commun detect mine social media research simul traffic accid scene opengl fuzzi expert system base intellig websit assess system research multi dagent base market cooper system simul blood vessel surgeri simul client dside load balanc cloud social network analysi approach detect suspici onlin financi activ converg influenti blogger topic discoveri blogospher sentiment propag social network case studi livejourn integr approach indentifi biolog relev gene remot network lab ddemand network cloud configur test predict collect behavior social dimens extract guest editor introduct social comput blogospher wiscol collect wisdom base blog cluster model data mine blogospher scalabl learn collect behavior base spars social dimens studi collect behavior understand individu behav social network environ ocean data gener social media facebook twitter flickr youtub present opportun challeng studi collect behavior larg scale work aim learn predict collect behavior social media inform individu infer behavior unobserv individu network social dimens base approach adopt address heterogen connect present social media network social media coloss size involv hundr thousand million actor scale network entail scalabl learn model collect behavior predict address scalabl issu propos edg centric cluster scheme extract spars social dimens spars social dimens socialdimens base approach effici handl network million actor demonstr compar predict perform scalabl method categori andsubjectdescriptor databas manag databas applic comput median valu cloud environ gridbatch mapreduc quantifi util trustworthi advic share onlin social media abstract grow popular social media recent year result creation enorm amount user develop content inform readili easi find content detect trustworthi casual observ differenti useless trustworthi untrustworthi work studi problem quantifi user share content focuss health content neg impact higher domain advic share health social network daili strength studi describ defin notion trustworthi util social media content identifi necess challeng assess propos framework help address challeng identifi relev featur provid empir mean meet requir evalu select relev variabl perform numer experi evalu model result demonstr promis perform possibl replic social media applic web server farm cloud perform evalu dynam architectur web server farm cloud perform evalu dynam architectur uncovern group heterogen interact analysi integr knowledg search biolog relev gene social ident approach identifi familiar stranger social network present problem search familiar stranger social network familiar stranger individu directli connect exhibit similar power law natur social network determin major individu directli connect small number fellow individu similar individu larg unknown individu social network local view network make problem aggreg familiar stranger challeng work formul problem show signific address challeng present approach innov emploi social ident individu competit approach blogger citat network showcas technic detail empir result relat issu futur work blogtrack tool sociologist track analyz blogospher present tool blogtrack assist sociologist track analyz blog interest design integr uniqu featur present overview blogtrack illustr function compon blogtrack outlin futur work expans meet grow sociologist connect ident commun interest challeng area social comput social media analysi call commun analysi barrier cross commun multipl websit analysi disconnected websit paper aim provid evid exist map ident multipl commun provid method connect websit studi shown simpl effect approach leverag social media collect pattern util find map emploi method successfulli reveal map accuraci multiclass probabilist kernel discrimin analysi kernel discrimin analysi kda effect approach supervis nonlinear dimension reduct probabilist model kda improv robust state art model handl binari class problem confin applic real world problem overcom limit propos nonparametr probabilist model base gaussian process kda handl multiclass problem model bayesian interpret kda paramet automat tune optim margin loglikelihood data empir studi demonstr efficaci propos model build distribut robot garden evalu trustworthi wikipedia articl qualiti credibl wikipedia popular destin web surfer seek knowledg wide varieti subject help articl accur inform consist unreli articl inaccur incomplet inform casual observ differenti good bad work identifi necess challeng trust assess wikipedia propos framework address challeng identifi relev featur provid empir mean meet requir evalu select relev variabl perform experi evalu approach result demonstr promis perform compar approach possibl replic social media applic relat learn latent social dimens social media blog facebook flickr present data network format classic iid distribut address interdepend data instanc relat learn propos collect infer base network connect adopt predict connect social media multi dimension actor connect actor due factor alumni colleagu live citi share similar interest collect infer differenti connect work propos extract latent social dimens base network inform util featur discrimin learn social dimens describ affili social actor hidden network subsequ discrimin learn automat determin affili align class label scheme prefer multipl divers relat network conduct extens experi social media data real world blog site popular content share site model outperform repres relat learn method base collect infer label data sensit model connect exist method carefulli examin remot network lab ddemand network cloud configur test manipul data dimens reduct method featur select trust blogospher cross dvalid leav cross valid scheme gener assess neural network model computation expens due replic train session linear unlearn exampl recent suggest approach approxim crossvalid briefli review linear unlearn scheme dub luloo illustr system identif address possibl extract confid inform error bar luloo ensembl introduct nonlinear regress output regress nonlinearli input vector report concern neural network implement output predict delta denot nonlinear map neural net vector network paramet condit input output distribut probabl distribut output condit test input basic object neural