Factor in the Neighbors Scalable and Accurate Collaborative Filtering

《Factor in the Neighbors Scalable and Accurate Collaborative Filtering》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、FactorintheNeighbors:ScalableandAccurateCollaborativeFilteringYehudaKoren∗Yahoo!Research,Haifayehuda@yahoo-inc.comABSTRACTandlatentfactormodels.Neighborhoodmethodsarecenteredoncomputingtherelation-Recommendersystemsprovideuserswithpersonalizedsuggestionsshipsbetweenitemsor,

2、alternatively,betweenusers.Anitem-itemforproductsorservices.ThesesystemsoftenrelyonCollaboratingapproachevaluatesthepreferenceofausertoanitembasedonFiltering(CF),wherepasttransactionsareanalyzedinordertoes-ratingsofsimilaritemsbythesameuser.Inasense,thesemeth-tablishconnect

3、ionsbetweenusersandproducts.ThemostcommonodstransformuserstotheitemspacebyviewingthemasbasketsapproachtoCFisbasedonneighborhoodmodels,whichisbasedofrateditems.Thisway,wenolongerneedtocompareuserstoonsimilaritiesbetweenproductsorusers.Inthisworkweintroduceitems,butratherdire

4、ctlyrelateitemstoitems.anewneighborhoodmodelwithanimprovedpredictionaccuracy.Latentfactormodels,suchasSingularValueDecomposition(SVD),Unlikepreviousapproachesthatarebasedonheuristicsimilarities,compriseanalternativeapproachbytransformingbothitemsandwemodelneighborhoodrelati

5、onsbyminimizingaglobalcostfunc-userstothesamelatentfactorspace,thusmakingthemdirectlytion.Furtheraccuracyimprovementsareachievedbyextendingthecomparable.Thelatentspacetriestoexplainratingsbycharacteriz-modeltoexploitbothexplicitandimplicitfeedbackbytheusers.ingbothproductsa

6、ndusersonfactorsautomaticallyinferredfromPastmodelswerelimitedbytheneedtocomputeallpairwisesimi-userfeedback.Forexample,whentheproductsaremovies,fac-laritiesbetweenitemsorusers,whichgrowquadraticallywithinputtorsmightmeasureobviousdimensionssuchascomedyvs.drama,size.Inparti

7、cular,thislimitationvastlycomplicatesadoptinguseramountofaction,ororientationtochildren;lesswelldefineddi-similaritymodels,duetothetypicallargenumberofusers.Ourmensionssuchasdepthofcharacterdevelopmentor“quirkiness”;newmodelsolvestheselimitationsbyfactoringtheneighborhoodorc

8、ompletelyuninterpretabledimensions.Forusers,eachfactormodel,thusmakingbothitem-ite