发布时间 : 星期一 文章2013 Context-aware item-to-item recommendation within the factorization framework更新完毕开始阅读7f3969cb804d2b160b4ec0a3
Table2:Mainpropertiesofthedatasets
DatasetDomainGroceryTV1TV2LastFM
E-groceryIPTVIPTVMusic
Trainingset
#Users#Items#Events2494770771449684
992
168837733398174091
Multi
Testset
#EventsLength
56449122962186617941
1111monthweekdayday
62382693.02795449471.000025282151.00001890859721.2715
Table3:Resultswithcontextmanuallyset
RECALL@20
Cosinesimilarity
DatasetGroceryTV1TV2LastFM
Non-CA0.07400.05870.21670.1035
CALevel1iTALSiTALSx0.07380.05910.19150.1098
0.07200.06210.21800.0936
iTALS0.07290.05140.01830.1051
CALevel2
iTALSxiTALSxN0.07330.06170.21510.1018
0.08360.08600.20970.1002
Non-CA0.07110.05480.19540.0920
Scalarproduct
CALevel1iTALSiTALSx0.13230.05990.15150.1073
0.08590.06290.17640.0768
CALevel2iTALSiTALSx0.12440.05220.03640.1099
0.08630.08430.17290.0827
MAP@20
Cosinesimilarity
DatasetGroceryTV1TV2LastFM
Non-CA0.12370.01470.06570.4883
CALevel1iTALSiTALSx0.12600.01600.05710.5096
0.11230.01490.06380.4226
iTALS0.20070.01420.00280.6325
CALevel2
iTALSxiTALSxN0.12440.01670.06230.5020
0.18110.02940.07400.5416
Non-CA0.13260.01420.05760.4571
Scalarproduct
CALevel1iTALSiTALSx0.40380.01720.04420.5410
0.16770.01510.04890.3670
CALevel2iTALSiTALSx0.40410.01440.00770.6424
0.16890.02420.04940.3459
COVERAGE
Cosinesimilarity
DatasetGroceryTV1TV2LastFM
Non-CA88.83?.84?.41.78%
CALevel1iTALSiTALSx94.06?.24?.64.57%
94.87?.13?.14 .98%
iTALS93.40?.79X.62%9.02%
CALevel2
iTALSxiTALSxN94.01?.01?.82.53%
89.61?.11?.70.87%
Non-CA61.75y.82@.32.94%
Scalarproduct
CALevel1iTALSiTALSx63.81W.70.48.36%
78.99y.30(.08.28%
CALevel2iTALSiTALSx85.44y.43 .66%3.08%
46.32u.29&.72%9.88%
resultistheL2iTALSx-basedsimilaritywithscalarproduct.Also,thecoveragevaluesfortheL2methodsandcosinesimilarityoutperformthatofthebasicmethod.
TV2dataset:Inearlierworkswefoundthatseasonality(withthesesettingsatleast)doesnotsuitthisdatasetasacontextdimension.Thusitissurprisingthatthecontext-awareapproachesareonparwiththebasicapproach.TheiTALS-basedmodelsareslightlyworse,whileiTALSx-modelsachieveapproximatelythesameresultswhencosinesimilar-ityisused.ITALSL2similarityhaspoorperformancewithscalarproduct,becausethisapproachisverysensitivetothequalityofcontext.
LastFMdataset:TheperformanceofiTALS-basedmeth-odsarealittlebitbetter,whileiTALSx-basedapproachesareslightlyoutperformedbythebasicmethod.However,thereisanotableincrementinthecoverageforcontext-awaremethods,exceptforL2iTALS-based.
