Measures on the Robustness and Accuracy of Collaborative

ImpactofRelevanceMeasuresontheRobustnessandAccuracyofCollaborativeFiltering⋆JJSandvig,BamshadMobasher,andRobinBurkeCenterforWebIntelligenceSchoolofComputerScience,TelecommunicationsandInformationSystemsDePaulUniversity,Chicago,Illinois,USA{jsandvig,mobasher,rburke}@ersthatcannotbereadilydistinguishedfromordinaryusersmayinjectbiasedprofiles,ndarduser-basedcollabora-tivefipaper,weexaminerelevancemeasuresthatcomplementneigh-borsimilarityandtheirinflicular,weconsidertwotechniques,significanceweightingandtrustweighting,thatattempchniqueshavebeenusedtoimprovepredictionaccuracyincollaborativefithatsignificanceweighting,inparticular,alsoresultsinimprovedrobustnessunderprofileinjectionattacks.1IntroductionAnadaptivesystemdependentonanonymous,unauthenticateduserprofindardcollaborativefilteringalgorithmbuildsarecommendationforatargeicioususerinjectstheprofiledatabasewithanumberoffictitiousidentities,suchattacksprofileinjectionattacks(alsoknownasshilling[1]).Recentresearchhasshownthatsurprisinglymodestat-tacksaresufficienttomanipulatethemostcommonCFalgorithms[2,1,3].Suchattacksdegradetheobjectivityandaccuracyofarecommendersystem,paperweeicular,weexaminevariantsthatcombilevanceweightingtechniquesapplyaweighttoeachneighbor’ssimilarityscore,basedonsome

valuereflsontwotypesofrelevancemeasures:signifificanceweighting[4]takesthesizeofprofievents-basedweighting[5]estimatestheutilityofaneighborasaratingpredictorionaluser-basedcollaborativefilteringalgorithmsfocusexclusivelyonthedegreeofsimilaritybetwer,the“reliability”oftheneighborprofimple,duetothesparsityofthedata,thesimilaritiesmayhavebeenobtainedbasedonveryfewco-rateditemsbetwrly,unreliableneighborsthathavemadepoorpredictionsintheptheapproachestorelevanceweightingmentionedabovewere,therefore,initiallyintroducedinordertoimprovethepredictionaccuracyinuser-basedcollaborativefirust-basedmodel[5]anexplicittrustvalueiscomputedforeachuser,reflectingthe“reputation”snotlimitedtothemacroprofilelevel,andcanbecalculatstvalues,inturn,[5],O’DonovanandSmythfurtherstudiedtheimpactoftrustweightingapproachontherobustnessofcollaborativerecommetherhand,thesignificanceweight-ingapproach,introducedinitiallyin[4],doesnotfocusontrust,butratheronthenumberofco-rateditemsbetweenthetargetuserandtheneighborsasameasureforthedegreeofreliabilityoftheneighborprofiproachhasbeenshowntohaveasignificantimpactontheaccuracyofpredictions,ghtheseandothersimilarapproacheshavebeenusedtoimprovethepredictionaccuracyofrecommendersystems,theimpactofneighborsignifi-canceweightingonalgomarycontributionofthispaperistodemonstratethatrelevanceweightingisanimportantfactorindeterminingtherobustnessofacollaborativefinganoptimalultsshowthatsignificanceweighting,inparticular,isnotonlymoreaccurate;italsoimprovesalgorithmrobustnessunderprofileinjectionattacksthathavecompactprofilesignatures.2AttacksinCollaborativeRecommendersWeassumethatanaybeintheformofanincreasednumberofrecommen-

dationsfortheattacker’sproduct,orfewerrecommendationsforacompetitor’borativerecommenderdatabaseconsistsofmanyuserprofiles,eachwithassignedratingstoanumberofproductsthatrepresenttheuser’-basedcollaborativefilteringalgorithmsattempttodiscoveraneigh-borhoodofuserprofigvalueispredictedforallmissingitemsinthetargetuser’sprofile,dlistisproduced,ndardk-nearestneighboralgorithmiswidelyusedandreasonablyaccurate[4].SimilarityiscomputedusingPearson’scorrelationcoefficient,andthekmostsimpliesatargetusermayhaveadiffsocommontofilterneighborswithsimilaritybelowaspecifidentifyinganeighborhood,weuseResnick’salgorithmtocomputethepredictionforatargetitemiandtargetuseru: