transfer increment learn pattern classif skif data imput framework concept drift data stream ensembl prune individu contribut order ensembl set learn model make decis collect ensembl accur singl learner exist ensembl method tend construct unnecessarili larg ensembl increas memori consumpt comput cost ensembl prune tackl problem select subset ensembl member form subensembl subject resourc consumpt respons time accuraci similar origin ensembl paper analyz accuraci divers trade prove classifi accur make predict minor group import subensembl construct base gain insight heurist metric consid accuraci divers propos explicitli evalu individu classifi contribut ensembl incorpor ensembl member decreas order contribut subensembl form user select top percent ensembl member depend resourc avail toler wait time predict experiment result uci data set show subensembl form propos epic ensembl prune individu contribut order algorithm outperform origin ensembl state ofth art ensembl prune method orient order oo rule synthes multipl relat databas featur select dataset imbalanc class distribut featur select dataset imbalanc class distribut normal dimension reduct nonneg matrix factor multipl criteria program model vip dmail behavior analysi impact gene select imbalanc microarrai express data bia dvarianc analysi ensembl regular multipl criteria linear program model mine data stream label unlabel train exampl vagu dclass learn data stream approxim repeat pattern mine gap requir error detect uncertainti model imprecis data multipl inform sourc cooper learn applic face problem learn object dataset inform auxiliari sourc benefici integr object dataset learn paper propos omni view learn approach enabl learn multipl data collect theme organ heterogen data sourc unifi tabl global data view achiev omni view learn goal object dataset auxiliari dataset share instanc level depend structur propos relat mean cluster instanc auxiliari dataset cluster build featur captur correl object auxiliari dataset experiment result demonstr omni view learn build model outperform learn object dataset comparison train algorithm assert omni view learn altern effect semi supervis learn aggreg ensembl mine concept drift data stream nois featur select bias sampl distribut aggreg ensembl mine concept drift data stream nois quantit associ rule vey behavior footprint dpropag worm detect profil contrast pattern mine gap constraint peptid fold predict cleans noisi data stream vob predictor vote bag classif vci predictor vote classif imput learn set empir studi supervis learn biolog sequenc profil microarrai express data analysi categor mine concept drift data stream ivesta interact visual evalu system drive test data imput dboost collabor filter machin learn classifi imput dboost collabor filter machin learn classifi select object function multipl criteria program classif data mine conceptu equival contrast mine classif learn learn occur compar classif learn order compar data group exist method compar raw instanc learn classif rule paper take approach conceptu equival group equival underli concept equival instanc space necessarili overlap rule set necessarili present appear methodolog compar propos learn represent group underli concept crossexam group instanc group concept represent innov fold quantifi degre conceptu equival group retrac sourc discrep level abstract level underli concept specif level instanc appli numer data categor data fourth circumv direct comparison possibl alarg number rule demand substanti effort reduc depend accuraci emploi classif algorithm empir evid suggest methodolog effect simpl scenario nois cleans concept chang learn mine nois knowledg error dawar data mine lazi bag approach classif paper propos lazi bag lb build bootstrap replic bag base characterist test instanc receiv test instanc xk lb trim bootstrap bag take consider xk nearest neighbor train data hypothesi unlabel instanc nearest neighbor provid valuabl inform enhanc local learn gener classifi refin decis boundari emphas test instanc surround region take full advantag xk nearest neighbor classifi reduc classif bia varianc classifi xk result lb built classifi significantli reduc classif error compar tradit bag tb approach investig lb perform carefulli design synthet dataset gain insight lb work condit outperform tb test lb rival algorithm larg suit real world benchmark dataset varieti statist test empir result confirm lb statist significantli outperform altern method term reduc classif error editori call amp calendareditori lazi bag approach classif paper propos lazi bag lb build bootstrap replic bag base characterist test instanc receiv test instanc xk lb trim bootstrap bag take consider xk nearest neighbor train data hypothesi unlabel instanc nearest neighbor provid valuabl inform enhanc local learn gener classifi refin decis boundari emphas test instanc surround region take full advantag xk nearest neighbor classifi reduc classif bia varianc classifi xk result lb built classifi significantli reduc classif error compar tradit bag tb approach investig lb perform carefulli design synthet dataset gain insight lb work condit outperform tb test lb rival algorithm larg suit real world benchmark dataset varieti statist test empir result confirm lb statist significantli outperform altern method term reduc classif error mmi mine multipl inform sourc workshop report mmi mine multipl inform sourc workshop report sail dapprox effici dline algorithm approxim pattern match wildcard length constraint discov relat pattern multipl databas activ learn data stream paper address research problem activ learn data stream data volum grow continu label data consid expens impract object label small portion stream data