Collaborative 64257ltering the most success ful recommendation approach makes recommendations based on past transactions and feedback from consumers sharing similar interests A major problem limiting the usefulness of collaborative 64257ltering is t ID: 35211
Download Pdf The PPT/PDF document "Applying Associative Retrieval Technique..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
ApplyingAssociativeRetrievalTechniquestoAlleviatetheSparsityProbleminCollaborativeFilteringZANHUANG,HSINCHUNCHEN,andDANIELZENGTheUniversityofArizona Recommendersystemsarebeingwidelyappliedinmanyapplicationsettingstosuggestproducts,services,andinformationitemstopotentialconsumers.CollaborativeÞltering,themostsuccess- Thisresearchwassupportedinpartbythefollowinggrants:NSFDigitalLibraryInitiative-II,ÒHigh-PerformanceDigitalLibrarySystems:FromInformationRetrievaltoKnowledgeMan-agement,ÓIIS-9817473,April1999ÐMarch2002,andNSFInformationTechnologyResearch,ÒDevelopingaCollaborativeInformationandKnowledgeManagementInfrastructure,ÓIIS-0114011,September2001ÐAugust2004.D.ZengisalsoafÞliatedwiththeKeyLabofComplex AssociativeRetrievalTechniquesfortheSparsityProblemCategoriesandSubjectDescriptors:H.1.2[ModelsandPrinciples]:User/Machinesystemshumaninformationprocessing;H.3.3[InformationStorageandRetrieval]:InformationSearchandRetrievalinformationÞltering;relevancefeedback;retrievalmodelsGeneralTerms:Algorithms,Design,ExperimentationAdditionalKeyWordsandPhrases:Recommendersystem,collaborativeltering,sparsityproblem,associativeretrieval,spreadingactivation 1.INTRODUCTIONRecommendationasasocialprocessplaysanimportantroleinmanyappli-cationsforconsumers,becauseitisoverlyexpensiveforeveryconsumertolearnaboutallpossiblealternativesindependently.Dependingonthespeciapplicationsetting,aconsumermightbeabuyer(e.g.,inonlineshopping),aninformationseeker(e.g.,ininformationretrieval),oranorganizationsearchingforcertainexpertise.Inaddition,recommendationasapersonalizedmarket-ingmechanismhasrecentlyattractedsignicantindustryinterest(e.g.,onlineshoppingandadvertising).Recommendersystemshavebeendevelopedtoautomatetherecommenda-tionprocess.Examplesofresearchprototypesofrecommendersystemsare:[Terveenetal.1997],SyskillsandWebert[PazzaniandBillsus1997],Fab[BalabanovicandShoham1997],and[Konstanetal.1997;Sarwaretal.1998].ThesesystemsrecommendvarioustypesofWebresources,onlinenews,movies,amongothers,topotentiallyinterestedparties.Large-scalecommercialapplicationsoftherecommendersystemscanbefoundatmanye-commercesites,suchas,andMovieFinderThesecommercialsystemsrecommendproductstopotentialconsumersbasedonprevioustransactionsandfeedback.Theyarebecomingpartofthestan-darde-businesstechnologythatcanenhancee-commercesalesbyconvert-ingbrowserstobuyers,increasingcross-selling,andbuildingcustomerloyalty[Schaferetal.2001].Oneofthemostcommonly-usedandsuccessfulrecommendationapproachesisthecollaborativelteringapproach.[Hilletal.1995;Resnicketal.1994;ShardanandandMaes1995].Whenpredictingthepotentialinterestsofagivenconsumer,suchanapproachrstidentiesasetofsimilarconsumersbasedonpasttransactionandproductfeedbackinformationandthenmakesapredictionbasedontheobservedbehaviorofthesesimilarconsumers.Despiteitswidespreadadoption,collaborativelteringsuffersfromseveralmajorlimitationsincludingsparsity,systemscalability,andsynonymy[Sarwaretal.2000a].Inthisarticle,wefocusonthesparsityproblem,whichreferstothelackofpriortransactionalandfeedbackdatathatmakesitdifcultandunreliabletopredictwhichconsumersaresimilartoagivenconsumer.Forinstance,therecommendersystemsusedbyonlinebookstoresusepastpurchasinghistorytogroupconsumersandthenmakerecommendationstoanindividualcon-sumerbasedonwhattheotherconsumersinthesamegrouphavepurchased.Whensuchsystemshaveaccessonlytoasmallnumberofpasttransactionrecords(relativetothetotalnumbersofthebooksandconsumers),however,ACMTransactionsonInformationSystems,Vol.22,No.1,January2004. AssociativeRetrievalTechniquesfortheSparsityProblemlteringhasbeenthemostsuccessfulrecommendationsystemapproachtodate[Sarwaretal.2000a]andhasbeenwidelyappliedinvariousapplications[Burke2000;Claypooletal.1999;Mobasheretal.2000;Nasraouietal.1999;Pazzani1999;Sarwaretal.1998].Despiteitssuccessinmanyap-plicationsettings,thecollaborativelteringapproachneverthelesshasbeenreportedtohaveseveralmajorlimitationsincludingthesparsity,scalability,andsynonymyproblems[Sarwaretal.2000b].Thesparsityproblemoccurswhentransactionalorfeedbackdataissparseandinsufcientforidentifyingneighborsanditisamajorissuelimitingthequalityofrecommendationsandtheapplicabilityofcollaborativelteringingeneral.Ourstudyfocusedonde-velopinganeffectiveapproachtomakinghigh-qualityrecommendationsevenwhensufcientdataisunavailable.Thenextsectionwilldiscussthesparsityproblemindetail.2.2TheSparsityProblemIncollaborativelteringsystems,usersorconsumersaretypicallyrepresentedbytheitemstheyhavepurchasedorrated.Forexample,inanonlinebookstoreselling2millionbooks,eachconsumerisrepresentedbyaBooleanfeaturevec-torof2millionelements.Thevalueforeachelementisdeterminedbywhetherthisconsumerhaspurchasedthecorrespondingbookinpasttransactions.Typ-icallythevalueof1indicatesthatsuchapurchasehadoccurredand0indicatesthatnosuchpurchasehasoccurred.Whenmultipleconsumersareconcerned,amatrixcomposedofallvectorsrepresentingtheseconsumerscanbeusedtocapturepasttransactions.Wecallthismatrixtheproductinter-actionmatrix.Thegeneraltermisusedtorefertothismatrixasopposedtothemorespecipurchasingbecausethereareothertypesofrelationssuchasexplicitandimplicitratingsbetweenconsumersandproductsforgeneralrecommendersystems.Wenowintroducesomenotationtobeusedthroughoutthearticle.Weusetodenotethesetofconsumersandthesetofitems.Wedenotetheconsumerproductinteractionmatrixbya),suchthat1,ifuserpurchaseditem0,otherwiseNotethat,inourstudy,wefocusedonactualtransactionsthatoccurred,isbinary.Inotherrecommendationscenariossuchasthosethatinvolveratings,cantakeothercategoricalorcontinuousvalues(e.g.,5-levelratingscalesandprobabilitiesofinterest).Inmanylarge-scaleapplicationssuchasmajore-commercewebsites,boththenumberofitems,,andthenumberofconsumers,,arelarge.Insuchcases,evenwhenmanytransactionshavebeenrecorded,theconsumerinteractionmatrixcanstillbeextremelysparse,thatis,thereareveryfewele-mentsinwhosevalueis1.Thisproblem,commonlyreferredtoasthesparsityproblem,hasamajornegativeimpactontheeffectivenessofacollaborativeteringapproach.Becauseofsparsity,itishighlyprobablethatthesimilarity(orcorrelation)betweentwogivenusersiszero,renderingcollaborativeACMTransactionsonInformationSystems,Vol.22,No.1,January2004. Z.Huangetal.useless[BillsusandPazzani1998].Evenforpairsofusersthatarepositivelycorrelated,suchcorrelationmeasuresmaynotbereliable.problemfurtherillustratestheimportanceofaddressingthesparsityproblem.Thecold-startproblemreferstothesituationinwhichanewuseroritemhasjustenteredthesystem[Scheinetal.2002].Collaborativelteringcannotgenerateusefulrecommendationsforthenewuserbecauseofthelackofsufcientpreviousratingsorpurchases.Similarly,whenanewitementersthesystem,itisunlikelythatcollaborativelteringsystemswillrecom-mendittomanyusersbecauseveryfewusershaveyetratedorpurchasedthisitem.Conceptually,thecold-startproblemcanbeviewedasaspecialinstanceofthesparsityproblem,wheremostelementsincertainrowsorcolumnsoftheproductinteractionmatrixare0.Manyresearchershaveattemptedtoalleviatethesparsityproblem.Sarwaretal.[2001]proposedanitem-basedapproachtoaddressingboththescalabilityandsparsityproblems.Basedonthetransactionalorfeedbackdata,itemsthataresimilartothosepurchasedbythetargetuserinthepastareidentiandthenrecommended.Itemsimilaritiesarecomputedasthecorrelationsbetweenthecorrespondingcolumn(item)vectors.Itisreportedthatincertainapplicationsthisitem-basedapproachachievedbetterrecommendationqualitythantheuser-basedapproach,thepredominantapproachusedinrecommendersystems,whichreliesoncorrelationsbetweenrow(user)vectors.Anotherproposedapproach,dimensionalityreduction,aimstoreducethedimensionalityoftheconsumerproductinteractionmatrixdirectly.Asimplestrategyistoformclustersofitemsorusersandthenusetheseclustersasbasicunitsinmakingrecommendations.Moreadvancedtechniquescanbeap-pliedtoachievedimensionalityreduction.ExamplesarestatisticaltechniquessuchasPrincipleComponentAnalysis(PCA)[Goldbergetal.2001]andinfor-mationretrievaltechniquessuchasLatentSemanticIndexing(LSI)[BillsusandPazzani1998;Sarwaretal.2000b].Empiricalstudiesindicatethatdi-mensionalityreductioncanimproverecommendationqualitysignicantlyinsomeapplications,butperformspoorlyinothers[Sarwaretal.2000b].Thedimensionalityreductionapproachaddressesthesparsityproblembyremov-ingunrepresentativeorinsignicantconsumersorproductstocondensetheproductinteractionmatrix.However,potentiallyusefulinformationmightbelostduringthisreductionprocess.Thismaypartiallyexplainthemixedresultsreportedontheperformanceofdimensionalityreduction-basedlteringapproaches.Researchershavealsoattemptedtocombinecollaborativelteringwithcontent-basedrecommendationapproachestoalleviatethesparsityproblem[BalabanovicandShoham1997;Basuetal.1998;Condliffetal.1999;Goodetal.1999;Huangetal.2002;Pazzani1999;Sarwaretal.1998].Suchanap-proachconsidersnotonlypastconsumerproductinteractionsbutalsosimilar-itiesbetweenproductsoritemsdirectlyderivedfromtheirintrinsicpropertiesorattributes.Werefertothisapproachasthehybridapproach.Mostpreviousstudiesusingthehybridapproachhavedemonstratedsignicantimprovementinrecommendationqualityovertheuser-basedapproachesdiscussedabove.However,thehybridapproachrequiresadditionalinformationregardingtheACMTransactionsonInformationSystems,Vol.22,No.1,January2004. AssociativeRetrievalTechniquesfortheSparsityProblemproductsandametrictocomputemeaningfulsimilaritiesamongthem.Inprac-tice,suchproductinformationmaybedifcultorexpensivetoacquireandarelatedsimilaritymetricmaynotbereadilyavailable.Ourresearchdealtwiththesparsityproblemunderadifferentframework.Insteadofreducingthedimensionoftheconsumerproductinteractionmatrix(thus,makingitlesssparse),weproposedtoexplorethetransitiveinter-actionsbetweenconsumersanditemstothematrixandmakeitforrecommendationpurposes.Theintuitionbehindtran-sitiveinteractionscanbeexplainedbythefollowingexample.Supposeusersboughtbookandusersboughtbook.Standardcol-lteringapproachesthatdonotconsidertransitiveinteractionswillandalsobutnot.Anapproachthatincorporatestransitiveinteractions,however,willrecognizetheassociativere-lationshipbetweenandwillsuchtransitiveinteractionsintotheconsumerproductinteractionmatrixforrecommendations.Ourresearchfocusesondevelopingacomputationalapproachtoexploringtransitiveuseranditemsimilaritiestoaddressthesparsityprobleminthecon-textofcollaborativeltering.Thenextsectionpresentsourgeneralmodelingframeworkanddiscussesexistingresearchrelatedtothecomputationandap-plicationoftransitiveassociationsinInformationRetrievalandRecommenderSystems.3.MODELINGRECOMMENDATIONASANASSOCIATIVERETRIEVALPROBLEM3.1AssociativeRetrievalandGraph-BasedModelsThepotentialvalueoftransitiveassociationshasbeenrecognizedbyre-searchersworkingintheeldofrecommendersystems[BillsusandPazzani1998;Sarwaretal.2000b].Theexplorationoftransitiveassociationsinthecontextofrecommendersystemsistypicallycarriedoutinagraph-basedrec-ommendationmodelfortworeasons.First,agraphornetwork-basedmodeliseasytointerpretandprovidesanaturalandgeneralframeworkformanydifferenttypesofapplicationsincludingrecommendersystems.Second,arichsetofgraph-basedalgorithmsisreadilyapplicablewhentherecommendationtaskisformulatedasagraph-theoreticproblem.Below,webrieysurveythreerepresentativegraph-basedmodelsthatex-ploretransitiverelationships.Aggarwaletal.[1999]introducedarecommen-dationmodelbasedonadirectedgraphofusers.Intheirmodel,adirectedlinkstartingfromuserandendingatuseresthatsbehaviorisstronglypredictiveofsbehavior.Recommendationsaremadebyexploringshort(in-dicatingstrongpredictability)pathsjoiningmultipleusers.Mirza[2001]andMirzaetal.