performsusuallybetterthanthemanuallyconstructedones.Theimprovedqualityofthecontextdimensionenablesthecontext-awaremethodstoperformbetter.ApparentlytheiTALS-basedapproachcanbene?tmorefromthisthantheothermodel,becauseitismoresensitive.Theusageofautomaticcontextselectionimprovedtheresultsbyanav-erageof1.62%and4.01%foriTALSandiTALSxmodelsonLevel1andby216.35%and11.06%respectivelyonLevel2.Wehighlighttwoextremecasesfromtheresults.The?rstiscontext-aware,iTALS-basedL2similarityonTV1.Hererecallis7–9timestherecallofthebasicapproach.Theac-tualratiodependsonthesimilaritymetric.TheotheroneisthesamemethodonTV2.Inthiscasetherecallisasmallfractionofthatofthebasicmethod.TheseexamplesdemonstrateshowthesimilaritymetricL2withITALSissensitivetothepropercontextvalues.
4.2Sensitivitytocontextquality
4.3Whichapproachtouse?
Thesecondexperimentexaminesthesensitivityoftheap-proachestocontextquality(seeTable4).Herethecontext-stateswerecreatedautomaticallyusingasimpleclusteringwithinthecontextdimensiontosuitthedatabetter[13].Asshownin[13],theautomaticallydeterminedcontext-states
Theexperimentsshowthateachmethodhasitsstrengthsandweaknesses,thusaclearwinnercannotbedeclared.Ifthequalityofthecontextdimension(w.r.t.theproblem)isacceptablethanL1methodswitheithermodelcanbeusedtoincreasecoveragewithoutlossinaccuracy.TherelativeperformanceofiTALSandiTALSxmodelsdependsonthe
23
Table4:Resultswithcontextautomaticallyset
RECALL@20
Cosinesimilarity
DatasetGroceryTV1TV2LastFM
Non-CA0.07400.05870.06570.1035
CALevel1iTALSiTALSx0.07690.05860.19920.1090
0.07660.06360.21750.0953
iTALS0.08990.52700.03830.1239
CALevel2
iTALSxiTALSxN0.09580.06300.21620.1011
0.10210.07960.20840.1044
Non-CA0.07110.05480.19540.0920
Scalarproduct
CALevel1iTALSiTALSx0.13410.06010.15340.1108
0.09560.06600.17470.0817
CALevel2iTALSiTALSx0.13730.39580.03090.1116
0.10130.11080.17350.0881
MAP@20
Cosinesimilarity
DatasetGroceryTV1TV2LastFM
Non-CA0.12370.01470.06570.4883
CALevel1iTALSiTALSx0.13860.01580.05920.5036
0.12940.01560.06130.4275
iTALS0.20400.10710.01180.8078
CALevel2
iTALSxiTALSxN0.27370.01650.06310.4792
0.32990.02930.07710.4849
Non-CA0.13260.01420.05760.4571
Scalarproduct
CALevel1iTALSiTALSx0.39900.01660.04450.5572
0.23500.01600.04730.3718
CALevel2iTALSiTALSx0.43270.09290.00730.6870
0.27900.04660.04940.3847
usedsimilaritymetricandthepropertiesofthedataset.Forexample:datawithstrongerpopularity-e?ectcantoleratemorenoiseinthecontextdimensionwhenscalarproductsimilarityisusedinsteadofcosinesimilarity.Generally,theiTALS-basedmodelshouldbeappliedinsuchasettingthattoleratesbetterthenoiseinthecontext.
Ifthecontextqualityisfair,thenL2similaritiescangreatlyimproveperformance.TheiTALSmodelisoftenbetterinthisscenario,andmayincreaserecallbeyondexpectations(seeTV1resultsinTable4).HoweverthesensitivityofiTALStocontextselectionisadouble-edgedsword,andthuscanyieldinlargeperformancedecrease(seeTV2re-sultsinTable4).Ifthecontextqualityispoor,thennoimprovementcanbeexpectedfromcontext-awaremethodsingeneral,includingcontext-awaresimilaritiesaswell.Insuchacase,oneshouldsticktobeoriginalcontext-unawareapproach.nario,context-awareitem-2-itemrecommendationalgorithmsareadvisedtobeusedsinceatleastasmallimprovementcanalwaysbeachievedwithlevel-1approaches.Fromlevel-2approaches,themorerobustiTALSx-basedmethodcanbeusedsafelyalmosteverytime,whileontheotherhand,themoresensitivelevel-2iTALS-basedmethodsshouldbeusedwithcare,butcangreatlyimprovetheperformanceifusedwithpropercontextdimension.
