【正文】
? GeneralizeSVDtomultipledimensionalarrays? Uncoverplexrelationshipsamongaspects,movies,andusers.36Linear Combination Tensor FactorizationRatingInference:tensorfactorization37l Objectivefunction:? CPdepositionisdefinedbyfactormatricsA,B,andCofsizeIxR,JxR,and(K+1)xRRatingInference:tensorfactorizationl Inmatrixnotationwherehasthesamesizewith,andandRatingInference:tensorfactorizationl Thepartialderivativesaregivenbyl Aftergettingthegradient,wecanuseafirstorderoptimizationmethodsuchas(NCG)orlimitedmemoryBFGStogetA,B,andC.Inferring aspect ratings40? Extractaspecttermsandopinionwords?Semisupervisedmethod:doublepropagation? ClusteraspecttermsintoKlatentaspects?LatentDirichletAllocation? Computeratingsonaspectsbasedonopinions?ofpositivewords/oftotalopinionwordsExtractingaspecttermsopinions41l DoublepropagationapproachbyG.Qiu,B.Liu.l Useanopinionlexiconasseed.l Usesyntacticrelationstoexpandopinionandaspectlexiconssimultaneously.l Iterateuntilthelexiconsdon’tincrease..:The film has a great plot.Movie aspect clustered by LDA42action story war film movie ……john role man edy timeguy oscar film fun watchdie play american humor storyplot music world cast peoplehero film men character characterscene academy people role thingcop picture death films scenecar cast kubrick batman plotpolice grant history show actorkill screen power time endchase hollywood battle performance partcrime performance epic script feel……Empirical study data set43l Collectstarratings(110stars)andtextualreviewsl Removeuserswhoratelessthan20movies.Raw reducedmovies 1606 1525users 83585 946reviews 193266 53353Sparseness % %Experimental Setupl Evaluationmetric:RootMeanSquaredError.l Twobaselines:l MF:usestarratingsonly.l MR:useopinionsfromreviewsasanemotionspace.l SparsenessExperimental result45演講完畢,謝謝觀看!