[2003]proposedasocialnetworkgraphofuserstoproviderecom-mendations.Linksinthissocialnetworkgraphareinducedbyhammockjumpsnedbetweentwouserswhohaveagreedratingsonatleastagivennum-berofitems).BothAggarwalsandMirzasmodelsemphasizeusingthegraphofusersandonlyemployuserassociationstoexploretransitiveassociations.ACMTransactionsonInformationSystems,Vol.22,No.1,January2004. Z.Huangetal.Inourpreviousresearch,wedevelopedanothergraph-basedmodelforcollab-ltering[Huangetal.2003]whichincludesbothusersanditemsinthegraph.Thismodelwasintendedtocaptureadditionaltypesofinputsandrecommendationapproachesinauniedframework.Theabovegraph-basedmodelsprovidethebasicrepresentationalandmod-elingframeworkforourresearchonthesparsityproblemandenableustodrawananalogybetweenrecommendersystemsandassociativeretrievalsys-tems.Thisanalogy,inturn,suggeststhatthesparsityproblemcanpotentiallybedealtwitheffectivelyusingcomputationalmethods,inparticular,spread-ingactivationalgorithms,whichhavebeensuccessfullyappliedinassociativeInthissection,wediscussindetailhowtherecommendationtaskcanbeformulatedasanassociativeretrievalproblemandhowspreadingactivationalgorithmscanbeusedtoexploreusefultransitiveassociationsandthushelptosolvethesparsityproblem.Weconcludethissectionbypresentingresearchquestionsdesignedtoevaluatetheideaofapplyingspreadingactivationalgo-rithmsinthecontextofrecommendersystems.3.2CollaborativeFilteringasAssociativeRetrievalAssociativeinformationretrievalhasitsorigininstatisticalstudiesofassoci-ationsamongtermsanddocumentsinatextcollection.Thebasicideabehindassociativeretrievalistobuildagraphornetworkmodelofdocumentsandindextermsandqueries,andthentoexplorethetransitiveassociationsamongtermsanddocumentsusingthisgraphmodeltoimprovethequalityofinfor-mationretrieval.Forexample,thegeneralizedvectorspacemodel[Wongetal.1985]representsadocumentbyavectorofitssimilaritiestoallotherdocu-mentsinthecorpus.Theassociations(similarities)amongdocuments,deastransitiveassociationsthroughcommonindexterms,areconstructedanddi-rectlyusedtosupportinformationretrieval.Anumberoftechniqueshavebeenproposedtoconstructandutilizesuchnetworksofassociationsininforma-tionretrieval.Examplesofthesetechniquesarevariousstatisticalapproaches[CrouchandYang1992],neuralnetworks[JungandRaghavan1990],geneticalgorithms[Gordon1988],andspreadingactivationapproaches[CohenandKjeldsen1987;SaltonandBuckley1988].Thesimilaritybetweenassociativeretrievalandcollaborativelteringhasbeenrecognizedbysomerecentstudies[SoboroffandNicholas2000].Inasso-ciativeretrieval,documentsarerepresentedbyindexterms.Atthesametime,thesemanticsofanindextermcanalsoberepresentedbythesetofdocumentsthatcontainit.Similarly,incollaborativeltering,userspreferencescanberepresentedbytheitemsandtheirinteractionswiththeitems.Theintrinsicfeaturesofanitemcanalsoberepresentedbytheusersandtheirinteractionswithit.Thefollowingexampleillustratestheideaofexploringtransitiveassocia-tionsinrecommendersystems.UsingthenotationdevelopedinSection2.2,thepasttransactionscanberepresentedinthefollowingconsumerinteractionmatrix.ACMTransactionsonInformationSystems,Vol.22,No.1,January2004. AssociativeRetrievalTechniquesfortheSparsityProblem Fig.1.Asimpleexamplefortransitiveassociationsincollaborativeltering.Notethat,inourwork,weassumethattheonlyinformationavailabletotherecommendersystemistheabovematrix.Hence,thegraphshowninFigure1isabipartitegraph.(Inabipartitegraph,nodesaredividedintotwodistinctivesets.Linksbetweenpairsofnodesfromdifferentnodesetsareadmissible,whilelinksbetweennodesfromthesamenodesetarenotallowed.)Supposetherecommendersystemneedstorecommendproductsforcon-.Thestandardcollaborativelteringalgorithmwillmakecommenda-tionsbasedonthesimilaritiesbetweenandotherconsumers().Thesimilaritybetweenisobviousbecauseofpreviouscommonpurchases).Asaresult,isrecommendedtohaspurchasedit.Nostrongsimilaritycanbefoundbetween.Therefore,,whichhasbeenpurchasedby,willnotberecommendedtoTheaboverecommendationapproachcanbeeasilyimplementedinagraph-basedmodelbycomputingtheassociationsbetweenproductnodesandcus-tomernodes.Inourcontext,theassociationbetweentwonodesisdeterminedbytheexistenceandlengthofthepath(s)connectingthem.Standardcollaborativelteringapproaches,includingboththeuser-basedanditem-basedapproaches,consideronlypathswithlengthequalto3.Forinstance,theassociationbetweenisdeterminedbyallpathsoflength3connecting.ItiseasytoseefromFigure