net model chi dsquar approxim distribut dneg definit quadrat form dcentral normal variabl search interact featur subset select special issu paper sdm introduct special issu social comput introduct special issu social comput understand effect sampl healthcar risk model predict futur high dcost patient gridbatch cloud comput larg dscale data dintens batch applic healthcar risk model medicaid patient impact sampl predict high dcost patient cluster blog collect wisdom blogospher expand unpreced speed understand blogospher greatli facilit develop social web serv user servic provid advertis import task process cluster blog site cluster blog site present challeng propos tap collect wisdom cluster blog site present statist visual result report find suggest futur work extend real world applic robust integr multipl inform sourc view complet applic multipl data sourc featur integr order perform infer task optim research shown task webpag classif imag classif pattern recognit combin data multipl inform sourc yield significantli result singl sourc task multipl data sourc thought provid view underli object domain view instanc view miss problem miss view affect perform machin learn task paper provid method view complet heurist predict miss view show view complet achiev significantli result show inform higher level term view lower level term featur achiev result demonstr compar method exist method miss view problem miss problem heterogen data fusion alzheim diseas studi commun evolut dynam multi dmode network multi mode network typic consist multipl heterogen social actor type interact occur identifi commun multi mode network understand structur properti network address data shortag unbalanc problem assist task target market find influenti actor group gener network membership group evolv gradual dynam multi mode network actor membership interact evolv pose challeng problem identifi commun evolut work address issu emploi tempor inform analyz multi mode network spectral framework scalabl issu carefulli studi experi synthet data real world larg scale network demonstr efficaci algorithm suggest gener solv problem complex relationship identifi biolog relev gene multipl heterogen data sourc select gene differenti express critic biolog process major challeng post arrai analysi recent develop bioinformat made data sourc mrna mirna express profil biolog pathwai gene annot effici effect integr multipl data sourc help enrich knowledg involv sampl gene select gene bear signific biolog relev work studi problem multi sourc gene select multipl heterogen data sourc data set select gene express profil integr inform data sourc investig effect emploi inform contain multipl data sourc extract intrins global geometr pattern covari analysi gene select design conduct experi systemat compar propos approach repres method term statist biolog signific show efficaci potenti propos approach promis find protect privaci increment mainten distribut associ rule mine socialtagg collabor tag blog long tail social bookmark process user share tag onlin resourc blog collabor tag provid valuabl metadata retriev system success collabor tag system demonstr popular websit del icio site cover small fraction blog web vast major blog collabor tag system tag author lack coverag collabor tag consider roadblock tag metadata web scale inform retriev system solv problem propos implement system automat recommend collabor tag blog automat gener tag surfac blog make social book mark site easili discov potenti tag wider popul identifi influenti blogger commun blog popular web user publish inform web blogger write blog post share like dislik voic opinion provid suggest report new form group blogospher blogger form virtual commun similar interest activ happen blogospher affect extern world understand develop blogospher find influenti blog site influenti blog site form long tail blog site influenti influenti blogger inspir high impact influenti physic commun studi problem identifi influenti blogger blog site activ blogger necessarili influenti influenti blogger impact fellow blogger wai paper discuss challeng identifi influenti blogger investig constitut influenti blogger present preliminari model attempt quantifi influenti blogger pave build robust model find type influenti illustr issu conduct experi data real world blog site evalu multi facet problem identifi influenti blogger discuss uniqu challeng conclud interest find futur work dimension reduct area scienc complet miss view multipl sourc web media physic topolog design doptic network abstract advantag deploi allopt network guidelin properli design physic topolog network propos effici physic topolog design algorithm asymptot growth rate provis capac metric compar design altern higher asymptot growth rate translat directli higher deploy cost larg network studi show take fiber length consider lead lower capac requir find suffici larg fiber node ratio order minim asymptot growth provis capac increas capac util minim wavelength convers studi real network find fiber node ratio low result larg provis capac requir capac usabl increas ratio reduc provis capac achiev close util blogospher research issu tool applic weblog blog facilit peopl express thought voic opinion share experi idea individu experi sens commun feel belong bond member matter nich met onlin interact open standard low barrier public transform inform consum produc creat plethora open sourc intellig collect wisdom act storehous overwhelm amount knowledg member environ symbiosi nonetheless vast amount knowledg remain discov exploit suitabl paper introduc state art research issu review kei element research tool methodolog blogospher present case studi identifi influenti blogger commun exemplifi integr major aspect discuss paper end compar contrast blogospher social network research introduct topic taxonomi adapt group profil predict futur high dcost patient real dworld risk model applic adapt distanc metric learn cluster good distanc metric crucial data mine task learn metric unsupervis set metric learn algorithm project observ data lowdimension manifold geometr relationship pairwis distanc preserv extend nonlinear case appli kernel trick emb data featur space kernel function comput dot product data point