Themainchallengesofdatastreamsclassificationincludeinfinitelength,concept-drifting,arrivalofnovelclassesandlackoflabeledinstances.Mostexistingtechniquesaddressonlysomeofthemandignoreothers.Soanensembleclassificationmodelbasedondecision-feedback(ECM-BDF)ispresentedinthispapertoaddressallthesechallenges.Firstly,adatastreamispidedintosequentialchunksandaclassificationmodelistrainedfromeachlabeleddatachunk.Toaddresstheinfinitelengthandconcept-driftingproblem,afixednumberofsuchmodelsconstituteanensemblemodelEandsubsequentlabeledchunksareusedtoupdateE.Todealwiththeappearanceofnovelclassesandlimitedlabeledinstancesproblem,themodelincorporatesanovelclassdetectionmechanismtodetectthearrivalofanovelclasswithouttrainingEwithlabeledinstancesofthatclass.Meanwhile,unsupervisedmodelsaretrainedfromunlabeledinstancestoprovideusefulconstraintsforE.AnextendedensemblemodelExcanbeacquiredwiththeconstraintsasfeedbackinformation,andthenunlabeledinstancescanbeclassifiedmoreaccuratelybysatisfyingthemaximumconsensusofEx.ExperimentalresultsdemonstratethattheproposedECM-BDFoutperformstraditionaltechniquesinclassifyingdatastreamswithlimitedlabeleddata.