1thatthereexisttwopathsconnecting.Thisstrongassociationleadstotherecommendationof.Associationbetweendoesnotexistbecausenopathoflength3exists.Intuitively,thehigherthenumberofdistinctivepathsconnectingaproductnodetoaconsumernode,thehighertheassociationbetweenthesetwonodes.Theproductthereforeismorelikelytoberecommendedtotheconsumer.Extendingtheaboveapproachtoexploreandincorporatetransitiveassoci-ationsisstraightforwardinagraph-basedmodel.ByconsideringpathswhoseACMTransactionsonInformationSystems,Vol.22,No.1,January2004. Z.Huangetal.activationreceivedfromthedirectlylinkedactivenodestotheirownactiva-tionlevels.Onlyaxednumberofnodeswiththehighestactivationlevelskeeptheiractivationlevelsin).Allotherelementsof)areresettovalue0.Thecontrolparameterswereheuristicallysetto0.2and0.8inourexperiments,afterobservingseveralalgorithmruns.StoppingCondition.Thealgorithmterminatesafteraxednumberofitera-tions.Thislimitoniterationsissetto10inthecurrentimplementation.Thetop50itemnodesthathavethehighestactivationlevelsintheactivationvectorofthenalstage(10)andthathavenotbeenpreviouslypurchasedformtherecommendationforthetargetedconsumer.4.2Branch-and-BoundAlgorithmOurimplementationofthebranch-and-boundalgorithm(thereaftertheBNBalgorithm)followsthatusedinChenandNg[1995],originallydevelopedinthecontextofconceptexploration.Ourimplementationstartswithausernodecorrespondingtothetargetuser.Neighboringnodes,thatis,itemnodesthatcorrespondwiththetargetuserspreviouspurchases,arethenactivated.Theactivatednodesareputintoapriorityqueuebasedontheiractivationlevelsandhigh-prioritynodesareusedtoactivatetheirneighbors.Themainstepsoftheimplementedbranch-and-boundalgorithmaresummarizedasfollows:.Thenodecorrespondingwiththetargetuserisinitializedtohavetheactivationlevelof1.Allothernodesareinitializedwithlevel0.Apriorityqueue,,iscreatedwithonlythetargetusernodeasitsinitialmember.Aninitiallyemptyoutputqueue,,iscreatedtostoreactivatednodes.ActivationandActivationLevelComputation.Duringeachiteration,thealgo-rithmremovesthefrontnodefrom(thisnodehasthehighestlevelofactivation),activatesitsneighboringnodes,andthencomputestheseneigh-activationlevelas,where)representstheactivationlevelofthefrontnoderemovedfromrepresentstheweightofthelinkconnectingthefrontnodewithaneighboringnode(weas-signedeachlinkaweightof0.5inthecurrentimplementation),andrepresentsthenewlycomputedactivationlevelforthisneighboringnode.Ac-tivatednodesthathavenotbeenrecordedearlierinareinsertedintotheoutputqueue.Iftheyalreadyexistin,theiractivationlevelwillbeincreasedbyStoppingCondition.Theaboveactivationprocessisrepeatedforaxednum-beroftimesbeforethealgorithmendsandoutputsthetop50itemnodesfrom.Inourexperimentsweheuristicallysetthelimitonthenumberoftheiterationsto70.4.3HopÞeldNetAlgorithmTheHopeldnetalgorithm(thereaftertheHopeldalgorithm)performsapar-allelrelaxationsearchtosupportspreadingactivation.Inourcontext,thegraphmodelofcollaborativelteringmapstointerconnectedneuronsandsynapsesACMTransactionsonInformationSystems,Vol.22,No.1,January2004. AssociativeRetrievalTechniquesfortheSparsityProblemintheHopeldnetwithneuronsrepresentingusersanditemsandsynapsesrepresentinginteractionbetweenusersanditems.TheimplementedHopnetactivationalgorithmisdescribedasfollows:.Theusernodecorrespondingtothetargetuserisinitializedtohavetheactivationlevel1.Allothernodesareinitializedwithlevel0.ActivationandActivationLevelComputation.AsintheLCMalgorithm,axednumberofnodeswithhighestactivationlevelsareactivated.Theacti-vationlevelforeachnodeiscomputedas1,(6)isthecontinuousSIGMOIDtransformationfunction[Knight1990] ,(7)1)istheactivationlevelofnodeatiteration1,andistheweightofthelinkconnectingnodetonode(similartothebranch-and-boundalgorithm,weassignedeachlinkaweightof0.5).Inaccordancewith(6),eachnewlyactivatednodecomputesitsactivationlevelbasedonthesummationoftheproductsofitsneighborsactivationlevelandtheirsynapses.ThecontroloftheSIGMOIDfunctionwereheuristicallysetto10and0.8inourexperiments.StoppingCondition.Theaboveprocessisrepeateduntilcondition(8)issat-edindicatingthatthereisnosignicantchangebetweenthelasttwoiterations.Inthiscondition,isasmallpositivenumber.Notethattheallowablechangesareproportionaltothenumberofiterationsperformedtospeeduptheconvergence.Asinallotherapproaches,topitemnodesthathavethehighestactivationlevelinthenalstateofthenetworkarerecommendedafterremovingitemsalreadypurchasedbythetargetuser.5.ANEXPERIMENTALSTUDYWeconductedanexperimentusingdatafromanonlinebookstoretoevaluatetheeffectivenessoftransitiveassociation-basedcollaborativelteringandan-swertheresearchquestionsdiscussedinSection3.4.Inthissection,wedescribetheexperimentaldataandpresenttheevaluationdesignandper-formancemeasuresusedinourstudy.Wethensummarizeourexperimentalndings.5.1ExperimentDataAmajorChineseonlinebookstore(www.books.com.tw)provideduswithdatacoveringaportionofveyearsofrecenttransactions.ThisdatasetcorrespondsACMTransactionsonInformationSystems,Vol.22,No.1,January2004. Z.Huangetal. Fig.4.Over-activationeffect(withgraphsenhancedbyitemassociations).Overall,allthreespreadingactivationalgorithmsconsistentlyoutperformedthe3-hopalgorithm.TheconclusionswehavedrawnbasedontheHopalgorithmalsoholdtruefortheLCMandBNBalgorithms.InFigure3,weobserveweakover-activationeffectsofthespreadingacti-vationalgorithmsinourexperiment.Therecommendationqualityofspread-ingactivation-basedcollaborativelteringincreasedfasterthanthatofthestandardcollaborativelteringapproachbecausethetransactionaldataac-cumulatesduringtheinitialdeploymentphaseoftherecommendersystem.Therecommendationqualitymoreorlesspeaks(withnoticeabledegradations)whentheconsumerproductinteractionmatrixbecomesrelativelydense(seeG11).InFigure4,thethreealgorithmsshowsomenoticeabledifferencesinperformancewhentheunderlyinggraphisdense.Forinstance,LCMshowsamoresignicantover-activationeffect,resultinginthedeteriorationoftherec-ommendationquality.WenoticeinFigure4thattherearesomeimprovementsintheperformanceofthealgorithmsbeforetheoveralldownwardtrendsstart.Thismaybeexplainedbythebenetofincludingcontentsimilarityinformation[BalabanovicandShoham1997;Sarwaretal.1998].Asmorecontentinforma-tionisadded,itseemsthattheover-activationeffectstartstoovershadowthetofusingadditionalinformation.5.4ComputationalIssueswithSpreadingActivationAlgorithmsInthissection,wefocusoncomputationaspectsofthespreadingactivationalgorithms.Werstexaminetheimpactofcontrolparametersettingsofthethreespreadingactivationalgorithms.Wethencomparethecomputationalciencyofthesealgorithms.SensitivityofControlParameters.Intheexperimentsreportedintheprevioussection,thecontrolparametersofvariousimplementedspreadingACMTransactionsonInformationSystems,Vol.22,No.1,January2004. AssociativeRetrievalTechniquesfortheSparsityProblemWearecurrentlyextendingtheresearchreportedinthispaperinthefollow-ingareas.Weareusingadditionaldatasetswithdifferentcharacteristicstocomparetheperformancesofthespreadingactivationalgorithmswithothercollab-lteringalgorithmsstudiedinthisarticle.Forinstance,ourinitialexperimentalresultsontheMillionMoviedataset,wheretheconsumerproductinteractionmatrixismuchdenserthanthatintheonlinebookstoredatasetinthisstudy,showedthattheitem-basedapproachachievedthebestperformance,followedbytheuser-basedapproaches.Thespreadingac-tivationalgorithmperformedslightlyworsethantheuser-basedapproaches.Thisresultprovidesfurtherevidenceoftheover-activationeffectandindi-catestheimportanceofspeciccharacteristicsofthedatasetandtheirim-pactontheselectionofanappropriatecollaborativelteringapproach.Ourfutureresearchisaimedatgainingacomprehensiveunderstandingoftheapplicabilityandeffectivenessofthespreadingactivation-basedcollabora-lteringapproach.Weareintheprocessofcomparingandcombiningthespreadingactivationalgorithmswiththehybridrecommendationapproaches.Byincludingitemanduserassociationsbasedoncontent-relatedinformation(e.g.,bookcon-tent,customerdemographics,etc.),thespreadingactivationalgorithmscanbedirectlyappliedtogeneratehybridrecommendations.Ourinitialexper-imentalresultsshowedthatthespreadingactivation-basedhybridrecom-mendationperformedsignicantlybetterthanalltheotherapproaches.Wearealsoworkingonincorporatinginverseuserfrequencyandinverseitemfrequencyintoourspreadingactivationframework.Byassigningtheseasweightstothenodesinthegraphmodel,wemayimprovetherecom-mendationqualityofthespreadingactivationalgorithmsandtosomeextentalleviatetheover-activationeffect.Lastly,weareextendingthespreadingactivationframeworksoitcandealwithsystemshavingfeedbackthattakemultiplevalues(e.g.,ratings)inadditiontobinarytransactionaldata.Wewillthendirectlycompareourap-proachwithAggarwalsandMirzasgraph-theoreticalapproaches.Wearealsoextendingourframeworktoincorporatetheusersfeedbackontherec-ommendationstofurtherimprovethequalityoftherecommendationusingthespreadingactivationapproach.ACKNOWLEDGMENTSWewishtothanktheanonymousreviewersfortheirdetailedandconstructivecommentsonthetwoearlierversionsofthisarticle.Wewouldalsoliketoac-knowledgebooks.com.twforprovidinguswiththedatasetandtheirassistanceduringtheproject.GGARWAL,C.C.,W,J.L.,W,K.-L.,,P.S.1999.Hortinghatchesanegg:Anewgraph-theoreticapproachtocollaborativeltering.InProceedingsofthe5thACMSIGKDDConferenceACMTransactionsonInformationSystems,Vol.22,No.1,January2004. Z.Huangetal.onKnowledgeDiscoveryandDataMining(KDD(SanDiego,Calif.).ACM,NewYork,201,R.ARABASI,A.-L.2002.Statisticalmechanicsofcomplexnetworks.Rev.Mod.Phys.,47,J.R.1983.Aspreadingactivationtheoryofmemory.J.Verb.Learn.Verb.Behav.22ALABANOVIC,M.,Y.1997.FAB:Content-based,collaborativerecommendation.mun.ACM40,3,66,C.,H,H.,,W.1998.Recommendationasclassication:Usingsocialandcontent-basedinformationinrecommendation.InProceedingsofthe15thNationalConferenceonArticialIntelligence,714,D.,M.J.1998.Learningcollaborativeinformationlters.InProceedingsofthe15thInternationalConferenceonMachineLearning,46,J.,V,H.,,L.M.1999.Miningassociativerelationsfromwebsitelogsandtheirapplicationtocontext-dependentretrievalusingspreadingactivation.IningsoftheWorkshoponOrganizingWebSpace(WOWS).ACMDigitalLibraries99.,J.S.,H,D.,KADIE,C.1998.Empiricalanalysisofpredictivealgorithmsforcollaborativeltering.InProceedingsofthe14thConferenceonUncertaintyinArticialIn-(Madison,Wisc.).Morgan-Kaufmann,Reading,Mass.43,R.2000.Semanticratingsandheuristicsimilarityforcollaborativeltering.Iningsofthe17thNationalConferenceonArticialIntelligence,H.,V.1991.Cognitiveprocessasabasisforintelligentretrievalsystemsdesign.InformationProcessingandManagement27,5,405,H.,L,K.J.,B,K.,,D.T.1993.Generating,integrating,andactivatingthesauriforconcept-baseddocumentretrieval.IEEEExp.,Spec.SeriesArtif.Intell.Text-basedInf.Systems8,2,25,H.,D.T.1995.Analgorithmicapproachtoconceptexplorationinalargeknowledgenetwork(automaticthesaurusconsultation):Symbolicbranch-and-boundsearchvs.Connection-istHopeldnetactivation.J.ASIS46,5,348LAYPOOL,M.,G,A.,M,T.,MURNIKOV,P.,N,D.,,M.1999.Combiningcontent-basedandcollaborativeltersinanonlinenewspaper.InProceedingsoftheACMSIGIRWorkshoponRecommenderSystems.ACM,NewYork.,P.R.,R.1987.Informationretrievalbyconstrainedspreadingactivationinsemanticnetworks.InformationProcessingandManagement23,4,255,A.M.,E.F.1975.Aspreadingactivationtheoryofsemanticprocessing.Psych.Rev.82,6,407,M.K.,L,D.D.,M,D.,,C.1999.Bayesianmixed-effectsmodelsforrecommendersystems.InProceedingsoftheACMSIGIRWorkshoponRecommenderSystemsACM,NewYork.RESTANI,F.,P.L.2000.Searchingthewebbyconstrainedspreadingactivation.Inf.Proc.Manage.36,585,C.,B.1992.Experimentsinautomaticstatisticalthesaurusconstruction.InProceedingsofthe15thAnnualInternationalACMSIGIRConferenceonResearchandDevelop-mentinInformationRetrieval,(Copenhagen,Denmark).ACM,NewYork,77,K.,R,T.,GUPTA,D.,,C.2001.Eigentaste:Aconstanttimecollabo-lteringalgorithm.Inf.Ret.4,2,133,N.,S,J.,KONSTAN,J.,B,A.,SARWAR,B.,H,J.,,J.1999.Combiningcollaborativelteringwithpersonalagentsforbetterrecommendations.Iningsofthe16thNationalConferenceonArticialIntelligence,439,M.1988.Probabilisticandgeneticalgorithmfordocumentretrieval.Commun.ACM,10,1208,W.,S,L.,R,M.,,G.1995.Recommendingandevaluatingchoicesinavirtualcommunityofuse.InProceedingsoftheACMCHI95ConferenceonHumanFactorsinComputingSystems.ACM,NewYork,194UANG,Z.,C,W.,,H.2003.Agraphmodelfore-commercerecommendersystems.J.ASIST,inpress.ACMTransactionsonInformationSystems,Vol.22,No.1,January2004. AssociativeRetrievalTechniquesfortheSparsityProblemUANG,Z.,C,W.,O,T.