Futureworkincludesthethroughexaminationofcontextdimensionsandcontextstatesthatwillenableustopredictwhichcontextdimensionwillproveusefulforagivenprob-lemandwithagivenmodel.Weassumethatsomecontextdimensionsaremoresuitableforcreatingcontext-awareitembiases(orinotherwords:toopromoteanddemoteitems),whileothersaremoreusefulinareweightingsetting,andweassumethatthispropertyofthecontextcanbeprede-terminedfromthedata.
5.CONCLUSION
6.REFERENCES
Thispaperproposedcontext-awaresimilarityfunctionsbasedonthefeaturevectorsofcontext-awarefactorizationmethods.Twolevelsofcontext-awarenesswerede?ned.Onthe?rstlevel,contextisonlyusedtoenablethealgorithmtobeabletolearntheitemfeaturesbetter,butthesimilaritiesthemselvesarenotcontext-aware.Themoreadvancedlevel-2approachesde?necontext-dependentsimilarityfunctions.Wemeasuredrecall,MAPandcoverageintheexperi-mentsperformedonfourimplicitfeedbackdatasets,butnotclearwinnercouldbespeci?ed.Theproposedapproachescanimproveeitheraccuracy(recallandMAP)orcoverage,orincertaincasesboth.Thecontextindependentlevel-1methods,wherecontextisonlyusedduringthelearning,performconsistently,butusuallyonlyslightimprovementcanbeachievedoverthebaseline.Ontheotherhand,con-textdependentsimilarityapproachesareverysensitivetothecontextquality.Thiscanresultbothinoutstandingandinverypoorperformance.Thesensitivityisdi?erentforthetwoinvestigatedmodels:theHadamardproductbasediTALSismuchmoresensitivethanthepairwiseproductbasediTALSxmodel.
Theresultssuggestthatinapracticalapplicationsce-
[1]Adomavicius,G.,Sankaranarayanan,R.,Sen,S.,and
Tuzhilin,A.Incorporatingcontextualinformationinrecommendersystemsusingamultidimensionalapproach.ACMTrans.Inf.Syst.23,1(2005),103–145.
[2]Adomavicius,G.,andTuzhilin,A.Context-aware
recommendersystems.InRecsys’08:ACMConf.onRecommenderSystems(Lausanne,Switzerland,2008),335–336.
[3]Agrawal,R.,Imieli′nski,T.,andSwami,A.Mining
associationrulesbetweensetsofitemsinlarge
databases.InSIGMOD’93:ACMSIGMODInt.Conf.onManagementofData(WashingtonDC,USA,1993),207–216.
[4]Bader,R.,Neufeld,E.,Woerndl,W.,andPrinz,V.
Context-awarePOIrecommendationsinan
automotivescenariousingmulti-criteriadecisionmakingmethods.InCaRR’11:Workshopon
Context-awarenessinRetrievalandRecommendation(PaloAlto,CA,USA,2011),23–30.
[5]Baltrunas,L.,andAmatriain,X.Towards
time-dependantrecommendationbasedonimplicit
24
[6]
[7][8]
[9]
[10]
[11]
[12]
[13][14]
[15]
[16]
[17]
[18]
[19]
[20]
feedback.InCARS’09:WorkshoponContext-awareRecommenderSystems(NewYork,NY,USA,2009),1–5.
Bogers,T.Movierecommendationusingrandomwalksoverthecontextualgraph.InCARS’10:2ndWorkshoponContext-AwareRecommenderSystems(Barcelona,Spain,2010),1–5.
