effici numer stabl spars learn cross valid framework choos model dataset transfer learn gener transport system immers video evalu futur cumulonimbu forecast base rough set artifici immun algorithm gener transport system immers video evalu futur commun outlier effici detect inform network gener heurist dfree featur construct improv accuraci multi dlabel classif multi dlabel cost predict model heterogen sourc analysi chernoff criterion linear dimension reduct svdppc effect singular decomposit dbase method conserv diverg dexpress gene modul identif visual data secur manag smart citi heterogen cross domain rank latent space relax transfer class spectral partit abstract exist transfer learn techniqu limit problem knowledg transfer task share set class label paper relax constraint propos spectral base solut aim unveil intrins structur data gener partit target data transfer eigenspac separ sourc data clusteringbas kl diverg propos automat adjust transfer evalu propos model text imag dataset class categori sourc target data explicitli class transfer class show propos approach improv baselin averag accuraci sourc code dataset author univers learn relat distribut adapt graph transduct thunderstorm forecast model base weight svm data field mine interest ratio pattern stream slide window cross domain distribut adapt kernel map label exampl limit difficult obtain transfer learn emploi knowledg sourc domain improv learn accuraci target domain assumpt made exist approach margin condit probabl directli relat sourc target domain limit applic origin space linear transform solv problem propos adapt kernel approach map margin distribut targetdomain sourc domain data common kernel space util sampl select strategi draw condit probabl domain closer formal show kernel map space differ distribut domain bound predict error propos approach bound experiment result demonstr propos method outperform tradit induct classifi state art boost base transfer algorithm domain includ text categor web page rate achiev higher accuraci approach text categor problem sourc code dataset author heterogen sourc consensu learn decis propag negoti nowadai enorm amount data continu gener massiv scale conflict view import consolid concept intellig decis make predict research area peopl result achiev combin consolid predict obtain public network authorship network textual content public multipl supervis unsupervis hypothes drawn inform sourc negoti differ consolid decis yield accur model due divers heterogen model paper address problem consensu learn compet hypothes reli knowledg supervis learn intern structur unsupervis cluster argu consensu learn np hard problem propos solv effici heurist method construct belief graph propag predict supervis model unsupervis negoti reach consensu final decis consolid calcul model weight base degre consist model experi conduct newsgroup data cora research paper dblp author confer network yahoo movi dataset result show propos method improv classif accuraci cluster qualiti measur nmi base model run time proport number instanc effici larg scale data set latent space domain transfer high dimension overlap distribut transfer knowledg domain challeng due number reason condit margin distribut train data test data ident model train domain directli appli domain low accuraci applic larg featur set text document sequenc data medic data imag data resolut domain featur introduc larg number miss valu consid union featur domain word margin distribut overlap time problem high dimension thousand featur combin high bag bag predictor method gener multipl version predictor aggreg predictor aggreg averag version predict numer outcom plural vote predict class multipl version form make bootstrap replic learn set learn set test real simul data set classif regress tree subset select linear regress show bag give substanti gain accuraci vital element instabl predict method perturb learn set signific predictor construct bag improv accuraci introduct learn set consist data ng class label numer respons procedur learn set form predictor input break comput barrier divid dconquer aggreg base approach alu insert site characteris fair mobil payment protocol activ transfer domain knowledg graph dbase iter hybrid featur select direct mine discrimin essenti frequent pattern model dbase search tree knowledg transfer multipl model local structur map effect knowledg transfer classif algorithm depend differ distribut gener train exampl test exampl drawn task difficult train exampl domain test domain paper propos local weight ensembl framework combin multipl model transfer learn weight dynam assign model predict power test integr advantag learn algorithm label inform multipl train domain unifi classif model appli domain importantli previous propos method base learn method requir specif design transfer learn show optim local weight ensembl framework gener approach combin multipl model domain transfer propos implement local weight assign map structur model structur test domain weight model local consist neighborhood structur test experiment result text classif spam filter intrus detect data set demonstr signific improv classif accuraci gain framework transfer learn task newsgroup messag categor propos local weight ensembl framework achiev accuraci singl model predict correctli test exampl summari improv accuraci problem forward semi dsupervis featur select forward semi dsupervis featur select cluster evolv data stream base onlin recent dbias approxim type dindepend correct sampl select bia structur discoveri dbalanc sampl select bia common problem real world applic train data obtain realist constraint make follow distribut futur test data applic hospit clinic studi common practic build model elig volunt train