adapt cost dsensit learn reject option gener queri ambigu oracl supervis learn minim effort real dcode biogeographi dbase optim mutat gener queri domain expert improv learn abstract assist domain expert activ learn select construct fewer exampl request label build accur classifi previou work activ learn gener specif queri real world applic domain expert oracl readili answer gener queri don care attribut power gener queri gener queri equival specif overli gener queri good answer domain expert oracl highli uncertain make learn difficult paper propos activ learn algorithm ask good gener queri extend algorithm construct hierarch featur nomin numer attribut demonstr experiment method ask significantli fewer queri compar previou work activ learn initi label dataset small oracl inaccur class probabl estim method readili deploi real world data mine task obtain label exampl costli extens paper publish ieee icdm keyword activ learn domain expert gener queri dtrain handwritten digit recognit dtrain handwritten digit recognit hybrid differenti evolut base fuzzi dmean cluster hybrid differenti evolut base fuzzi dmean cluster activ learn gener queri abstract activ learn activ select construct exampl label reduc number label exampl need build accur classifi previou work activ learn specif queri predict osteoarthr patient dataset attribut specif queri valu attribut irrelev natur gener queri don care attribut peopl knee pain osteoarthr attribut ag type pain assum oracl human expert readili answer gener queri return probabilist label power gener queri gener queri equival specif overli gener queri receiv highli uncertain label oracl make learn difficult paper propos activ learn algorithm ask gener queri demonstr experiment method ask significantli fewer queri compar previou work activ learn method readili deploi real world task obtain label exampl costli keyword activ learn gener queri supervis learn dtrain work real data abstract train paradigm semi supervis learn promis allevi effect shortag label exampl supervis learn standard view train requir dataset view featur previou studi shown train work view satisfi suffici independ assumpt practic assumpt ensur view commonli supervis dataset set attribut view split view order appli standard twoview train paper propos approach empir verifi assumpt train view design method split singl view dataset view order make train work reliabl empir result show larg label train set view verif split method effect train call precis label train set small small label train set show train assumpt difficult verifi view split unreli conclus train effect mix view satisfi assumpt train work base small label train set verifi assumpt split singl view view unreli uncertain standard train work semi dsupervis learn reconstruct manifold space caricatur gener advanc data mine applic advanc data mine applic requir support relat databas engin cluster applic high dimension featur space demand proper support multipl top queri order perform project cluster research tackl problem optim restrict rank top queri solut singl rank criterion deficit optim multipl topk queri join target paper perspect hand propos minim handi extens sql express multipl top queri hand propos optim hash join strategi effici execut type queri extens experi conduct context show feasibl propos semi dsupervis dtrain sentenc subject classif proper model select signific test abstract model select import ubiquit task machin learn select model futur classif perform measur goal metric anevalu metric select classif model compet common practic goal evalu metric recent studi claim evalu metric auc goal metric accuraci result select correct model paper point flaw experiment design studi propos improv method test claim extens experi show convincingli goal metric reliabl select correct classif model discrimin paramet learn bayesian network bayesian network classifi wide classif problem fix bayesian network structur paramet learn approach gener discrimin learn gener paramet learn effici discrimin paramet learn effect paper propos simpl effici effect discrimin paramet learn method call discrimin frequenc estim dfe learn paramet discrimin comput frequenc data empir studi show dfe algorithm integr advantag gener discrimin learn perform state art discrimin paramet learn method elr accuraci significantli effici activ learn direct queri construct activ learn hold kei solv data scarciti problem supervis learn lack label data label data costli process activ learner request label select instanc reduc label work dramat previou work activ learn pool base pool unlabel exampl learner select exampl pool queri label type activ learn weak paper propos activ learn algorithm construct exampl directli queri label studi specif activ learner base decis tree algorithm gener activ learner work base learn algorithm restrict exampl queri method shown queri fewer exampl reduc predict error quickli cast doubt us pool pool base activ learn method easili adapt work pool unlabel exampl activ learn direct queri construct activ learn hold kei solv data scarciti problem supervis learn lack label data label data costli process activ learner request label select instanc reduc label work dramat previou work activ learn pool base pool unlabel exampl learner select exampl pool queri label type activ learn weak paper