model deriv predict newli arriv instanc accur order tackl challeng rais data stream dynam natur propos classifi ensembl base activ learn framework select label instanc data stream build accur classifi minim varianc principl introduc guid instanc label data stream addit weight updat rule deriv ensur instanc label process adapt adjust dynam drift concept data experiment result synthet real world data demonstr perform propos effort comparison simpl approach lazi bag classifi imbalanc data empir studi nois impact cost dsensit learn mine complex pattern sequenc gap requir mine frequent pattern wildcard biolog sequenc push frequenc constraint util mine model wordnet disambigu word sens text classif abstract paper propos automat text classif method base word sens disambigu hood algorithm remov word ambigu word replac sens context nearest ancestor sens stopword give document select class document appli algorithm brown corpu effect evalu compar classif result classif result manual disambigu offer princeton univers keyword disambigu word sens text classif wordnet hybrid collabor filter algorithm mixtur expert mine maxim frequent itemset data stream editori special issu mine low dqualiti data mine sequenti pattern time sequenc acceler kernel featur analysi error awar data mine correct classif classifi ensembl correct divers base learner scalabl repres instanc select rank bridg local global data cleans identifi class nois larg classif mine anticip concept chang proactiv dreactiv predict data stream abstract predict stream data import activ modern societi major challeng pose data stream data grow limit difficult retain long histori raw data underli concept data chang time novelti paper fold measur conceptu equival organ data histori histori concept contrast common practic recent raw data concept histori compact retain essenti inform learn learn concept transit pattern concept histori anticip concept case concept chang proactiv prepar predict model futur chang contrast convent methodolog passiv wait chang incorpor proactiv reactiv predict anticip turn correct proper predict model launch instantli class nois handl effect cost dsensit learn cost dguid iter classif filter effici string match wildcard length constraint effect classif noisi data stream attribut dorient dynam classifi select abstract recent mine data stream import challeng task real world applic credit card fraud protect sensor network popular solut separ stream data chunk learn base classifi chunk integr base classifi effect classif paper propos dynam classifi select dc mechan integr base classifi effect mine data stream propos algorithm dynam select singl classifi classifi test instanc run time scheme statist inform attribut valu attribut partit evalu set disjoint subset procedur evalu classif accuraci base classifi subset test instanc attribut valu determin subset similar instanc evalu set construct classifi highest classif accuraci subset select classifi test instanc experiment result compar studi demonstr effici efficaci method dc scheme appear promis mine data stream dramat concept drift signific amount nois base classifi conflict low confid sequenti pattern mine multipl stream paper deal mine sequenti pattern multipl data stream build state art sequenti pattern mine algorithm prefixspan mine transact databas propos mile effici algorithm facilit mine process mile recurs util knowledg exist pattern avoid redund data scan effect speed pattern discoveri process uniqu featur mile incorpor prior knowledg data distribut data stream mine process improv perform extens empir result show mile significantli faster prefixspan mile consum memori prefixspan present solut balanc memori usag time effici memori constrain environ ac aggress classifi ensembl error detect scalabl induct learn partit data combin proactiv reactiv predict data stream mine data stream import scienc commerc major challeng data grow limit difficult retain long histori underli concept data chang time common practic recent raw data paper measur conceptu equival organ data histori histori concept journei concept chang identifi concept appear learn transit pattern concept predict convent methodolog passiv wait concept paper incorpor proactiv reactiv predict proactiv mode anticip concept futur concept chang take place prepar predict strategi advanc anticip turn correct proper predict model launch instantli concept chang promptli resort reactiv mode adapt predict model data system repro propos implement idea experi compar system repres exist predict method benchmark data set repres diversifi scenario concept chang empir evid demonstr propos methodolog effect effici solut predict data stream cost dconstrain data acquisit intellig data prepar video data mine semant index event detect associ perspect appendic video data mine semant index event detect associ perspect abstract advanc media entertain industri includ stream audio digit tv present challeng manag access larg audio visual collect current content manag system support retriev low level featur motion color textur low level featur mean naiv user prefer identifi content high level semant concept creat gap system user bridg system effect end paper present knowledg base video index content manag framework domain specif video basketbal video provid solut explor video knowledg mine associ video data explicit definit evalu measur tempor support confid video associ propos integr inher featur video data approach video process techniqu find visual audio cue court field camera motion activ applaus introduc multilevel sequenti associ mine explor associ audio visual cue classifi associ assign class label appear video construct video indic experiment result demonstr perform propos