data appli model entir popul volunt select random train set drawn distribut test set dataset suffer sampl select bia covari shift past year work propos reduc sampl select bia static match distribut train set test set paper explor distribut directli propos classifi data stream skew class distribut concept drift forecast skew bias stochast ozon dai analys solut forecast skew bias stochast ozon dai analys solut optim suppli chain servic dorient architectur fair payment protocol dcommerc transact research construct soa enterpris china imag hallucin neighbor embed visual primit manifold assumpt mine data stream analysi practic recent year wit increas number studi stream mine aim build accur model continu arriv data exist work make implicit assumpt train data test data sampl distribut distribut evolv time demonstr true distribut model fit observ distribut unsatisfactori accuraci incom data practic assum bare minimum learn observ data random guess predict class label importantli formal multi dagent optim design multi dresourc job shop schedul problem semi dsupervis cast index featur dlength film gener framework mine concept ddrift data stream skew distribut recent year interest studi predict model data stream studi assum balanc stabl data stream handl skew posit lot neg stochast distribut typic data stream applic paper propos approach mine data stream estim reliabl posterior probabl ensembl model match distribut sampl neg repeat sampl posit formal show interest import properti propos framework reliabl estim probabl skew posit class accuraci estim probabl effici scalabl experi perform synthet real world dataset skew distribut demonstr framework substanti advantag exist approach estim reliabl predic accuraci sampl select bia effici correct model averag unlabel exampl challeng experi prototyp multi dmodal stream analyt monitor applic system fgf web tool fish gene famili genom databas traceabl electron cash model base group signatur local linear model face appear manifold applic dual dsubspac base classif face recognit imag set hierarch extract exemplar appear manifold unsupervis nonparametr approach propos automat extract repres face sampl exemplar video sequenc imag set multipleshot face recognit motiv nonlinear dimension reduct algorithm call isomap local neighborhood inform approxim geodes distanc face imag hierarch agglom cluster hac algorithm appli group similar face base estim geodes distanc approxim locat appear manifold defin exemplar cluster center templat match subsequ test stage final recognit outcom major vote scheme combin decis individu frame test set experiment result subject video databas demonstr effect flexibl propos method discov unrev properti probabl estim tree algorithm select perform explan forecast skew bias stochast ozon dai analys solut optim cooper agent organ distribut schedul system facial express recognit advanc local binari pattern paper propos facial express recognit approach base set featur extract face imag textur featur global appear featur set obtain extend local binari pattern intens gradient map comput tsalli entropi gabor filter respons set featur obtain perform nullspac base linear discrimin analysi train face imag propos method evalu extens experi jaff databas compar wide facial express recognit approach experiment result show propos approach maintain high recognit rate wide rang resolut level outperform altern method index term face recognit machin vision gener framework accur fast regress data summar random decis tree revers test effici framework select classifi sampl select bia import assumpt made classif algorithm train test set drawn distribut call stationari distribut assumpt futur past data set ident probabilist standpoint domain real world applic market solicit fraud detect drug test loan approv popul survei school enrol rare case label sampl train bias wai due varieti practic reason limit circumst tradit method evalu expect gener error classif algorithm structur risk minim ten fold cross valid leav valid return poor estim classif algorithm train bias dataset accur futur unbias dataset number compet candid estim order learn algorithm accuraci poor random guess method determin accur learner need data mine sampl select bia real world applic present approach determin learner perform unbias test set possibl bias train set fraction comput cost cross valid base approach effici model averag dperform boost bag abstract bayesian model averag bay optim classifi boc ensembl techniqu extens statist literatur compar ensembl techniqu bag boost boc rare data mine partli due model averag perceiv ineffici bag boost consist outperform singl model rais question boc datamin show answer question illustr recent effici model averag approach significantli outperform bag boost realist difficult situat extens class label nois sampl select bia class problem knowledg insight model averag outperform bag boost situat publish machin learn mine statist commun video dbase face recognit bayesian infer model effect estim posterior probabl explain accuraci random decis tree approach improv categor classifi sensit sampl select bia recent paper categor classifi learn algorithm sensit common type sampl select bia chanc select train sampl depend featur vector directli class label classifi learner categor local insensit type sampl select bia consid global paper true model distinguish model algorithm output discuss bayesian classifi logist regress hard margin support vector machin true model model gener true class label implicitli assum contain model space learner true class probabl model estim class probabl assum asymptot converg train data set size increas discuss naiv bay decis tree soft margin support vector machin model space assum true model classif