User1 realtime interest prediction based Collaborative Filtering and Interactive Computing.docx

User1 realtime interest prediction based Collaborative Filtering and Interactive Computing.docx

《User1 realtime interest prediction based Collaborative Filtering and Interactive Computing.docx》由会员分享，可在线阅读，更多相关《User1 realtime interest prediction based Collaborative Filtering and Interactive Computing.docx（10页珍藏版）》请在冰豆网上搜索。

User1realtimeinterestpredictionbasedCollaborativeFilteringandInteractiveComputing

Userreal-timeinterestpredictionbasedCollaborativeFilteringandInteractiveComputinginAcademicRecommendation

JieYu,HaihongZhao,andFangfangLiu

SchoolofComputerEngineeringandScience,ShanghaiUniversity,ShangHai200072,China

{jieyu,zhaohaihong,ffliu}@

Abstract.WiththerapiddevelopmentofInternet,theinformationandresourceofWebacademicdatabaseisgreatandexplosivegrowth,soitisdifficulttoquicklyandaccuratelyobtaininformationwhichmeetsindividualuser’sneeds.Webpersonalizationserviceseffectivelyalleviateuser'scognitiveburdenandinformationoverloadproblem.First,thisarticleproposesconceptsofuserknowledgeunitanduserknowledgeflowthatrepresentsusershort-terminterestandlong-terminterestrespectively.Second,existingmethodshavesomedefectswhichcan’tbesensitivetoperceiveuserinterestchangeandaccuratelypredictuserreal-timeinterest,weputforwardCollaborativeTimeWeight（CTW）andCollaborativeRelationWeight（CRW）tosolvethoseproblems.Meanwhileuserreal-timeinterestpredictionalgorithmisproposedbasedoncollaborativefilteringandinteractivecomputing.Finally,experimentalresultsdemonstratethatourmethodcanaccuratelycaptureuserreal-timeinterestsandalleviatetheuser’scognitiveburdeneffectively.

Keywords:

Webpersonalizationservices,Userknowledgeflow,Interactivecomputing,Collaborativefiltering,Userinterestprediction.

1Introduction

WiththerapiddevelopmentandpopularityofWebAcademicdatabase,itisattractingmoreandmoreattentionandbecomingthemainwayofobtainingandpropagationinformationandtechnology.Peoplecangetallthenecessaryresourcesfromtheacademicdatabase[1].ButduetoinformationandresourceofWebacademicdatabaseisgreatandexplosivegrowth,itisdifficulttoobtaintheinformationquicklyandaccuratelywhichmeetsindividualuser’sneeds.Webpersonalizedserviceeffectivesolutiontotheproblemofinformationoverloadandusercognitiveburden.Sohowtopredictandrepresentuserreal-timeinterestisakeyissue.

Inrecentyears,lotsofresearchworkabouthowtopredictandrepresentuser’sreal-timeinterestisbeingundertakenorhascompleted.First,inacquisitionandrepresentationofuserinterest,[2]proposeamethodofuserinformationextraction,representationandintegrationfrommultipledatasourcebasedrandomfield.[3]introduceshowtobuildaprobabilitymodel,andthentorepresentpreferencesoftheuser'sinterestaccordingtothismodel.[4]raisestopicofuserinformationofuserpreferencestorepresentuserinterestsolvingtheproblemofheavy-tailednetworkresourceswithpositiveandnegativeinteresttheory.Second，accordingtoresearchalotofliteratureinuserinterestprediction,[5]extenttheideaofbipartitenetworktoformathrivenetworkthentousethisnetworktopredictandexpressedinterest.[6-8]researchthedriftinguser’sinterestsandlawofchanginguser’sinterestsanduserinteresttrackingindividually.[9]usesBayesiannetworkstopredictusers'interestsandpreferences.[10]usesneuralnetworkstopredictusers'interestsandpreferences.[11]istopredicttheuserinterestsanddiscoveruserinterestvariationpointbymixedmodelwhichisprobabilitytheoryandquantumtheory.[12]utilizesbipartitenetworktodiscoverthecollaborativeusersthentopredictuserreal-timeinterestsandpreferencesaccordingtocollaborativeusers.

Mostofmethodsexistingcurrentlycan’teffectivelyandexactlypredicttheuserreal-timeinterestintheacademicdatabase.Forexample,predictionmethodsbasedonBayesian,Neuralnetworkaremostlyofflinethatcan’texactlypredicttheuser'sreal-timeinterest.Onthebipartitenetworks,quantumtheory,althoughtheycanachievereal-timeefficiency,butalgorithmcomplexityandimplementationdifficultiescan’tneglect.Atthesametimeinthefieldofacademicrecommendation,studyofuserinterestpredictionisrelativelyrare,especiallyinuserreal-timeinterest.Basedonthisbackground,thispaperproposesanewmethodtodescribeandrepresentuserinterestandpredictuserreal-timeinterest.First,semanticcontentofsomesessionsinteractingwithWebScienceDatabaseistobuilduserknowledgeunittoaccuratedescriptusershort-terminterest.Computingthesemanticinformationsimilaritybetweenuserknowledgeunitandlogicalreasoningtogenerateuserknowledgeflowwhichistorepresentsuserlong-terminterest.Secondly,basedonsimilarityofuserknowledgeflowsandclusteringalgorithmtodiscovercollaborativeusersofcurrentactivityuser.Finally,accordingtocollaborativeusers,consideringCTWandCRWbetweencollaborativeusers,weproposeaalgorithmcallinguserreal-timeinterestpredictingbasedoncollaborativefilteringandinteractivecomputing.ThegreatestcontributionofthispapertakesintoaccountnotonlyCTWwhichmeanstimesemanticsofcollaborativeusers,butalsoCRWwhichstandardforrelationshipsemanticofcollaborativeusers.

Therestofthispaperisorganizedasfollows.Somerelatedtermsareintroducedinsection2.Section3describesourapproachhowtopredictuserreal-timeinterest.Experimentalresultsaregiveninsection4.Finally,weconcludeourworkinSection5.

2RelateTerms

Inthispaper,weproposeuserreal-timeinterestingpredictionbasedcollaborativefilteringandinteractivecomputinginacademicrecommendation.Beforethepredictionalgorithmisdescribed,wefirstintroductionsomerelativeterms.Thedefinitionsaregivenasfollows:

Definition1.（UserKnowledgeUnitforPersonalizedService,UKUPS）

UKUPSisatwo-tupleN=（C,R）where

--C={c1,c2,…,cn}isafinitenon-emptysetofconcepts.Eachconceptsisatwo-tupleci=（s,cw）where

--sdefinesthesemanticscontentoftheconcept;

--cwdefinestheweightoftheconcept.

--R={r1,r2,…,rm}isafinitenon-emptysetofassociationrelationbetweenconcepts,Eachconceptsisatwo-tupleri={,rw}where

--

Candisapairwhichdefinesthesemanticrelationbetweentwoconcepts;

--rwdefinestherelationdegreebetweencpandcq.

Definition2.（UserKnowledgeFlowforPersonalizedService,UKFPS）

UKFPSisasequentiallinkwithrichsemantics,whichisactivatedbyuser’sdemandsandchangeswiththedemands.

ThesequenceandsemanticofUKFPSisdefinitebyF=（U,S）where

--U={u1,u2,…,un]isasequenceofUKUPS.ThenodesinUKFPSrepresenttheWebresourceswhichmeetuser’sspecificrequirementsinsomesessionsduringafixedtimesection;

--S=definesthesemanticrelationshipsbetweenadjacentUKUPSsi=（,sw）where

--definesisapairwhichdefinesthesemanticrelationbetweentwoUKUPS

--swdefinestherelationshipdegreebetweenun-1andun

Definition3.（CollaborativeTimeWeight,CTW）

Infact,theinterestsofusersarechangingallthetime,andtherecentUKUPScanreflectuser’spreferencebetterthantheearlyones.Forexample,aresearchermaybeinterestedinthetheoryofknowledgefortheuserinterestingpredictionandcollaborativefiltering,buthemayprobablyswitchhisattentiontoresearchersofcurrentfield.GiventheunreliabilityofearlyUKUPS,theimpactcausedbyoutdatedUKUPSshouldbereducedtoitsimpactfactor.

Inthispaper,wewillconsiderthesequenceofUKUPSsthatrecentUKUPSshouldbeincreasetotheimpactfactor,incontrasttheoutdatedUKUPSreduced.WedefinitethosefactorsasCollaborativeTimeWeight（CTW）.

Definition4.（CollaborativeRelationWeight,CRW）

TheoreticalstudyandquantitativeanalysisbasedonFacebooksitesfoundthatsocialnetworkinguserstosharemoreinformationfromtheweakties.Inthispaper,everycollaborativeuser’sinfluencefactorsofthecurrentlyactiveuserisdifferent,wethinkthatUKFPSoflowsimilarityuserwiththecurrentactiveuserismaybecurrentuserinterestedinformation,wewillincreasetheinfluencefactors.Viceversa,wereduce.ThisfactorisdefiniteasCollaborativeRelationWeight（CRW）

Definition5.（InteractiveComputingModel,ICM）

InteractiveComputingModelisatetradM=（S,I,O,F）where

-Sisanenumerablesetofstatesdescribingusers’browsingstates,eachofwhichreflectsuser’srecentinterestthatisdescribedbyUKUPS.

-Iisanenumerablesetofinputstates,whichdescribesuser’sbrowsingbehaviors.

-Oisanenumerablesetofoutputstates,whichdescribesthegeneratedUKUPS.

-F:

S×I->S×OisacomputablefunctionwithsemanticsofUKUPS.

3UserReal-TimeInterestingPrediction

Infact,userinterestischangingallthetime,andtherecentinterestcanreflectuser’spreferencebetterthantheearlyones.Forexample,aresearchermaybeinterestedinthetheoryoftheuserinterestpredictionandcollaborativefiltering,butthenhemayprobablyswitchhisattentiontoresearchersofcurrentfield.Manyeffortshavebeendoneonuserinterestpredictionandsomeachievementshavebeengained.However,existingmethodshavesomedefects.Theyareinsensitivetouserinterestchangeandcan’tbeaccuratepredictionuserinterest.Inaddition,theyareindependentofuser’sinteraction,whichmakeWebpersonalizedservicesunabletoprovideuserreal-timerequirementsinformationeffectively.Aimingatsolvingthesedefects,thispaperpresentsUKUPSandUKFPSwhichisgeneratebyICMtorepresentuserinterest,thenwegetthecollaborativeusersofcurrentactivityuseraccordingtosimilarityofUKFPSs.Finallywepredictthecurrentuserreal-timeinterestbasedoncollaborativefilteringandinteractivecomputingwhichisconsideringtheCT