unifi approach activ dual supervis label featur exampl optim debt collect constrain reinforc learn uncertainti sampl transduct experiment design activ dual supervis dual supervis refer gener set learn label exampl label featur label featur natur task text classif frequent provid domain knowledg form word associ strongli class paper problem activ dual supervis optim queri featur label oracl simultan collect form supervis object build classifi cost effect manner appli classic uncertainti experiment design base activ learn scheme graph kernel base dual supervis model empir studi confirm potenti scheme significantli reduc cost acquir label data train high qualiti model sentiment analysi blog combin lexic knowledg text classif activ featur dvalu acquisit induct algorithm build predict model input train data form featur vector acquir valu featur costli simpli acquir valu wast prohibit expens activ featur acquisit afa select featur increment attempt improv predict model cost effect paper present framework afa base estim inform straightforward principl estim approxim made appli framework practic present acquisit polici sampl expect util seu emploi estim enabl effect rank potenti acquisit set inform underli domain present experiment result show compar polici repres sampl featur acquisit seu reduc cost produc model desir accuraci exhibit consist perform domain extend framework gener model set featur valu class label miss costli acquir predict analysi improv invoic dto dcash collect custom target model activ dselect web content problem predict likelihood compani purchas product seller statist model develop ibm purpos reli histor transact data coupl structur firmograph inform compani revenu number employe paper extend methodolog includ addit text base featur base analysi content compani websit empir result demonstr incorpor web content significantli improv custom target present method activ select web content improv model reduc cost acquisit process breast cancer identif kdd cup winner report activ learn probabl estim jensen dshannon diverg abstract activ select good train exampl import approach reduc data collect cost machin learn exist method focu maxim classif accuraci applic unequ misclassif cost produc good class probabl estim cpe import optim classif accuraci introduc approach activ learn base algorithm bootstrap lv activedecor jensen shannon diverg similar measur probabl distribut improv sampl select optim cpe comprehens experiment result demonstr benefit approach combin bia varianc reduct techniqu regress tree abstract gradient boost bag appli regressor reduc error due bia varianc altern stochast gradient boost sgb iter bag ib attempt simultan reduc contribut bia varianc error provid extens empir analysi method altern bia varianc reduct approach bag gradient boost baggb bag stochast gradient boost bagsgb experiment result demonstr sgb perform ib altern approach result show baggb bagsgb perform competit low bia learner gener iter bag effect method expect util approach activ featur dvalu acquisit creat divers ensembl artifici data divers ensembl classifi import factor determin gener error present method gener ensembl decor divers ensembl creation opposit relabel artifici train exampl directli construct divers hypothes addit artifici construct train exampl techniqu simpl gener meta learner strong learner base classifi build divers committe experiment result decis tree induct base learner demonstr approach consist achiev higher predict accuraci base classifi bag random forest decor obtain higher accuraci boost small train set achiev compar perform larger train set divers ensembl activ learn queri committe eectiv approach select sampl disagr ensembl hypothes select data label queri bag queri boost practic implement approach bag boost build committe eectiv activ learn critic committe made consist hypothes dierent decor recent develop method directli construct divers committe arti cial train data paper introduc activ decor decor committe select good train exampl experi ensembl miss noisi data potenti advantag multipl classifi system increas robust nois imperfect data construct divers classifi ensembl artifici train exampl ensembl method bag boost combin decis multipl hypothes strongest exist machin learn method divers member ensembl import factor determin gener error paper present method gener ensembl directli construct divers hypothes addit artifici construct train exampl techniqu simpl gener meta learner strong learner base classifi build divers committe experiment result decis tree induct base learner demonstr approach consist achiev higher predict accuraci base classifi bag obtain higher accuraci boost earli learn curv train data limit content dboost collabor filter improv recommend recommend system collabor filter content base method predict item interest user method advantag individu fail provid good recommend situat incorpor compon method hybrid recommend system overcom shortcom natur languag assist dialog system onlin product recommend emerg commerc system success inform access commerc websit essenti menu driven navig keyword search provid commerci site consider limit tend overwhelm frustrat user lengthi rigid effect interact provid effici solut inform access built natur languag assist nla web base natur languag dialog system user find relev product commerc site system bring technolog natur languag process human comput interact creat faster intuit interact web site combin statist pars techniqu tradit ai rule base technolog creat dialog system accommod custom busi requir system embed applic recommend laptop deploi pilot ibm websit 