algorithm argu global learner argu classifi learn algorithm includ discuss affect sampl select bia depend dataset heurist induct bia impli classifi learn algorithm appropri dataset make earlier experiment result produc addit result illustr claim local manifold match face recognit maxim tree divers build complet drandom decis tree make data mine model model dpai custom exchang carrier optim probabl estim random decis tree mine extrem skew trade anomali combin null space dbase gabor featur face recognit propos face recognit strategi combin discrimin gabor featur multi scale multi orient bank chosen gabor filter appli imag construct group featur vector null space base lda nlda perform simultan orient channel origin imag give compon classifi output combin increas final recognit rate experiment result feret databas demonstr effect flexibl propos method decis tree evolut limit number label data item drift data stream previous propos mine method data stream make unrealist assumpt label data stream readili mine anytim real world problem label data stream rare immedi due reason model reconstruct label data period passiv stream mine model drawback propos concept demand driven activ data mine activ mine loss model continu guess true class label estim small number instanc actual class label verifi pai afford cost estim loss toler threshold model evolv small number instanc verifi true class label previou work activ mine concentr error guess estim paper discuss approach decis tree evolut combin null space dbase gabor featur face recognit propos face recognit strategi combin discrimin gabor featur multi scale multi orient bank chosen gabor filter appli imag construct group featur vector null space base lda nlda perform simultan orient channel origin imag give compon classifi output combin increas final recognit rate experiment result feret databas demonstr effect flexibl propos method systemat data select mine concept ddrift data stream activ mine data stream previous propos mine method data stream make unrealist assumpt label data stream readili mine anytim real world problem label data stream rare immedi due reason model reconstruct label data period passiv stream mine model drawback propos concept demand driven activ data mine activ mine loss model continu guess true class label estim small number instanc actual class label verifi pai afford cost estim loss toler threshold model evolv small number instanc verifi true class label previou work activ mine concentr error guess estim paper discuss approach decis tree evolut fast algorithm subspac cluster pattern similar streammin classifi ensembl dbase engin mine concept ddrift data stream demonstr streammin random decis tree ensembl base engin mine data stream fundament challeng data stream mine applic credit card transact author secur buysel transact phone call record concept drift discrep previous learn model true model data basic problem abil judici select data adapt model accur match chang concept data stream streammin techniqu support mine data stream concept drift demonstr kei function streammin detect concept drift fly train stream model classifi incom data stream know ground truth systemat data select data data chunk comput optim model fit chang data stream artifici anomali detect unknown network intrus intrus detect system idss capabl detect unknown attack anomali studi problem build detect model pure anomali detect combin misus anomali detect detect unknown intrus propos algorithm gener artifici anomali coerc induct learner discov accur boundari class normal connect intrus anomali empir studi show pure anomali detect model train normal artifici anomali capabl detect unknown intrus class accuraci intrus class combin misus anomali detect model accur pure misus detect model detect intrus capabl detect unknown intrus class accuraci measur class cross dfeatur analysi detect ad dhoc rout anomali index weight dsequenc larg databas present index structur manag weight sequenc larg databas weight sequenc defin dimension structur element sequenc weight seri network event instanc weight sequenc event timestamp queri larg sequenc databas event occurr pattern step understand tempor causal relationship event index structur propos paper enabl effici retriev databas subsequ possibl contigu match queri sequenc event weight index method take consider uniform frequenc distribut event sequenc data addit method find broad rang applic index scientif data consist multipl numer column discoveri correl column instanc index dna micro arrai record express level gene condit enabl search gene respons experiment perturb follow pattern demonstr real world data set method effect effici random model accuraci effici induct learn sequenti data scan mine concept ddrift data stream ensembl classifi recent mine data stream concept drift action insight import challeng task wide rang applic includ credit card fraud protect target market network intrus detect convent knowledg discoveri tool face challeng overwhelm volum stream data concept drift paper propos gener framework mine concept drift data stream weight ensembl classifi train ensembl classif model ripper naiv bayesian sequenti chunk data stream classifi ensembl judici weight base expect classif accuraci test data time evolv environ ensembl approach improv effici learn model accuraci perform classif empir studi show propos method substanti advantag singl classifi approach predict accuraci ensembl framework effect varieti classif model vist dynam index method queri xml data tree structur research provid flexibl queri facil extract data structur xml document paper propos vist index structur search xml document repres xml document xml queri structur encod sequenc show queri xml data equival