propos activ learn algorithm construct exampl directli queri label studi specif activ learner base decis tree algorithm gener activ learner work base learn algorithm restrict exampl queri method shown queri fewer exampl reduc predict error quickli cast doubt us pool pool base activ learn method easili adapt work pool unlabel exampl cost dsensit decis tree pre dprune cost dsensit decis tree pre dprune multi dobject genet algorithm base densiti machin learn stock select partial acquisit cost dsensit learn extract action knowledg decis tree abstract data mine algorithm tool stop discov custom model produc distribut inform custom profil techniqu appli industri problem custom relationship manag crm point custom attritor custom loyal requir human expert postprocess discov knowledg manual postprocess techniqu limit produc visual result interesting rank directli suggest action lead increas object function profit paper present algorithm suggest action chang custom undesir statu attritor desir loyal maxim object function expect net profit algorithm discov costeffect action transform custom undesir class desir approach integr data mine decis make tightli formul decis make problem directli top data mine result postprocess step improv effect approach present ensembl decis tree shown robust train data empir test conduct realist insur applic domain uci benchmark data index term phrase decis make data mine machin learn custom classif learn base queri project cost dsensit test strategi medic diagnosi doctor determin medic test rai blood test order patient minim total cost medic test misdiagnosi paper design cost sensit machin learn algorithm model learn diagnosi process medic test attribut machin learn valu obtain cost attribut cost misdiagnos misclassif incur cost misclassif cost propos improv decis tree learn algorithm minim sum attribut cost misclassif cost design test strategi request obtain valu unknown attribut cost similar doctor order medic test cost order minim total cost test exampl patient empir evalu compar test strategi show effect outperform previou method case studi heart diseas threshold make classifi cost dsensit od fast robust differenti evolut base orthogon design keyphras extract semant network structur analysi keyphras plai kei role text index summar categor exist keyphras extract approach requir human label train set paper propos automat keyphras extract algorithm featur weight supervis unsupervis task algorithm treat document semant network hold syntact statist inform structur dynam network easili identifi kei node extract keyphras unsupervisedli experi demonstr propos keyphras extract algorithm averag improv effect effici unsupervis task perform compar supervis extractor appli algorithm supervis task develop accur classifi classifi assembl syntact statist featur experi show precis supervis extract construct ensembl rank featur acquisit test sequenti batch test algorithm medic diagnosi doctor order set medic test sequenc order make accur diagnosi patient diseas make trade cost test misdiagnosi paper cost sensit learn model process assum test exampl patient miss valu actual valu acquir cost similar medic test order reduc misclassif error misdiagnosi propos sequenti batch test algorithm acquir set attribut valu sequenc similar set medic test order doctor sequenc goal algorithm minim total cost trade acquir attribut valu misclassif demonstr effect algorithm show outperform previou method significantli algorithm readili appli real world diagnosi task case studi heart diseas paper maximum profit mine applic softwar develop softwar defect bug correct test part lengthi softwar develop cycl enterpris softwar vendor releas softwar product report defect correct due deadlin limit resourc small number defect escal custom resolv immedi softwar vendor high cost paper develop escal predict ep system mine histor defect report data predict escal risk defect maximum net profit specif describ simpl gener framework convert maximum net profit problem cost sensit learn appli compar cost sensit learn approach ep experi suggest cost sensit decis tree method produc highest posit net profit comprehens result ep system deploi successfulli product group enterpris softwar vendor note paper accept kdd indextoolkit open sourc toolbox index protein databas high dthroughput proteom discov classif data multipl sourc abstract larg commerc organ multipl data sourc describ custom import consolid data multipl sourc intellig busi decis make paper propos method predict classif data multipl sourc class label sourc test method artifici real world dataset show classifi data accur machin learn perspect method remov fundament assumpt provid class label supervis learn bridg gap supervis unsupervis learn learn contextu depend network model link dbase classif link object rich semant help classifi object irrelev link found real world link data web page noisi irrelev link provid predict inform categor import automat identifi link relev categor paper present contextu depend network cdn model classifi link object presenc noisi irrelev link cdn model make depend function character contextu depend link object cdn differenti impact relat object classif reduc effect irrelev link classif show learn cdn model effect gibb infer framework learn model collect classif multipl link