approach index term video mine multimedia system databas manag knowledg base system insightvideo hierarch video content organ effici brows hierarch video brows featur base video retriev standard method access video content research address benefit integr method effect effici video content access error detect impact dsensit instanc rank noisi dataset noisi dataset locat erron instanc attribut rank suspici instanc base impact system perform interest import research issu provid paper error detect impact sensit instanc rank edir mechan address problem noisi dataset train benchmark classifi instanc effect classifi treat suspici forward subset attribut switch class label train classifi ap instanc ap benchmark classifi locat erron attribut quantit rank instanc defin impact measur base inform gain ratio ir calcul ir attribut ir impact sensit weight sum impact sensit weight locat erron attribut total impact experiment result demonstr effect strategi cost dguid class nois handl effect cost dsensit learn recent research machin learn data mine relat area produc wide varieti algorithm costsensit cs classif maxim classif accuraci minim misclassif cost object method assum train set signific nois rare case real world environ paper systemat studi impact class nois cs learn propos cost guid class nois handl algorithm identifi nois effect cs learn call cost guid iter classif filter cicf seamlessli integr cost exist classif filter nois identif put equal weight handl nois class exist effort cicf put emphasi expens class make success deal dataset larg cost ratio experiment result compar studi real world dataset exist nois corrupt perform cs classifi adopt propos cicf algorithm significantli reduc misclassif cost cs classifi noisi environ data acquisit activ impact dsensit instanc select oidm onlin interact data mine deal predict dbut dunpredict attribut noisi data sourc abstract attribut nois affect classif learn previou work handl attribut nois focus predict attribut predict class attribut attribut predict unpredict predict essenti classif learn import handl nois unpredict requir strategi predict attribut paper present studi identifi cleans measur nois predict unpredict attribut strategi propos theoret analysi empir evid suggest strategi effect effici previou altern class nois attribut nois quantit studi classview hierarch video shot classif index access recent advanc digit video compress network made video access exist content base video retriev system suffer problem semant sensit video classif problem semant gap low level visual featur high level semant visual concept integr video access problem lack effici video databas index automat video annot concept orient summari organ techniqu paper propos framework call classview make advanc effici video databas index access hierarch semant sensit video classifi propos shorten semant gap hierarch tree structur semant sensit video classifi deriv domain depend concept hierarchi video content databas relev analysi select discrimin visual featur suitabl import expect maxim em algorithm determin classif rule visual concept node classifi hierarch video databas index summari present techniqu propos support effect video access larg scale databas hierarch tree structur video databas index scheme determin domain depend concept hierarchi video classif present visual summari integr inher hierarch video databas index tree structur integr video access effici databas index tree structur provid great opportun support power video search engin vdbm testb facil research video databas benchmark real world video base applic requir databas technolog capabl store digit video form video databas provid content base video search retriev method handl tradit data storag queri search retriev present extend provid function vdbm research initi motiv requir videobas applic search retriev portion video data base content testb facil facilit research area video databas manag paper describ vdbm video databas research platform system support comprehens effici databas manag digit video fundament concept provid full rang function video defin abstract databas data type descript paramet applic method research problem address vdbm support handl video data includ mpeg standard multimedia content represent algorithm imag base shot detect imag process techniqu extract lowlevel visual featur high dimension index techniqu access high dimension featur vector extract imag preprocess multimedia queri process optim queri oper real time stream manag search base buffer manag polici access control model select content base access stream video vdbm environ test correct scope video process techniqu measur perform algorithm standard compar perform implement algorithm compon elimin class nois larg dataset paper present approach identifi elimin mislabel instanc larg distribut dataset partit dataset subset small process induct algorithm time construct good rule subset good rule evalu dataset instanc ik error count variabl count number time identifi nois subset instanc higher error valu higher probabl mislabel threshold scheme major object identifi nois experiment result compar studi real world dataset report evalu effect effici propos approach mine video associ effici databas manag support effici video databas manag paper explor concept video associ mine associ pattern character sequenti video shot cluster inform continu video sequenc video shot segment mechan introduc pars discret shot cluster shot visual distinct group construct shot cluster sequenc class label shot associ mine scheme design mine sequenti cluster sequenc detect associ convei valuabl knowledg video databas manag experiment result demonstr effect design hierarch access control model video