資源描述:
《2005-A framework for learning predictive structures from multiple tasks and unlabeled data》由會(huì)員上傳分享,免費(fèi)在線閱讀,更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫(kù)。
1、JournalofMachineLearningResearch6(2005)1817–1853Submitted5/05;Revised8/05;Published11/05AFrameworkforLearningPredictiveStructuresfromMultipleTasksandUnlabeledDataRieKubotaAndorie1@us.ibm.comIBMT.J.WatsonResearchCenterYorktownHeights,NY10598,U.S.A.TongZhangtzhang@yahoo-inc.c
2、omYahooResearchNewYork,NY,U.S.A.Editor:PeterBartlettAbstractOneofthemostimportantissuesinmachinelearningiswhetheronecanimprovetheperformanceofasupervisedlearningalgorithmbyincludingunlabeleddata.Methodsthatusebothlabeledandunlabeleddataaregenerallyreferredtoassemi-supervise
3、dlearning.Althoughanumberofsuchmethodsareproposed,atthecurrentstage,westilldon’thaveacompleteunderstandingoftheire?ectiveness.Thispaperinvestigatesacloselyrelatedproblem,whichleadstoanovelapproachtosemi-supervisedlearning.Speci?callyweconsiderlearningpredictivestructuresonh
4、ypothesisspaces(thatis,whatkindofclassi?ershavegoodpredictivepower)frommultiplelearningtasks.Wepresentageneralframeworkinwhichthestructurallearningproblemcanbeformulatedandanalyzedtheoretically,andrelateittolearningwithunlabeleddata.Underthisframework,algorithmsforstructura
5、llearningwillbeproposed,andcomputationalissueswillbeinvestigated.Experimentswillbegiventodemonstratethee?ectivenessoftheproposedalgorithmsinthesemi-supervisedlearningsetting.1.IntroductionInmachinelearningapplications,onecanoften?ndalargeamountofunlabeleddatawithoutdi?culty
6、,whilelabeleddataarecostlytoobtain.Thereforeanaturalquestioniswhetherwecanuseunlabeleddatatobuildamoreaccurateclassi?er,giventhesameamountoflabeleddata.Thisproblemisoftenreferredtoassemi-supervisedlearning.Ingeneral,semi-supervisedlearningalgorithmsusebothlabeledandunlabele
7、ddatatotrainaclassi?er.Althoughanumberofmethodshavebeenproposed,theire?ectivenessisnotalwaysclear.Forexample,Vapnikintroducedthenotionoftransductiveinference(Vapnik,1998),whichmayberegardedasanapproachtosemi-supervisedlearning.Al-thoughsomesuccesshasbeenreported(e.g.,seeJoa
8、chims,1999),therehasalsobeencriticismpointingoutthatthismethodmaynotbehavewellunde