featur space paper propos unsupervis nonlinear adapt metric learn algorithm call naml perform cluster distanc metric learn simultan naml map data high dimension space kernel function appli linear project find low dimension manifold separ data maxim final perform cluster low dimension space perform naml depend select kernel function project show joint kernel learn dimension reduct cluster formul trace maxim problem solv iter procedur em framework experiment result demonstr efficaci propos algorithm spectral featur select supervis unsupervis learn featur select aim reduc dimension build comprehens learn model good gener perform featur select algorithm larg studi separ type learn supervis unsupervis work exploit intrins properti underli supervis unsupervis featur select algorithm propos unifi framework featur select base spectral graph theori propos framework gener famili algorithm supervis unsupervis featur select show exist power algorithm relieff supervis laplacian score unsupervis special case propos framework knowledg work attempt unifi supervis unsupervis featur select enabl joint studi gener framework experi demonstr efficaci algorithm deriv framework search interact featur present web search interact featur queri link websit frequent visit user similar inform popular destin complement tradit search result allow direct navig authorit resourc queri topic destin identifi histori search brows behavior user extend time period collect behavior basi comput sourc author describ user studi compar suggest destin previous propos suggest relat queri tradit unaid web search result show search enhanc destin suggest outperform system exploratori task perform obtain mine past user behavior queri level granular nonlinear adapt distanc metric learn cluster good distanc metric crucial data mine task learn metric unsupervis set metric learn algorithm project observ data lowdimension manifold geometr relationship pairwis distanc preserv extend nonlinear case appli kernel trick emb data featur space kernel function comput dot product data point featur space paper propos unsupervis nonlinear adapt metric learn algorithm call naml perform cluster distanc metric learn simultan naml map data high dimension space kernel function appli linear project find low dimension manifold separ data maxim final perform cluster low dimension space perform naml depend select kernel function project show joint kernel learn dimension reduct cluster formul trace maxim problem solv iter procedur em framework experiment result demonstr efficaci propos algorithm time ddepend event hierarchi construct semi dsupervis featur select spectral analysi traffic groom wdm sonet ring multipl line speed physic topolog design doptic network abstract advantag deploi allopt network guidelin properli design physic topolog network propos effici physic topolog design algorithm asymptot growth rate provis capac metric compar design altern higher asymptot growth rate translat directli higher deploy cost larg network studi show take fiber length consider lead lower capac requir find suffici larg fiber node ratio order minim asymptot growth provis capac increas capac util minim wavelength convers studi real network find fiber node ratio low result larg provis capac requir capac usabl increas ratio reduc provis capac achiev close util queri select techniqu effici crawl structur web sourc featur subset select bia classif learn featur select appli highdimension data prior classif learn train dataset select learn result socal featur subset select bia bia put exacerb data overfit neg affect classif perform current practic separ dataset seldom emploi select learn divid train data dataset featur select classifi learn reduc amount data task work attempt address dilemma formal select bia classif learn analyz statist properti studi factor affect select bia bia impact classif learn experi research endeavor provid illustr explan bia neg impact classif expect regress balanc ensembl approach weight classifi text classif featur subset select bia classif learn featur select appli highdimension data prior classif learn train dataset select learn result socal featur subset select bia bia put exacerb data overfit neg affect classif perform current practic separ dataset seldom emploi select learn divid train data dataset featur select classifi learn reduc amount data task work attempt address dilemma formal select bia classif learn analyz statist properti studi factor affect select bia bia impact classif learn experi research endeavor provid illustr explan bia neg impact classif expect regress power dlaw tradeoff optic electron switch current issu stream data mine discov map hierarch data multipl sourc inher structur block objection web content leverag multipl inform sourc world wide web humong archiv content inordin amount inform found web present challeng deliv inform user hand abund inform freeli access web denizen hand inform irrelev deleteri user control filter mechan desir prevent inappropri offens materi pornograph websit reach children wai access websit term access scenario access scenario includ search engin imag search textual content url redirect websit directli type porn websit url paper propos framework analyz websit aspect inform sourc gener classif model aim accur classifi content irrespect access scenario extens experi perform evalu result system illustr promis propos approach advanc knowledg discoveri data mine biolog browser base collabor resourc navig assist biomed research abstract realiz commun colleagu studi similar topic help identifi inform relev area studi found acceler acquisit collect knowledg defin area identifi specif sphere inquiri sphere correspond group peopl expert field paper provid systemat gain knowledg onlin search activ enabl organ share search find analysi built prototyp system biolog biomed research share implicit knowledg peer store access pattern central system reus biolog deploi lab tgen pilot studi data gather analyz preliminari text mine collabor filter method error dsensit grade model combin sensor sequenc model drive bia analysi text classif highli skew data learn label session approach model gener heterogen data classif ensembl method bag boost successfulli appli classif problem