object experi show cdn model demonstr high robust data set irrelev link test dcost sensit classif data miss valu area cost sensit learn induct learn algorithm extend handl type cost repres misclassif error previou work focus deal misclassif cost paper address equal import issu handl test cost queri miss valu test case attribut miss test case worthwhil extra effort order obtain attribut attribut depend benefit bring increas accuraci paper integr test cost sensit learn handl miss valu unifi framework includ model build test strategi test strategi determin attribut perform test order minim sum classif cost test cost show instanti framework popular machin learn algorithm decis tree naiv bayesian method empir evalu test cost sensit method handl miss valu data set test strategi cost dsensit decis tree abstract studi cost sensit learn decis tree incorpor test cost misclassif cost propos lazi decis tree learn minim total cost test misclassif assum test exampl unknown attribut valu obtain cost test cost design test strategi attempt minim total cost test misclassif test empir evalu treebuild test strategi show effect result readili appli real world diagnosi task medic diagnosi doctor determin test blood test order patient minim total cost test misclassif misdiagnosi case studi heart diseas paper simpl test strategi cost dsensit decis tree abstract studi cost sensit learn decis tree incorpor test cost misclassif cost propos lazi decis tree learn minim total cost test misclassif assum test exampl unknown attribut valu obtain cost test cost design test strategi attempt minim total cost test misclassif test empir evalu treebuild test strategi show effect result readili appli real world diagnosi task medic diagnosi doctor determin test blood test order patient minim total cost test misclassif misdiagnosi case studi heart diseas paper partial ensembl classifi select rank predict softwar escal maximum roi enterpris softwar vender releas softwar product report defect correct small number report defect escal custom busi impact escal defect quickli resolv high cost softwar vendor total cost greater includ loss reput satisfact loyalti repeat revenu paper develop escal predict ep system mine histor defect report data predict escal risk current defect report maximum roi return invest specif describ simpl gener framework convert maximum roi problem cost sensit learn appli compar cost sensit learn approach ep ep system produc promis result deploi product group enterpris softwar vendor conclus drawn studi provid guidelin mine imbalanc dataset costsensit learn hybrid cost dsensit decis tree abstract cost sensit decis tree cost sensit na ve bay cost sensit learn model propos recent minim total cost test misclassif advantag disadvantag paper propos cost sensit learn model hybrid cost sensit decis tree call dtnb reduc minimum total cost integr advantag cost sensit decis tree cost sensit na ve bay empir evalu test strategi experi show dtnb outperform cost sensit decis cost sensit na ve bay significantli minim total cost test misclassif base sequenti test strategi singl batch strategi rank measur order abstract data mine applic requir rank mere classif case exampl applic widespread includ internet search engin rank page return custom relationship manag rank profit custom theoret foundat practic guidelin establish assess merit rank measur order paper review gener criteria judg merit singl number measur propos rank measur compar commonli rank measur criteria lead prefer order rank measur conduct experi real world dataset confirm prefer order result paper evalu compar rank algorithm dynam ensembl dconstruct rank abstract ensembl learn shown success data mine work ensembl learn concern task classif work construct ensembl aim improv rank paper propos approach construct ensembl base ensembl purpos improv rank perform crucial data mine task experi real world data set show approach achiev signific improv rank origin bag adaboost ensembl pfind databas dsearch softwar system autom peptid protein identif tandem mass spectrometri miss miss valu cost dsensit decis tree abstract real world data set machin learn data mine miss valu previou research problem attempt imput miss valu train test paper studi issu costsensit learn consid test cost misclassif cost attribut test expens obtain valu cost effect miss valu similar skip expens riski test miss valu patient diagnosi classif miss miss valu reduc total cost test misclassif meaning imput valu discuss compar strategi util valu miss cost reduct cost sensit decis tree learn index term induct knowledg acquisit machin learn auc accuraci evalu learn algorithm appendic auc accuraci evalu learn algorithm area roc receiv oper characterist curv simpli auc recent propos altern singl number measur evalu predict abil learn algorithm formal argument auc prefer accuraci paper establish formal criteria compar measur learn algorithm show theoret empir auc gener measur defin precis accuraci reevalu establish claim machin learn base accuraci auc obtain interest surpris result show auc directli net profit accuraci direct market suggest learn algorithm optim auc accuraci real world applic conclus drawn paper make signific impact machin learn data mine applic note paper integr result paper publish ijcai icdm includ result concept indiffer section