Celma,O.MusicRecommendationandDiscoveryintheLongTail.Springer,2010.
Cremonesi,P.,andTurrin.,R.Analysisofcold-startrecommendationsiniptvsystems.InProceedingsofthe2009ACMConferenceonRecommenderSystems(2009).
Davidson,J.,Liebald,B.,Liu,J.,etal.TheYouTubevideorecommendationsystem.InRecsys’10:ACMConf.onRecommenderSystems(Barcelona,Spain,2010),293–296.
Ge,M.,Delgado-Battenfeld,C.,andJannach,D.
Beyondaccuracy:evaluatingrecommendersystemsbycoverageandserendipity.InRecsys’10:4thACMConf.onRecommenderSystems,ACM(NewYork,NY,USA,2010),257–260.
He,Q.,Pei,J.,Kifer,D.,Mitra,P.,andGiles,L.Context-awarecitationrecommendation.In
WWW’10:Int.Conf.onWorldWideWeb(Raleigh,NC,USA,2010),421–430.
Herlocker,J.L.,Konstan,J.A.,Terveen,L.G.,andRiedl,J.T.Evaluatingcollaborative?lteringrecommendersystems.ACMTransactionsonInformationSystems22,1(2004),5–53.
Hidasi,B.TechnicalreportoniTALSx.Tech.ReportSeries2012-2,GravityR&DInc.,2012.
Hidasi,B.,andTikk,D.Fastals-basedtensor
factorizationforcontext-awarerecommendationfromimplicitfeedback.InProceedingsoftheEuropeanConferenceonMachineLearningandPrinciplesandPracticeofKnowledgeDiscoveryinDatabases(ECMLPKDD’12)(Bristol,UK,September24–28,2012).Hu,Y.,Koren,Y.,andVolinsky,C.Collaborative?lteringforimplicitfeedbackdatasets.InICDM-08:IEEEInt.Conf.onDataMining(Pisa,Italy,2008),263–272.
Karatzoglou,A.,Amatriain,X.,Baltrunas,L.,andOliver,N.Multiverserecommendation:N-dimensionaltensorfactorizationforcontext-awarecollaborative?ltering.InRecsys’10:ACMConf.onRecommenderSystems(Barcelona,Spain,2010),79–86.
Linden,G.,Smith,B.,andYork,J.Amazon.comrecommendations:Item-to-itemcollaborative?ltering.InternetComputing,IEEE7,1(2003),76–80.Panniello,U.,Tuzhilin,A.,Gorgoglione,M.,Palmisano,C.,andPedone,A.Experimental
comparisonofpre-vs.post-?lteringapproachesincontext-awarerecommendersystems.InRecsys’09:ACMConf.onRecommenderSystems(NewYork,NY,USA,2009),265–268.
Rendle,S.,andSchmidt-Thieme,L.Pairwise
interactiontensorfactorizationforpersonalizedtagrecommendation.InProceedingsofthethirdACMinternationalconferenceonWebsearchanddatamining,WSDM’10(2010),81–90.
Shani,G.,andGunawardana,A.Evaluating
recommendationsystems.InRecommenderSystemHandbook,F.Ricci,L.Rokach,B.Shapira,andP.B.Kantor,Eds.Springer,Berlin–Heidelberg,2011,257–297.
[21]Wermser,H.,Rettinger,A.,andTresp,V.Modeling
andlearningcontext-awarerecommendationscenariosusingtensordecomposition.InAdvancesinSocialNetworksAnalysisandMining(ASONAM),2011InternationalConferenceon,IEEE(2011),137–144.[22]Zhou,T.,Kuscsik,Z.,Liu,J.-G.,Medo,M.,Wakeling,
J.R.,andZhang,Y.-C.Solvingtheapparent
diversity–accuracydilemmaofrecommendersystems.ProceedingsoftheNationalAcademyofSciences107,10(2010),4511–4515.
25