databas system paper propos approach support multilevel access control video databas access control techniqu combin video databas index mechan hierarch organ visual concept video databas index unit class user access video element video element qualiti level permiss video element access control mechan author object semant cluster subclust video scene video shot video frame salient object region interest paper introduc techniqu obtain multilevel hierarch video content descript summar unifi semant visual similar vide increasingli medium choic varieti commun channeca renneca primarili incr kw le vex ofneh ork multimest system kew ourhesw vide se provid summari tractabl format many xistingapproach limit xplore import low le ve lfe hxl hxl unit summar unfortunatew sew wstpk contew andstructur vide corr stp low le ve lfennww direww ve close stplxpk sce dese stp audio signal proceo drawback of xist mete instel unfold seold structur withinth videi low le ve unit addrei onlyth de import unitsetshhxn strate gybas low le ve lfe ture cannotb appli togehww videw providingusen ov ofth vide contex variousl ve summar eskl wst formor ficiep databas reaba val brows thispap prewpw ahiexwst xk vide contew detewst summarizationstr gysupport nove joint sent visual similaritystr gy todep wk vide contew ficiewk accuratet avid contep de scription ontolog isadoptel variousvid procehxlw teoc niqu util construct asewx st hpkw vide annotationfram work byint grate acquir conte de scription data hienlsthxhw vide contel structur construct groupm ging clustewsth final lay vide summari diffesh granulariti isassewpk assistus unfoldingth vide contei aprogr sive expesthw onrehwl ordvidek valid effect propos approach access video content kei object ip classmin mine medic video scalabl skim summar paper defin clinic oper group shot speaker chang east shot scene blood red close skin region skin region size larger total frame size half shot scene skin region classmin mine medic video content structur event effici access scalabl skim distribut databas server continu media facial featur extract verif omni dface detect video fimag qualit camera motion classif content dbase video index semi dautomat video content annot video databas manag system advanc video databas research environ advanc research develop video databas provid complet video databas manag includ video preprocess content represent index storag manag video metadata indic imag semant base queri process realtim buffer manag continu media stream environ support entir process investig implement analyz evalu techniqu identifi concret techniqu practic robust paper present video databas research initi culmin success develop vdbm video databas research platform support comprehens effici databas manag digit video describ kei video process compon system illustr vdbm research platform describ research project carri vdbm environ includ mpeg document support video featur import export queri oper optim multi featur imag similar match secur access control stream video mine medic video data hierarch content organ seed imag segment content dbase imag retriev applic hierarch video summar medic data provid user overview medic video content level abstract effici databas brows access hierarch video summar strategi develop present paper gener overview kei frame video preprocess extract special frame black frame slide clip art sketch draw special region face skin blood red area shot group method appli merg spatial tempor relat shot group visual featur knowledg video shot integr assign group predefin semant categori base video group semant categori video summari level construct group merg hierarch group cluster semant categori select base strategi user select layer summari access higher layer concis video summari lower layer greater detail contain summari model dbase video classif hierarch represent seed semant object gener content dbase video index automat scene detect new program integr visual featur rule automat model dbase semant object extract algorithm queri refin semant integr imag retriev system semiautomat annot scheme relev feedback power wide techniqu content base imag retriev system relev feedback approach weight featur sum feedback imag optim queri refin imag similar assess approach work case user express relationship queri paper propos method weight distanc sum minim distanc md minim distanc rank mdr measur similar imag databas feedback imag queri refin experiment comparison propos relev feedback scheme mdr method md method describ user multipl intent base scheme imag retriev semiautomat annot system ifind lu hu zhu zhang yang unifi framework semant featur base relev feedback imag retriev system proc acm intern confer pp zhu liu zhang wu imag retriev semi automat annot scheme larg imag databas web spie photon west jan electron imag integr queri refin semant inform present experi show propos method result substanti improv retriev accuraci retriev annot larg imag databas multiview multilevel video content represent retriev articl practic algorithm propos support content base video analysi model represent summar index access multilevel video databas model advantag model reason approach bridg gap low level repres featur high level semant concept human point view model base video analysi techniqu propos order detect video shot present techniqu adapt threshold scene cut detect activ variant video video shot seed region aggreg tempor track techniqu propos gener semant video object semant video scene gener extract video access unit shot object domain knowledg order categor video content set semant cluster integr video classif techniqu develop support effici multilevel video represent summar index access techniqu spie amp doi 