[NIPS 2012] Multilabel Classification using Bayesian Compressed Sensing

《[NIPS 2012] Multilabel Classification using Bayesian Compressed Sensing》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、MultilabelClassificationusingBayesianCompressedSensingAshishKapoory,PrateekJainzandRaajayViswanathanzyMicrosoftResearch,Redmond,USAzMicrosoftResearch,Bangalore,INDIAfakapoor,prajain,t-rviswag@microsoft.comAbstractInthispaper,wepresentaBayesianframeworkformultilabelclassifica

2、tionusingcompressedsensing.Thekeyideaincompressedsensingformultilabelclassi-ficationistofirstprojectthelabelvectortoalowerdimensionalspaceusingarandomtransformationandthenlearnregressionfunctionsovertheseprojections.Ourapproachconsidersbothofthesecomponentsinasingleprobabili

3、sticmodel,therebyjointlyoptimizingovercompressionaswellaslearningtasks.Wethenderiveanefficientvariationalinferenceschemethatprovidesjointposteriordistri-butionoveralltheunobservedlabels.Thetwokeybenefitsofthemodelarethata)itcannaturallyhandledatasetsthathavemissinglabelsandb

4、)itcanalsomeasureuncertaintyinprediction.Theuncertaintyestimateprovidedbythemodelallowsforactivelearningparadigmswhereanoracleprovidesinformationaboutlabelsthatpromisetobemaximallyinformativeforthepredictiontask.Ourexperimentsshowsignificantboostoverpriormethodsintermsofpre

5、dictionperformanceoverbenchmarkdatasets,bothinthefullylabeledandthemissinglabelscase.Finally,wealsohighlightvarioususefulactivelearningscenariosthatareenabledbytheprobabilisticmodel.1IntroductionLargescalemultilabelclassificationproblemsariseinseveralpracticalapplicationsan

6、dhasrecentlygeneratedalotofinterestwithseveralefficientalgorithmsbeingproposedfordifferentsettings[1,2].Aprimaryreasonforthrustinthisareaisduetoexplosionofweb-basedapplications,suchasPicasa,Facebookandotheronlinesharingsites,thatcanobtainmultipletagsperdatapoint.Forexample,

7、usersonthewebcanannotatevideosandimageswithseveralpossiblelabels.Suchapplicationshaveprovidedanewdimensiontotheproblemastheseapplicationstypicallyhavemillionsoftags.Mostoftheexistingmultilabelmethodslearnadecisionfunctionorweightvectorperlabelandthencombinethedecisionfunct

8、ionsinacertainmannertopredictlabelsforanunseenpoint[3,4,2,5,6].However,suchapproachesquic