import issu ensembl approach gener model construct ensembl combin classif paper focu problem model gener heterogen data classif partit heterogen data number homogen partit gener reliabl accur classif model homogen partit examin wai form homogen subset propos method data point assign multipl time order gener homogen partit ensembl learn present detail algorithm empir studi uci benchmark dataset dataset imag classif show propos approach effect heterogen data classif traffic groom wdm sonet upsr ring multipl line speed research paper recommend system subspac cluster approach research lab spend consider amount time search publish articl relev current project similar interest conduct independ time consum search share result unabl leverag previou search result search process propos research paper recommend system avoid time consum search augment exist search engin recommend base previou search perform lab exist recommend system develop commerci domain million user research paper domain user compar larg number onlin research paper major challeng type data larg number dimens spars data contribut paper scalabl subspac cluster algorithm scuba thattackl problem synthet benchmark dataset evalu cluster algorithm demonstr perform tradit collabor filter approach recommend research paper cubesvd approach person web search competit web search market increas high demand person web search conduct retriev incorpor web user inform paper focus util clickthrough data improv web search million search conduct everydai search engin accumul larg volum clickthrough data record submit queri page click clickthrough data highli spars di erent type object user queri web page relationship object complic perform analysi data attempt discov web user interest pattern user locat inform paper approach cubesvd propos improv web search clickthrough data repres order tensor perform mode analysi higher order singular decomposit techniqu automat captur latent factor govern relat multi type object user queri web page tensor reconstruct base cubesvd analysi reflect observ interact object implicit associ web search activ carri base cubesvd analysi experiment evalu real world data set collect msn search engin show cubesvd achiev encourag search result comparison standard method evolv featur select describ evolutionari approach activ vision system dynam featur select summar recent work evolut simul activ retina complex shape discrimin describ detail experi extend approach terrain mobil robot equip mobil camera show evolv robot capabl select simpl visual featur activ maintain retin posit larg simplifi recognit task order gener effici navig trajectori extrem simpl neural control system analysi evolv solut robot develop simpl integr featur select algorithm classif cluster paper introduc concept algorithm featur select survei exist featur select algorithm classif cluster group compar algorithm categor framework base search strategi evalu criteria data mine task reveal unattempt combin guidelin select featur select algorithm categor framework continu effort build integr system intellig featur select unifi platform propos intermedi step illustr present show exist featur select algorithm integr meta algorithm advantag individu algorithm ad advantag user emploi suitabl algorithm know detail algorithm real world applic includ demonstr featur select data mine conclud work identifi trend challeng featur select research develop switch approach ensembl learn heterogen data redund base featur select microarrai data gene express microarrai data analysi select small number discrimin gene thousand gene import problem accur classif diseas phenotyp problem challeng due larg number featur gene small sampl size tradit gene select method select top rank gene individu discrimin power handl high degre redund gene latest research show remov redund gene select achiev represent characterist target phenotyp lead improv classif accuraci studi paper relationship featur relev redund propos effici method effect remov redund gene effici effect method comparison repres method demonstr empir studi public microarrai data set compact dual ensembl activ learn class dspecif ensembl activ learn real world task imag classif limit amount label data train automat classifi extens human expert involv requir instanc label detect egeria densa digit imageri real world classif task present addit challeng due subtl spectral egeria make difficult find singl accur classifi solut propos emploi ensembl classifi class class specif ensembl combin activ learn scheme class specif ensembl implicitli divers divers requir increas accuraci combin predict combin predict ensembl reduc uncertainti detect egeria iter activ learn suggest adapt ensembl imag unseen activ learner solut build compact ensembl present need expedit train activ learner combin result accur compact ensembl requir significantli expert involv imag region classif empir studi build compact ensembl abstract ensembl method achiev excel perform reli member classifi accuraci divers conduct empir studi relationship ensembl size ensembl accuraci divers experi benchmark data set show feasibl small ensembl maintain accuraci divers similar full ensembl propos heurist method effect select member classifi form compact ensembl idea compact ensembl motiv effect activ learn task classif unlabel data select sampl approach activ featur select featur select preprocess step machin learn effect reduc dimension remov irrelev data increas learn accuraci improv result comprehens tradit featur select method resort random sampl deal data set huge number instanc paper introduc concept activ featur select investig select sampl approach activ featur select filter model set present formal select sampl base data varianc appli wide featur select algorithm relief show realiz activ featur select reduc requir number train instanc achiev time save perform deterior design object evalu measur perform conduct extens experi synthet benchmark data set observ consist signific improv suggest work base studi experi effici featur select analysi relev redund featur select appli reduc number featur applic data hundr thousand featur exist featur select method focu find relev featur paper show featur relev insuffici effici featur select high dimension data defin featur redund propos perform explicit redund analysi featur select framework introduc decoupl relev analysi redund analysi develop correl base method relev redund analysi conduct empir studi effici effect compar repres method activ learn detect spectral variabl subject color infrar imageri subspac cluster high dimension data review mine hidden phrase definit web keyword search common form document search web web publish manual annot meta tag titl page frequent queri phrase order improv placement rank hidden phrase defin phrase occur meta tag web page bodi paper present algorithm mine definit hidden phrase web document phrase definit publish find relev phrase high queri frequenc ii search engin test content bodi document match phrase cooccurr cluster associ rule mine algorithm learn phrase definit high dimension data set provid experiment result situat dawar person inform retriev mobil internet featur select high ddimension data fast correl dbase filter solut featur select preprocess step machin learn effect reduc dimension remov irrelev data increas learn accuraci improv result comprehens recent increas dimension data pose sever challeng exist featur select method respect effici effect work introduc concept predomin correl propos fast filter method identifi relev featur redund relev featur pairwis correl analysi effici effect method demonstr extens comparison method real world data high dimension activ learn ensembl imag classif real world task imag classif limit amount label data train automat classifi extens human expert involv requir verif effici handl featur redund high ddimension data high dimension data pose sever challeng data mine featur select frequent techniqu preprocess high dimension data success data mine tradition featur select focus remov irrelev featur high dimension data remov redund featur equal critic paper provid studi featur redund high dimension data propos correl base approach featur select filter model extens empir studi real world data show propos approach effici effect remov redund irrelev featur activ featur select class abstract featur select frequent data pre process data mine train data set larg sampl commonli overcom difficulti work investig applic activ sampl featur select filter model set object partit data take advantag class inform achiev perform featur select fewer relev instanc random sampl version activ featur select emploi class inform propos empir evalu comparison random sampl conduct extens experi benchmark data set analyz reason class base activ featur select work result deal larg data set provid idea scale featur select algorithm consist dbase search featur select fast hierarch cluster valid cluster task group similar object cluster promin class algorithm hierarch agglom cluster hac iter agglomer closest pair data point belong cluster output dendrogram show level agglomer number object dataset hac method drawback high time memori complex cluster ineffici inaccur cluster valid paper show drawback allevi close studi dendrogram empir studi show hac algorithm follow trend number top level dendrogram lower level agglomer cluster small size close proxim cluster method propos exploit characterist reduc time memori complex significantli make valid effici accur analysi experi show effect propos method effici map rang classifi ternari dcam packet classif inher multi dimension search problem comput intens memori intens softwar implement hardwar base solut gigabit line rate process paper standard ternari content address memori tcam hardwar classif engin tradition approach deem ineffici rang broken prefix store tcam result larg expans propos scheme effici map rang tcam propos expans expans width constrain applic propos enabl high speed determinist classif low cost commod hardwar featur select cluster filter solut process applic larg number dimens challeng kdd commun featur select effect dimension reduct techniqu essenti pre process method remov noisi featur literatur method propos featur select cluster method wrapper techniqu requir cluster algorithm evalu candid featur subset wrapper approach larg unsuit real world applic due heavi relianc cluster algorithm requir paramet number cluster due lack suitabl cluster criteria evalu cluster subspac paper propos filter method independ cluster algorithm propos method base observ data cluster point point distanc histogram data cluster propos entropi measur low data distinct cluster high entropi measur suitabl select import subset featur invari number dimens affect qualiti cluster extens perform evalu synthet benchmark real dataset show effect featur select select sampl featur select preprocess step machin learn shown effect reduc dimension remov irrelev data increas learn accuraci improv comprehens paper problem activ featur select lter model set describ formal activ featur select call select sampl demonstr appli wide featur select algorithm relief show realiz activ featur select reduc requir number train data relief achiev time save perform deterior design object evalu measur conduct extens experi bench mark data set observ consist signi improv featur extract imag mine issu instanc select digit technolog comput advanc boom internet led massiv data collect corpor data data warehous web inform misinform explos szalai grai phenomenon drown data szalai grai report year detector cern particl collid switzerland record petabyt data research area scienc astronomi human genom face problem choke inform natur question gather data raw data rare direct manual analysi simpli pace fast growth data data mine knowledg discoveri kdd emerg field compris disciplin databas statist machin learn rescu kdd attempt turn raw data nugget creat special edg competit world scienc discoveri busi intellig kdd process defin fayyad al nontrivi process identifi valid potenti ultim understand pattern data data mine process includ data select preprocess data mine interpret evalu discret enabl techniqu rout tabl compact ternari cam sampl databas dtree user intent model web applic data mine problem infer user intent machin human interact kei research issu provid person experi servic paper propos approach model infer user action comput linguist featur keyword concept featur extract semant