.

2.1Implicittrustmeasurement

Inthispaper,weprovideanimprovedmethodmergingtheexplicitandimplicittrusttoimprovetheoverallperformanceofrecommendationandmitigatethecold-startproblem.Firstlyinthetrustnetworkwecangettheactiveuser’sdirecttrustneighborssetandthetrustvalues.Thenwecomputetheimplicittrustvaluescorrespondingtothesetandmergetheexplicitandimplicittrusttogetthetrustvalues,finallyweaveragetheratingsaccordingtothetrustvaluesofthetrustedneighbors.

ThedatasetweexploitinourexperimentsisderivedfromtheEWebsite,ExtendedEpinionsdateset.UserscanexpresstheirWebofTrust,reviewswhosereviewsandratingstheyhaveconsistentlyfoundtobevaluable.Inferringfromtheusers’pastbehavior,itisverylikelythatthemorehelpfultheusers’reviewsare,themoretrusttheymayreceiveastheymaybegoodatusingthesystemandmakingcorrectrecommendations,andwetendtobelievetheyaremoretrustworthybecauseofthecontributionfortheotherstheymake.

Sowedefineatrustmetricbasedonthereviewsratingwhichisreflectionofthehelpfulnesstheactiveuserathinkthetargetuserbas.Therangeofratingscoresis1to5（1-Nothelpful,2-SomewhatHelpful,3-Helpful4-VeryHelpful5-MostHelpful）,whilewecomputethehelpfulnessofthetargetuserforallitsreviewsinthesystem,andweregardthehelpfulnessasthetrustworthinessreceivedfromtheuserwhohaveratingforthereviews.itisalsotheimplicittrustorglobaltrustaswedefined.Findingthetrustneighborsetcorrespondingtowhichwesearchthereviewratingoftheeachneighbor.TakingneighboruserT1forexample,itiseasyforustolookupthereviewsratingscoresRi（i=1,2,3,4,5forthereviewratingscale）andallofitsreviewnumberninthesystem.NiisthenumberofratingRi.Wedefineasimplemethodfortheimplicittrustasfollows:

（1）

Theamountoftrustthatuserbpossessisequivalenttotheproportionoftimesthatbgenerateshelpfulrecommendationsovertheperformanceinthesystem.Thatistosay,theideahereisnottocomputeuserb’scontributiontoacertainuser,buttoalltheusersthatithavemaderecommendations.Wecallthistrustvalueasb’sglobaltrust.

2.2mergetheimplicitandexplicittrust

Giventheusersetin

whichrepresentsdirecttrustedneighborsoftheactiveusera,wethencomputetheeveryusers’implicittrustvalueas[1]correspondingtotheset.Morespecifically,weadoptasimplyandeffectivemergingmethodasfollows:

（2）

Where

istheoveralltrustvalue,andDTistheexplicittrustspecifiedbytheactiveusera,IDTistheimplicittrustvalueofbintermsofthesystem.

isaparameterwhichweneedtoexploretheeffectoftheparameter,wetunethevalue

from0to0.9withstep0.1.Finally,thepredictedratingpofitemifortheactiveuseraisgeneratedbyaveragingtheratingsaccordingtothetrustvaluesofthetrustedneighborsspecifiedbytheactiveusera,asfollows:

p=

（3）

Theneighborswhoaretrustedmorebyuseraandhaveahigherglobaltrustvaluewillhavehigherimpactonthepredictedratings.

2.3Analysisofthemergemethod

Onecommoncharacteristicofthedatasparsityandcold-startproblemsisthatthesmallnumberoftrustinformationspecifiedbytheactiveusermakeitdifficu