ii section iii iii iv iv unpublish index term evalu learn algorithm auc accuraci roc guest editor introduct special issu machin learn bioinformat part guest editor introduct special issu machin learn bioinformat dpart multi dobject genet algorithm base quick sort artifici ag face support vector machin test dcost sensit naiv bay classif induct learn techniqu naiv bay decis tree algorithm extend past handl type cost distinguish cost classif error equal import issu handl test cost queri miss valu test case attribut miss test case worthwhil effort obtain miss depend result potenti gain classif accuraci paper show obtain test cost sensit naiv bay classifi csnb includ test strategi determin unknown attribut select perform test order minim sum misclassif cost test cost propos evalu potenti test strategi includ test empir evalu csnb method show compar favor decis tree counterpart decis tree minim cost propos simpl effect method build test decis tree minim sum misclassif test cost specif put forward origin simpl split criterion attribut select tree build treebuild algorithm desir properti cost sensit learn system account type cost assum test case larg number miss valu design intellig test strategi suggest wai obtain miss valu cost order minim total cost experiment compar strategi demonstr algorithm significantli outperform variat addit algorithm complex similar lower previou work work diagnost task factor misclassif test cost obtain miss inform method construct ddomin set multi dobject genet algorithm kernel dbase case retriev algorithm applic bioinformat exploit kernel trick correl fragment ion peptid identif tandem mass spectrometri motiv correl fragment ion tandem mass spectrum crucial reduc stochast mismatch peptid identif databas search effici score algorithm consid correl inform tunabl comprehens manner lack result paper promis approach util correl inform improv peptid identif accuraci kernel trick root statist learn theori exploit address issu low comput effort common score method tandem mass spectral dot product sdp extend kernel sdp ksdp experi dataset report previous demonstr effect ksdp implement consecut fragment show decreas error rate compar sdp softwar tool pfind simpl score function base ksdp outperform sdpbase softwar tool sequest sonar ms ms term identif accuraci auc measur accuraci compar learn algorithm predict accuraci wide main criterion compar predict abil classif system neural network naiv bay classifi produc probabl estim classif complet accuraci measur grant train test set provid class label paper establish rigour set area roc receiv oper characterist curv simpli auc measur accuraci result signific reason establish time rigour criteria compar evalu measur learn algorithm suggest auc replac accuraci measur compar classif system result prompt evalu establish conclus base accuraci machin learn accept machin learn commun term predict accuraci naiv bay decis tree similar auc show experiment naiv bay significantli decis tree learn algorithm fundament issu naiv bay naiv bay effici effect induct learn algorithm machin learn data mine condit independ assumpt base rare true real world applic research extend naiv bay repres depend explicitli propos relat learn algorithm base depend paper argu classif point view depend distribut plai crucial role depend propos explan superb classif perform naiv bay verifi idea design conduct experi extend chow liu algorithm depend distribut construct tan mutual inform reflect depend attribut empir result provid evid support explan compar naiv bay decis tree svm auc accuraci predict abil classif algorithm typ ical measur predict accuraci error rate minu accuraci test exampl classifi includ naiv bay produc probabl estim confid classpredict inform complet accuraci postprocess decis tree extract action knowledg data mine algorithm tool stop discov custom model produc distribut inform custom profil techniqu appli industri problem custom relationship manag crm point custom attritor custom loyal requir human expert postprocess mine inform manual postprocess techniqu limit produc visual result interesting rank directli suggest action lead increas object function profit paper present algorithm suggest action chang custom undesir statu attritor desir loyal maxim object function expect net profit develop algorithm resourc constraint abound realiti contribut work take output exist matur techniqu decis tree produc action knowledg automat postprocess decis tree rank auc area curv roc receiv oper characterist recent measur rank perform learn algorithm paper present probabl estim algorithm improv auc decis tree estim probabl singl leaf fall method averag probabl estim leav tree auc statist consist discrimin measur accuraci predict accuraci main evalu criterion predict perform classif learn algorithm recent year area roc receiv oper characterist curv simpli auc propos altern singl number measur evalu learn algorithm paper prove auc measur accuraci specif present rigour definit consist discrimin compar evalu measur learn algorithm present empir evalu formal proof establish auc statist consist discrimin accuraci result signific formal prove time auc measur accuraci evalu learn algorithm intellig protein structur retriev system mine optim action profit crm data mine appli crm custom relationship manag industri