Toward the Next Generation of Recommender Systems A Survey o

recommdation engine

734IEEETRANSACTIONSONKNOWLEDGEANDDATAENGINEERING,VOL.17,NO.6,JUNE2005

TowardtheNextGenerationofRecommenderSystems:ASurveyoftheState-of-the-Artand

PossibleExtensions

GediminasAdomavicius,Member,IEEE,andAlexanderTuzhilin,Member,IEEE

Abstract—Thispaperpresentsanoverviewofthefieldofrecommendersystemsanddescribesthecurrentgenerationof

recommendationmethodsthatareusuallyclassifiedintothefollowingthreemaincategories:content-based,collaborative,andhybridrecommendationapproaches.Thispaperalsodescribesvariouslimitationsofcurrentrecommendationmethodsanddiscussespossibleextensionsthatcanimproverecommendationcapabilitiesandmakerecommendersystemsapplicabletoanevenbroaderrangeofapplications.Theseextensionsinclude,amongothers,animprovementofunderstandingofusersanditems,incorporationofthecontextualinformationintotherecommendationprocess,supportformultcriteriaratings,andaprovisionofmoreflexibleandlessintrusivetypesofrecommendations.

IndexTerms—Recommendersystems,collaborativefiltering,ratingestimationmethods,extensionstorecommendersystems.

æ

1

INTRODUCTION

ECOMMENDER

R

systemshavebecomeanimportant

researchareasincetheappearanceofthefirstpapersoncollaborativefilteringinthemid-1990s[45],[86],[97].Therehasbeenmuchworkdonebothintheindustryandacademiaondevelopingnewapproachestorecommendersystemsoverthelastdecade.Theinterestinthisareastillremainshighbecauseitconstitutesaproblem-richresearchareaandbecauseoftheabundanceofpracticalapplicationsthathelpuserstodealwithinformationoverloadandprovidepersonalizedrecommendations,content,andservicestothem.Examplesofsuchapplica-tionsincluderecommendingbooks,CDs,http://www.77cn.com.cn[61],moviesbyMovieLens[67],andnewsatVERSIFITechnologies(http://www.77cn.com.cn)[14].Moreover,someofthevendorshaveincorporatedrecommendationcapabilitiesintotheircommerceservers[78].

However,despitealloftheseadvances,thecurrentgenerationofrecommendersystemsstillrequiresfurtherimprovementstomakerecommendationmethodsmoreeffectiveandapplicabletoanevenbroaderrangeofreal-lifeapplications,includingrecommendingvacations,certaintypesoffinancialservicestoinvestors,andproductstopurchaseinastoremadebya“smart”shoppingcart[106].Theseimprovementsincludebettermethodsforrepresent-inguserbehaviorandtheinformationabouttheitemstoberecommended,moreadvancedrecommendationmodeling

methods,incorporationofvariouscontextualinformationintotherecommendationprocess,utilizationofmultcriteriaratings,developmentoflessintrusiveandmoreflexiblerecommendationmethodsthatalsorelyonthemeasuresthatmoreeffectivelydetermineperformanceofrecommen-dersystems.

Inthispaper,wedescribevariouswaystoextendthecapabilitiesofrecommendersystems.However,beforedoingthis,wefirstpresentacomprehensivesurveyofthestate-of-the-artinrecommendersystemsinSection2.Then,weidentifyvariouslimitationsofthecurrentgenerationofrecommendationmethodsanddiscusssomeinitialap-proachestoextendingtheircapabilitiesinSection3.

2THESURVEY

OF

RECOMMENDERSYSTEMS

.G.AdomaviciusiswiththeCarlsonSchoolofManagement,UniversityofMinnesota,32119thAvenueSouth,Minneapolis,MN55455.E-mail:gedas@umn.edu.

.A.TuzhiliniswiththeSternSchoolofBusiness,NewYorkUniversity,44West4thStreet,NewYork,NY10012.E-mail:atuzhili@stern.nyu.edu.Manuscriptreceived8Mar.2004;revised14Oct.2004;accepted10Dec.2004;publishedonline20Apr.2005.

Forinformationonobtainingreprintsofthisarticle,pleasesende-mailto:tkde@http://www.77cn.com.cn,andreferenceIEEECSLogNumberTKDE-0071-0304.

1041-4347/05/$20.00ß2005IEEE

Althoughtherootsofrecommendersystemscanbetracedbacktotheextensiveworkincognitivescience[87],approximationtheory[81],informationretrieval[89],forecastingtheories[6],andalsohavelinkstomanagementscience[71]andtoconsumerchoicemodelinginmarketing[60],recommendersystemsemergedasanindependentresearchareainthemid-1990swhenresearchersstartedfocusingonrecommendationproblemsthatexplicitlyrelyontheratingsstructure.Initsmostcommonformulation,therecommendationproblemisreducedtotheproblemofestimatingratingsfortheitemsthathavenotbeenseenbyauser.Intuitively,thisestimationisusuallybasedontheratingsgivenbythisusertootheritemsandonsomeotherinformationthatwillbeformallydescribedbelow.Oncewecanestimateratingsfortheyetunrateditems,wecanrecommendtotheusertheitem(s)withthehighestestimatedrating(s).

Moreformally,therecommendationproblemcanbeformulatedasfollows:LetCbethesetofallusersandletSbethesetofallpossibleitemsthatcanberecommended,suchasbooks,movies,orrestaurants.ThespaceSof

PublishedbytheIEEEComputerSociety

recommdation engine

ADOMAVICIUSANDTUZHILIN:TOWARDTHENEXTGENERATIONOFRECOMMENDERSYSTEMS:ASURVEYOFTHE

STATE-OF-THE-ART...735

TABLE1

AFragmentofaRatingMatrixforaMovieRecommenderSystem

possibleitemscanbeverylarge,ranginginhundredsofthousandsorevenmillionsofitemsinsomeapplications,suchasrecommendingbooksorCDs.Similarly,theuserspacecanalsobeverylarge—millionsinsomecases.Letubeautilityfunctionthatmeasurestheusefulnessofitemstouserc,i.e.,u:CÂS!R,whereRisatotallyorderedset(e.g.,nonnegativeintegersorrealnumberswithinacertainrange).Then,foreachuserc2C,wewanttochoosesuchitems02Sthatmaximizestheuser’sutility.Moreformally:

8c2C;

s0c¼argmaxuðc;sÞ:

s2S

ð1Þ

Inrecommendersystems,theutilityofanitemisusuallyrepresentedbyarating,whichindicateshowaparticularuserlikedaparticularitem,e.g.,JohnDoegavethemovie“HarryPotter”theratingof7(outof10).However,asindicatedearlier,ingeneral,utilitycanbeanarbitraryfunction,includingaprofitfunction.Dependingontheapplication,utilityucaneitherbespecifiedbytheuser,asisoftendonefortheuser-definedratings,oriscomputedbytheapplication,ascanbethecaseforaprofit-basedutilityfunction.

EachelementoftheuserspaceCcanbedefinedwithaprofilethatincludesvarioususercharacteristics,suchasage,gender,income,maritalstatus,etc.Inthe …… 此处隐藏：79719字，全部文档内容请下载后查看。喜欢就下载吧 ……