2024年5月7日发(作者：)

Item-basedCollaborativeFilteringRecommendation

Algorithms

BadrulSarwar,GeorgeKarypis,JosephKonstan,andJohnRiedl

GroupLensResearchGroup/ArmyHPCResearchCenter

DepartmentofComputerScienceandEngineering

UniversityofMinnesota,Minneapolis,MN55455

{sarwar,karypis,konstan,riedl}@

AppearsinWWW10,May1-5,2001,HongKong.

Abstract

Recommendersystemsapplyknowledgediscoverytechniquestotheproblemofmakingpersonalizedrecom-

mendationsforinformation,ystems,especiallythek-nearest

neighborcollaborativefilteringbasedones,mendousgrowthin

theamountofavailableinformationandthenumberofvisitorstoWebsitesinrecentyearsposessomekeychallenges

re:producinghighqualityrecommendations,performingmanyrecommendations

persecondformillionitional

collaborativefilteringsystem

recommendersystemtechnologiesareneededthatcanquicklyproducehighqualityrecommendations,evenforvery

esstheseissueswehaveexploreditem-basedcollaborativefi-

basedtechniquesfirstanalyzetheuser-itemmatrixtoidentifyrelationshipsbetweendifferentitems,andthenuse

theserelationshipstoindirectlycomputerecommendationsforusers.

Inthispapeintodifferent

,similaritiesbetweenitemvec-

tors),sionmodel).

Finally,weexperimentallyeval

experimentssuggestthatitem-basedalgorithmsprovidedramaticallybetterperformancethanuser-basedalgorithms,

whileatthesametimeprovidingbetterqualitythanthebestavailableuser-basedalgorithms.

1Introduction

Theamountofinformatishave

knownthefeelingofbeingoverwhelmedbythenumberofnewbooks,journalarticles,andconferenceproceedings

logyhasdramaticallyreducedthebarrierstopublishinganddistributinginformation.

Nowitistimetocreatethetechnologiesthatcanhelpussiftthroughalltheavailableinformationtofindthatwhichis

mostvaluabletous.

Oneofthemostpromisingsuchtechnologiesiscollaborativefiltering[19,27,14,16].Collaborativefilteringworks

er,Neo,ismatchedagainstthedatabasetodiscover

neighbors,hattheneighborslikearethen

1

recommendedtoNeo,orativefilteringhasbeenverysuccessfulinboth

researchandpractice,andinbothinformationfir,therere-

mainimportantresearchquestionsinovercomingtwofundamentalchallengesforcollaborativefilteringrecommender

systems.

Thefirstchallengeistoimprovethescalabilityofthecollaborativefilgorithmsareable

tosearchtensofthousandsofpotentialneighborsinreal-time,butthedemandsofmodernsystemsaretosearchtens

r,existingalgorithmshaveperformanceproblemswithindividualusersfor

tance,ifasiteisusingbrowsingpatternsasindicationsof

contentpreference,“longuserrows”slow

downthenumberofneighborsthatcanbesearchedpersecond,furtherreducingscalability.

Thesecoeedrecommendations

theycantrusttohelpthemfiill”votewiththeirfeet”byrefusingtouserecommender

systemsthatarenotconsistentlyaccurateforthem.

Insomewaysthesetwochallengesareinconflict,sincethelesstimeanalgorithmspendssearchingforneighbors,

themorescalableitwillbe,sreason,itisimportanttotreatthetwochallenges

simultaneouslysothesolutionsdiscoveredarebothusefulandpractical.

Inthispaper,weaddresstheseissuesofrecommendersystemsbyapplyingadifferentapproach–item-basedal-

tleneckinconventionalcollaborativefilteringalgorithmsisthesearchforneighborsamongalarge

userpopulationofpotentialneighbors[12].Item-basedalgorithmsavoidthisbottleneckbyexploringtherelation-

shipsbetweenitemsfirst,endationsforusersarecomputedby

fietherelationshipsbetweenitemsarerelatively

static,item-basedalgorithmsmaybeabletoprovidethesamequalityastheuser-basedalgorithmswithlessonline

computation.

1.1RelatedWork

Inthissectionwebrieflypresentsomeoftheresearchliteraturerelatedtocollaborativefiltering,recommendersystems,

dataminingandpersonalization.

Tapestry[10]isoneoftheearliestimplementationsofcollaborativefi

systemreliedontheexplicitopinionsofpeoplefromaclose-knitcommunity,suchasanoffi-

ever,recommendersy,several

upLensresearchsystem[19,16]providesa

pseudonymouscollaborativefi[27]andVideoRecommender[14]

areemailandweb-basealissue

ofCommunicationsoftheACM[20]presentsanumberofdifferentrecommendersystems.

Othertechnologieshavealsobeenappliedtorecommendersystems,includingBayesiannetworks,clustering,and

annetworkscreateamodelbasedonatrainingsetwithadecisiontreeateachnodeandedges

ultingmodel

isverysmall,veryfast,andessentiallyasaccurateasnearestneighbormethods[6].Bayesiannetworksmayprove

practicalforenvironmentsinwhichknowledgeofuserpreferenceschangesslowlywithrespecttothetimeneeded

tobuildthemodelbutarenotsuitableforenvironmentsinwhichuserpreferencemodelsmustbeupdatedrapidlyor

frequently.

Clusteringtechniquese

clustersarecreated,predictionsforanindividualcanbemadebyaveragingtheopinionsoftheotherusersinthat

usteringtechnidiction

isthenanaverageacrosstheclusters,ringtechniquesusuallyproduce

less-personalrecommendationsthanothermethods,andinsomecases,theclustershaveworseaccuracythannearest

neighboralgorithms[6].Oncetheclusteringiscomplete,however,performancecanbeverygood,sincethesize

ringtechniquescanalsobeappliedasa”firststep”for

2