witha limit success data mine tool discov custom model profil custom attritor custom loyal action improv custom relationship chang attritor loyal custom describ algorithm suggest action chang custom undesir statu attritor desir loyal algorithm take account cost action attempt maxim expect net profit knowledg data miningalgorithmsor tool todai accomplish import task crm algorithm implement advanc featur special highli effect data mine softwar call proactiv solut represent upper bound bayesian network fundament issu bayesian network represent power reflect kind function repres paper prove upper bound represent power augment naiv bay extend result gener bayesian network bayesian classifi accur probabl data mine applic accur rank probabl estim essenti tradit classifi aim high classif accuraci low error rate produc probabl estim high predict accuraci impli rank probabl estim evalu method classifi classif accuraci purpos data mine applic answer area roc receiv oper characterist curv simpli auc improv encarta search engin perform mine user log editori note learn prosod pattern mandarin speech synthesi duminda wijesekera daniel barbara session discuss variat multimedia data mine simeon simoff discuss multimedia mine encompass open issu coffe break sigkdd confer open award viii ix tabl content geo spatial cluster user constraint anthoni tong raymond ng lak lakshmanan jiawei han multi level index gi enhanc learn satellit imageri krzysztof koperski giovanni marchisio predict locat map similar plum framework spatial data mine sanjai chawla shashi shekhar weili wu uygar ozesmi learn prosod pattern mandarin speech synthesi yiqiang chen wen gao tingshao zhu represent power discret bayesian network import fundament properti bayesian network represent power reflect kind function repres paper establish associ structur complex bayesian network represent power maximum number node parent measur structur complex bayesian network maximum xor contain target function measur function complex represent upper bound establish prove roughli speak discret bayesian network node parent repres function xor theoret result gain deeper understand capac limit bayesian network learn good prototyp classif filter abstract instanc geometr properti naiv bay nomin domain numer approach learn classifi decis tree neural network instanc base learn studi recent year probabl approach learn classifi extens investig bay theorem probabl delta delta delta class mine gener queri pattern web log user log popular search engin track user activ includ user queri user click return list user brows behavior knowledg user queri discov user log improv perform search engin propos data mine approach produc gener queri pattern templat raw user log popular commerci knowledg base search engin simul show templat improv search engin speed precis cover queri ask previous templat comprehens web editor easili discov topic user interest introduct data mine process discov implicit knowledg larg dataset fayyad al knowledg dataset simpl statist inform mean deviat data mine aim find higher level power learnabl augment naiv bay nonim domain data mine direct market problem solut direct market process identifi buyer product promot product improv learn algorithm augment naiv bay data mine applic requir learn algorithm high predict accuraci scale larg dataset produc comprehens outcom naiv bay classifi receiv extens attent due effici reason predict accuraci simplic assumpt attribut depend class naiv bay violat produc incorrect probabl affect success data mine applic extend naiv bay classifi depend relat attribut compar previou extens naiv bay algorithm effici problem larg number attribut produc simpler depend relat comprehens maintain similar predict accuraci learnabl naiv bay naiv bay effici effect learn algorithm previou result show represent abil sever limit repres linearli separ function binari domain give suffici condit linearli separ function binari domain learnabl naiv bay uniform represent show learnabl error rate naiv bay affect dramat sampl distribut result gain deeper understand seemingli simpl power learn algorithm graphem gener learn read english word data mine direct market problem solut direct market process identifi buyer product promot product comput cognit model sourc power relat issu align algorithm learn read aloud comput optim attribut weight set nearest neighbor algorithm learn classif multipl sourc unsupervis data decis dtree model balanc scale develop present altern model human cognit develop balanc scale task studi task inspir wide rang human comput work task requir children predict outcom place discret number weight distanc side fulcrum model featur symbol learn algorithm transit mechan exhibit regular found human data includ orderli stage progress shape develop torqu differ effect unlik previou success model task current model singl free paramet restrict size balanc scale accommod requir assumpt highli structur output represent train environ bias weight distanc inform model make number predict differ previou comput effort keyword cognit develop bala overfit gener learn discret pattern invert implic small train set present algorithm induc recurs claus invers implic invers resolut underli gener method approach appli class logic program similar class primit recurs function induct