-H.,,H.2002.Agraph-basedrecommendersystemfordigitallibrary.InProceedingsofthe2ndACM/IEEE-CSJointConferenceonDigitalLibraries(Portland,Ore.).ACM,NewYork,65,G.AGHAVAN,V.1990.Connectionistlearninginconstructingthesaurus-likeknowledgestructure.InProceedingsoftheAAAISpringSymposiumonText-basedIntelligentSystems,G.2001.Evaluationofitem-basedtop-nrecommendationalgorithms.InProceedingsofthe10thInternationalConferenceonInformationandKnowledgeManagement(CIKM),K.1990.Connectionistideasandalgorithms.Commun.ACM33,11,59ONSTAN,J.A.,M,B.N.,MALTZ,D.,H,J.L.,G,L.R.,,J.1997.Group-Lens:ApplyingcollaborativelteringtoUsenetnews.Commun.ACM40,3,77,W.,ALVAREZ,S.A.,,C.2002.Efcientadaptive-supportassociationruleminingforrecommendersystems.DataMiningKnowl.Disc.6,1,83,B.J.2001.Jumpingconnections:Agraph-theoreticmodelforrecommendersys-tems.ComputerScienceDepartment,VirginiaPolytechnicInstituteandstateuniversity,(http://scholar.lib.vt.edu/theses/available/etd-02282001-175040/unrestricted/etd.pdf).,B.J.,K,B.J.,,N.2003.Studyingrecommendationalgorithmsbygraphanalysis.J.Intel.Inf.Syst.20,2,131OBASHER,B.H.,D,T.L.,NAKAGAWA,M.,S,Y.,ILTSHIRE,J.2000.Discoveryofaggregateusageprolesforwebpersonalization.InProceedingsoftheWorkshoponWebMiningforE-ChallengesandOpportunities,O.,F,H.,J,A.,,R.1999.Miningwebaccesslogsusingrelationalcompetitivefuzzyclustering.InProceedingsofthe8thInternationalFuzzySystemsAssociationWorldCongressIFSA99,M.1999.Aframeworkforcollaborative,content-basedanddemographicltering.Artif.Intel.Rev.136,393,M.,D.1997.Learningandrevisinguserproles:Theidenticationofin-terestingwebsites.Mach.Learn.27,3,313,P.,P,J.,,R.1996.Silkfromasowsear:Extractingusablestructuresfromtheweb.InProceedingsoftheACMCHI96ConferenceonHumanFactorsinComputingSystems,P.,IACOVOU,N.,S,M.,B,P.,,J.1994.GroupLens:Anopenar-chitectureforcollaborativelteringofNetNews.InProceedingsoftheACMCSCW94ConferenceonComputer-SupportedCooperativeWork.ACM,NewYork,175ALTON,G.,C.1988.Ontheuseofspreadingactivationmethodsinautomaticin-formation.InProceedingsofthe11thACMSIGIRInternationalConferenceonResearchandDevelopmentinInformationRetrieval.ACM,NewYork,147ARWAR,B.,K,G.,KONSTAN,J.,,J.2000a.Analysisofrecommendationalgorithmsfore-commerce.InProceedingsoftheACMConferenceonElectronicCommerce.ACM,NewYork,ARWAR,B.,K,G.,KONSTAN,J.,,J.2000b.Applicationofdimensionalityreductioninrecommendersystems:Acasestudy.InProceedingsoftheWebKDDWorkshopattheACM.ACM,NewYork.ARWAR,B.,KONSTAN,J.,B,A.,H,J.,M,B.,,J.1998.UsingagentstoimprovepredictionqualityintheGroupLensresearchcollaborativelteringsystem.ProceedingsoftheACMConferenceonComputerSupportedCooperativeWork(CSCW).ACM,NewYork,345ARWAR,B.M.,K,G.,KONSTAN,J.A.,,J.T.2001.Item-basedcollaborativerecommendationalgorithms.InProceedingsofthe10thInternationalWorldWideWebConference,J.,KONSTAN,J.,,J.2001.E-commercerecommendationapplications.Min.Knowl.Disc.52,115,A.I.,P,A.,U,L.H.,,D.M.2002.Methodsandmetricsforcold-startrecommendations.InProceedingsofthe25thAnnualInternationalACMSIGIRConferenceonResearchandDevelopmentinInformationRetrieval(SIGIR2002).(Tampere,Finland),253ACMTransactionsonInformationSystems,Vol.22,No.1,January2004. Z.Huangetal.,U.,P.1995.Socialinformationltering:Algorithmsforautomatingofmouth.InProceedingsoftheACMCHI95ConferenceonHumanFactorsinComputingSys-.ACM,NewYork,210,I.,C.2000.Collaborativelteringandthegeneralizedvectorspacemodel.InProceedingsofthe23rdAnnualInternationalConferenceonResearchandDevelop-mentinInformationRetrieval(Athens,Greece).351,L.,H,W.,AMENTO,B.,M,D.,,J.1997.PHOAKS:Asystemforsharingrecommendations.Commun.ACM40,3,59,S.K.M.,Z,W.,,P.C.N.1985.Generalizedvectorspacesmodelinin-formationretrieval.InProceedingsofthe8thAnnualInternationalACMSIGIRConferenceonResearchandDevelopmentinInformationRetrieval.ACM,NewYork,18ReceivedJanuary2003;revisedJune2003andSeptember2003;acceptedSeptember2003ACMTransactionsonInformationSystems,Vol.22,No.1,January2004.