find subsequ match unlik index method disassembl queri multipl queri join result queri provid final answer vist tree structur basic unit queri avoid expens join oper vist unifi index content structur xml document perform advantag method index content structur vist support dynam index updat reli sole tree special data structur support dbmss experi show vist ectiv scalabl cient support structur queri prune dynam schedul cost dsensit ensembl index structur pattern similar search dna microarrai dat fulli distribut framework cost dsensit data mine paper propos fulli distribut system compar central partial distribut system cost sensit data mine experiment result shown approach achiev higher accuraci central partial distribut learn method incur train time commun comput overhead progress model present induct learn perform frustrat batch process user interact system control final accuraci train time accuraci produc model low comput resourc misspent paper propos progress model framework progress model learn algorithm estim onlin accuraci final model remain train time estim accuraci expect user termin train prior complet wast resourc user choos complet learn process progress model comput model expect accuraci expect time describ implement progress model ensembl classifi empir comparison reinforc learn strategi sequenti target market empir evalu perform reinforc learn method applic sequenti target market propos evalu progress reinforc learn method rang direct batch method indirect simul base method call semidirect method fall conduct number control experi evalu perform compet method result indirect method perform situat perfect model realist situat system model paramet restrict attent indirect method perform tend degrad show semi direct method effect reduc amount comput attain level perform result profit polici framework scalabl cost dsensit learn base comb probabl benefit ensembl dbase adapt intrus detect cost dsensit model intrus detect respons intrus detect system idss maxim realiz secur goal minim cost paper studi problem build cost sensit intrus detect model examin major cost factor id includ develop cost oper cost damag cost due success intrus cost manual autom respons intrus cost factor qualifi defin attack taxonomi site specif secur polici prioriti defin cost model formul total expect cost id present cost sensit machin learn techniqu produc detect model optim user defin cost metric empir experi show cost sensit model deploy techniqu effect reduc cost intrus detect artifici anomali detect unknown network intrus intrus detect system idss capabl detect unknown attack anomali studi problem build detect model pure anomali detect combin misus anomali detect detect unknown intrus propos algorithm gener artifici anomali coerc induct learner discov accur boundari class normal connect intrus anomali empir studi show pure anomali detect model train normal artifici anomali capabl detect unknown intrus class accuraci intrus class combin misus anomali detect model accur pure misus detect model detect intrus capabl detect unknown intrus class accuraci measur class mine system audit data opportun challeng intrus detect essenti compon comput secur mechan requir accur effici analysi larg amount system network audit data applic area data mine characteri tic audit data abund raw data rich system network semant amp quot stream amp quot ac cordingli develop data mine approach focu featur extract con struction custom gener algorithm ac cord semant inform optim execut effici output model paper describ data mine framework min ing audit data intrus detect model discuss advantag limit outlin open research problem data mine dbase intrus detector overview columbia id project multipl model cost dsensit approach intrus detect intrus detect system idss maxim secur minim cost paper studi problem build cost sensit intrus detect model realtim detect briefli discuss major cost factor id includ consequenti oper cost propos multipl model cost sensit machin learn techniqu produc model optim user defin cost metric empir experi line analysi show reduct approxim oper cost singl model approach reduct approxim consequenti cost pure accuraci base approach web dcollabor filter recommend music crawl web adacost misclassif cost dsensit boost adacost variant adaboost misclassif cost sensit boost method cost misclassif updat train distribut success boost round purpos reduc cumul misclassif cost adaboost formal show adacost reduc upper bound cumul misclassif cost train set empir evalu shown signific reduct cumul misclassif cost adaboost consum addit comput power applic adaboost distribut scalabl dline learn introduct learn larg distribut databas impos major perform challeng data mine databas grown larg fit main memori learn monolith classifi slow requir data held main memori databas inher distribut brought singl site varieti reason hard learn singl classifi constraint applic learn fraud detect model particip credit card compani dataset huge strictli distribut databas larg amount data period real time learn model data set time increment provid expens ineffici learn page dindepend heurist extract data web page bottleneck implement system intellig queri web develop wrapper program extract data web page describ method learn gener page independ heurist extract data html document input learn system set work wrapper program pair html page correctli wrap output gener procedur extract data work format page experi collect constrain realist extract problem demonstr problem handl perfectli learn extract heurist handl accept demonstr learn page independ extract heurist substanti improv perform method learn page specif wrapper keyword inform integr machin learn extract introduct number recent system oper take informatio 