perform small number posit exampl resolut path algorithm implement system name crustacean locat match list gener term determin pattern decomposit exhibit target recurs claus theoret analysi defin class logic program approach complet term characterist ilp approach current implement consider faster previous report present evid demonstr randomli select input increas number posit exampl increas accuraci reduc number output relat approach similar learn recurs relat randomli select small train set evalu crustacean induct logic program algorithm invers implic induc recurs claus exampl approach suit learn class recurs claus commonli logic program search common substructur exampl evid exist invers implic approach perform randomli select posit neg exampl show crustacean learn recurs relat higher accuraci golem reason effici demonstr increas number randomli select posit neg exampl increas accuraci randomli select test exampl increas frequenc output target relat reduc number output prove theorem defin class logic program approach complet motiv paper extend previou work ah learn past tens english verb symbol pattern associ connectionist model learn past tens english verb seemingli minor aspect languag acquisit gener heat debat landmark task test adequaci cognit model artifici neural network ann implement challeng symbol model pose paper present gener purpos symbol pattern associ spa base decis tree learn algorithm id conduct extens head head comparison gener abil ann model spa represent conclud spa gener past tens unseen verb ann model wide margin offer insight case discuss default strategi decis tree learn algorithm introduct learn past tens english verb seemingli minor aspect languag acquisit gener heat deb symbol model nonconsci acquisit inform paper present counter evid smolenski theori human intuit nonconsci cognit process accur explain term subsymbol comput carri artifici neural network present symbol learn model studi complic cognit task involv nonconsci acquisit inform learn product rule artifici finit state grammar result demonstr intuit learn impli subsymbol comput establish perceiv correl consciou symbol hand nonconsci subsymbol exist learn past tens english verb symbol pattern associ connectionist model learn past tens english verb seemingli minor aspect languag acquisit gener heat debat landmark task test adequaci cognit model artifici neural network ann implement challeng symbol model pose paper present gener purpos symbol pattern associ spa base decis tree learn algorithm id conduct extens head head comparison gener abil ann model spa represent conclud spa gener past tens unseen verb ann model wide margin offer insight case discuss default strategi decis tree learn algorithm introduct learn past tens english verb seemingli minor aspect languag acquisit gener heat deb learn control dynam system automat quantiz learn control dynam system unknown model challeng research problem previou work learn qualit control rule construct qualit state proper partit continu state variabl design human user learn program design learn method learn qualit state represent control rule simultan method aggress partit continu state variabl finer discret rang control rule base rang learn case studi appli method benchmark control problem cart pole balanc invert pendulum experiment result show method deriv partit cart pole system paramet learn control system extend period time random initi posit construct induct logic program autom construct semant grammar difficult interest problem machin learn paper show semant grammar acquisit problem view learn search control heurist logic program control rule learn order induct algorithm automat invent syntact semant categori empir result show learn parser gener sentenc perform previou approach base connectionist techniqu introduct design comput system understand natur languag input difficult task labori hand craft comput grammar support natur languag applic ineffici incomplet ambigu difficulti construct adequ grammar knowledg acquisit bottleneck motiv research machin learn numer research studi symbol model learn past dtens english verb support elimin connection argu pattern associ base explan languag learn languag process deni explicit rule symbol represent plai role languag process cognit gener argument base larg extent artifici neural network ann model claim learn past tens english verb rumelhart mcclelland macwhinnei leinbach articl critic review rumelhart mcclelland macwhinnei leinbach ann model conclud succe assign task learn past tens english verb order answer challeng symbol process approach present symbol pattern associ spa gener purpos pattern associ learn associ arbitrari discret pattern carri experi spa set verb induct learn good exampl studi kind data eas comput complex learn horn claus theori gold paradigm boolean function pac learn paradigm give definit good data basic gener repres set develop data driven algorithm learn faster good exampl degener learn limit worst exampl show horn claus theori kterm dnf gener dnf boolean function polynomi learnabl gener repres present introduct induct learn model data target theori suppli learn program crucial assumpt identif limit gold assum seri exampl admiss enumer posit neg exampl target concept requir learn algorithm produc correct hypothesi finit time comput time 