context intent model concept featur conceptu gener keyword associ rule mine find proper concept keyword modifi nave bay classifi intent model experiment result shown propos approach achiev averag accuraci predict user intent close precis human predict merg distanc densiti base cluster effici determin start sampl size progress sampl larg data set classif learn algorithm progress sampl ps increasingli larger random sampl learn model accuraci longer improv shown techniqu remark effici compar entir data set start sampl size ps open problem show improp start sampl size make ps expens comput due run learn algorithm larg number instanc sequenc random sampl achiev converg excess databas scan fetch sampl data suitabl start sampl size improv effici ps paper present statist approach effici find size call statist optim sampl size soss sens sampl size suffici resembl entir data introduc inform base measur resembl sampl qualiti defin soss show effici obtain scan data prove learn sampl soss produc model accuraci asymptot approach highest achiev accuraci entir data empir result number larg data set ucikdd repositori show soss suitabl start size progress sampl effici determin start sampl size progress sampl larg data set classif learn algorithm progress sampl ps increasingli larger random sampl learn model accuraci longer improv shown techniqu remark effici compar entir data set start sampl size ps open problem show improp start sampl size make ps expens comput due run learn algorithm larg number instanc sequenc random sampl achiev converg excess databas scan fetch sampl data suitabl start sampl size improv effici ps paper present statist approach effici find size call statist optim sampl size soss sens sampl size suffici resembl entir data introduc inform base measur resembl sampl qualiti defin soss show effici obtain scan data prove learn sampl soss produc model accuraci asymptot approach highest achiev accuraci entir data empir result number larg data set ucikdd repositori show soss suitabl start size progress sampl effici accur cluster effici accur cluster web document great interest web user kei compon search accuraci web search engin achiev paper introduc approach cluster web document call maxim frequent itemset mfi approach iter cluster algorithm mean expect maxim em sensit initi condit mfi approach firstli locat center point high densiti cluster precis center point initi point mean algorithm experiment result test web document set show mfi approach outperform method compar case case larg number categori web document set effici hierarch cluster algorithm partial overlap partit cluster import data explor task promin cluster algorithm agglom hierarch cluster roughli iter merg closest pair cluster propos back numer modif good featur natur simpl nonparametr group similar object capabl find cluster shape spheric arbitrari larg cpu time high memori requir limit larg data paper show geometr metric centroid median minimum varianc algorithm obei relationship roughli iter spent merg cluster distanc maximum merg distanc characterist exploit partial overlap partit shown experi analys type exist algorithm benefit excel drastic model classif perform larg data set learn algorithm learn accuraci increas size train data increas form learn curv learn curv fit interpol extrapol point model obtain learn curv predict maximum achiev learn accuraci estim amount data need achiev expect learn accuraci meaning data mine larg data set model propos model learn curv test applic larg data set paper focu issu empir compar potenti model fit learn curv typic classif algorithm decis tree log logist discrimin larg uci benchmark data set data learn fit full length learn curv small portion data fit part length learn curv model compar term perform fit full length learn curv fit part length learn curv predict learn accuraci full length experiment result show power law model perform algorithm data set result support applic learn curv data mine multidatabas mine identifi relev databas tool system knowledg discoveri data mine develop applic rule mine prior knowledg belief network approach exist data mine method classif tree neural network associ rule drawback user prior knowledg easili incorpor knowledg discoveri process rule mine databas lack quantit analys paper propos belief network method rule mine take advantag belief network direct acycl graph languag function numer represent probabilist depend variabl databas overcom drawback belief network provid natur represent captur causal relationship set variabl propos method mine gener correl rule captur relationship attribut variabl potenti applic propos method demonstr detail case studi benchmark databas keyword belief ne resourc descript framework metadata applic knowledg discoveri data mine rule mine learn belief network sampl databas dtree featur select cluster cluster import data mine task data mine concern larg high dimension data cluster algorithm literatur sensit larg high dimension featur affect cluster differ import cluster hinder cluster task effici handl select subset import featur help find cluster effici understand data reduc data size effici storag collect process task find origin import featur unsupervis data larg untouch tradit featur select algorithm work supervis data class inform unsupervis data class inform princip compon pc pc requir featur difficult understand approach featur except rule mine rel interesting measur paper present method mine except rule base measur estim interesting rel common sens rule refer rule mine interest rule import data mine task interest rule bring knowledg help decis maker advantag action true interesting rel issu depend prior knowledg estim bias due incomplet inaccur knowledg domain estim interesting trivial judg interesting huge set mine rule autom system requir exploit knowledg extractact data measur interesting extract knowledg data find measur unbias user belief unbias measur estim interestingn consist base featur select featur select effect techniqu deal dimension reduct classif task main compon data mine search optim subset featur search strategi consider complet heurist probabilist exist algorithm adopt measur evalu good featur subset work focus measur call consist studi properti comparison major measur wai measur search featur subset conduct empir studi examin pro con search method consist extens exercis aim provid comprehens view measur relat measur guidelin measur search strategi face applic introduct classif import data mine task basi custom retent data mine custom retent increasingli press issu todai evercompetit commerci arena relev import sale servic relat industri motiv real world problem face larg compani propos solut integr techniqu data mine featur select induct deviat analysi mine multipl concept level associ rule form intuit approach gaug custom loyalti predict likelihood defect action trigger earli warn result data mine kei eventu custom retent keyword custom retent featur select deviat analysi multipl level associ rule data mine introduct decad increas depend widespread databas busi scientif govern organ led explos growth data inst handl larg unsupervis data dimension reduct mine weak rule featur select consist measur studi support vector model independ select databas real world problem increas size situat select relev train data ecient storag process reason support vector machin svm reportedli exhibit desir properti select preserv train data support vector paper attempt quantifi extent svm train behav model independ select procedur common machin learn train databas compar predict result obtain dierent classier train data select svm select method ib random sampl interest observ made explan introduct base learn attract framework extract knowledg empir data goal gener input pattern realworld process formul classi cation task solv handl concept drift increment learn support vector machin increas size real world databas increas scale induct learn algorithm increment learn techniqu solut scalabl problem paper propos criteria evalu robust reliabl increment learn method studi robust increment train method support vector machin provid empir result benchmark machin learn dataset show support vector form succinct sucient set block block increment learn introduct develop classier learn method train data reduc risk learn process computation intract issu evid larg amount data research machin learn data mine scale classic induct learn algorithm handl effici search reliabl except find pattern data set fundament task data mine categor pattern strong weak random convent data mine techniqu design find strong pattern hold numer object consist expect expert strong pattern help predict unexpected contradict exhibit weak pattern repres small number object paper address problem find weak pattern reliabl except databas simpl effici approach propos deviat analysi identifi interest except explor reliabl flexibl handl subject object except demonstr effect propos approach set real life data set present interest find introduct data mine attract featur transform multivari decis tree induct univari decis tree udt inher problem replic repetit fragment multivari decis tree mdt propos overcom problem close examin convent wai build mdt reveal fragment problem persist approach suggest minim fragment problem separ hyperplan search decis tree build achiev featur transform initi featur vector featur vector featur transform obtain mdt build udt replac featur node combin initi featur elabor advantag approach detail expect perform experi conduct order confirm analysi result compar oc cart introducti monoton measur optim featur select featur select problem choos subset relev featur research search optim featur select method branch bound focu repres gener exhaust search bring optim subset condit exhaust search avoid sacrif subset optim condit exist monoton measur branch bound guarante optim subset error distanc base measur monoton measur emploi work monoton fast comput measur search relev featur guarante complet exhaust empir studi conduct show algorithm live claim discuss end introduct basic problem classif classifi pattern identifi relev databas multidatabas mine tool system knowledg discoveri data mine develop applic immers heap databas question face practition start mine paper break convent data mine assumpt databas join argu step multidatabas mine identifi databas relev applic mine process lengthi aimless ineffect relev measur propos identifi relev databas mine task object find pattern regular attribut effici implement identifi relev databas experi conduct valid measur perform show promis applic keyword multipl databas data mine queri relev measur introduct re cfmd conflict dfree multivari discret algorithm hybrid search featur subset featur select search problem optim subset featur class separ basic featur select criteria maxim class separ literatur work adopt criterion aim maintain discrimin power data examin pro con exist algorithm featur select propos hybrid algorithm probabilist complet search advantag algorithm begin run lvf probabilist search reduc number featur run automat branch amp bound abb complet search impos limit amount time algorithm run obtain approxim algorithm empir studi suggest divid time equal phase yield perform hybrid search algorithm substanti outperform earlier method gener introducti data mine applic custom retent port singapor author psa increment featur select featur select problem find relev featur number featur dataset larg number pattern huge effect method featur select dimension reduct increment probabilist algorithm design implement altern exhaust heurist approach theoret analysi support idea probabilist algorithm find optim optim subset featur experiment result suggest probabilist algorithm effect obtain optim suboptim featur subset increment version expedit featur select number pattern larg scale sacrif qualiti select featur keyword pattern recognit machin learn featur select dimension reduct appli intellig kluwer academ publish introduct featur select analysi hidden represent greedi cluster guest editor introduct featur transform subset select neurolinear system extract obliqu decis rule neural network present neurolinear system extract obliqu decis rule neural network train classif pattern condit obliqu decis rule correspond partit attribut space hyperplan necessarili axi parallel allow set hyperplan form boundari decis region lead signific reduct number rule gener maintain accuraci rate network describ compon neurolinear detail exampl artifici dataset experiment result real world dataset show system effect extract compact comprehens rule high predict accuraci neural network keyword rule extract obliqu rule prune discret introduct neural network wide appli solv classif problem comparison neural network decis tre featur select discret discret turn numer attribut discret featur select elimin irrelev redund attribut chi simpl gener algorithm statist discret numer attribut repeatedli inconsist found data achiev featur select discret handl mix attribut work multiclass data remov irrelev redund attribut keyword discret featur select pattern classif introduct featur select elimin irrelev redund attribut relev featur classif algorithm gener improv predict accuraci shorten learn period form simpler concept abund featur select algorithm method principl compon compos smaller number featur select subset origin attribut paper consi featur select classif featur select focu interest time work creation huge databas consequ requir good machin learn techniqu problem aris approach featur select demand survei comprehens overview exist method present identifi step typic featur select method categor exist method term gener procedur evalu function reveal hitherto unattempt combin gener procedur evalu function repres method chosen categori detail explan discuss benchmark dataset characterist compar studi strength weak method explain guidelin appli featur select method base data type domain characteri neurolinear neural network obliqu decis rule present neurolinear system extract obliqu decis rule neural network train classif pattern condit obliqu decis rule correspond partit attribut space hyperplan necessarili axi parallel allow set hyperplan form boundari decis region lead signific reduct number rule gener maintain accuraci rate network describ compon neurolinear detail exampl artifici dataset experiment result real world dataset show system effect extract compact comprehens rule high predict accuraci neural network keyword rule extract obliqu rule prune discret introduct neural network wide appli solv classif problem comparison neural network decis tre probabilist approach featur select filter solut featur select defin problem find minimum set relev attribut describ dataset origin attribut examin problem exhaust heurist approach featur select paper propos probabilist approach theoret analysi experiment studi show propos approach simpl implement guarante find optim resourc permit fast obtain result effect select featur improv perform learn algorithm site applic involv huge dataset conduct independ prove effect scalabl propos algorithm discuss aspect applic featur select algorithm introduct problem featur select defin find relev attribut origin attrib featur select classif probabilist wrapper approach featur select defin problem find minimum set featur induct algorithm achiev highest predict accuraci data origin featur probabilist wrapper model propos method exhaust search heurist approach aim model avoid local minima exhaust search highest predict accuraci criterion search smallest analysi experi show model effect find relev featur remov irrelev context improv predict accuraci induct algorithm simpl straightforward provid fast solut search optim applic model futur work relat issu discuss introduct problem featur select defin find relev featur origin attribut improv backpropag learn featur select exist redund irrelev noisi data proper data train network speed train simplifi learn structur improv perform phase train algorithm propos phase number input unit network determin inform base method attribut meet criteria inclus consid input network phase number hidden unit network select automat base perform network train data hidden unit ad time experiment result show algorithm achiev faster learn time simpler network improv perform symbol represent neural network backpropag neural network gener predict decis tree pattern classif problem regard black box predict explain decis tree applic explicit knowledg need human expert work deriv symbol represent neural network make epxlicit predict network algorithm propos implement extract symbol rule neural network effect data mine neural network classif data mine problem receiv great attent recent databas commun paper present approach discov symbol classif rule neural network neural network thought suit data mine classif made explicitli state symbol rule suitabl verif interpret human propos approach concis symbol rule high accuraci extract neural network network train achiev requir accuraci rate redund connect network remov network prune algorithm activ valu hidden unit network analyz classif rule gener result analysi effect propos approach demonstr experiment result set standard data mine test problem introduct dimension reduct discret exist numer data larg amount record databas pose challeng task explicit concept extract raw data paper introduc method reduc data vertic horizont discrimin power origin data pave extract concept method base discret vertic reduct featur select horizont reduct experiment result show data effect reduc propos method predict accuraci classifi improv data dimension reduct classif rule learn simpler kei word dimension reduct discret knowledg discoveri introduct wide comput bring prolifer databas aid comput flood raw data exploit human knowledg discoveri system databas understand neural network rule extract backpropag neural network gener predict decis tree pattern classif problem regard black box predict interpret decis tree paper argu proper techniqu enabl algorithm extract rule draw parallel decis tree show predict network explain rule extract network understood experi demonstr rule extract neural network compar decis tree term predict accuraci number rule averag number condit rule preserv high predict accuraci origin network effect classfic rule extract abstract neurorul connectionist approach data mine classif involv find rule partit data set disjoint group class data mine problem approach propos mine classif rule larg databas decis tree base symbol learn method connectionist approach base neural network thought suit data mine major reason cite knowledg gener neural network explicitli repres form rule suitabl verif interpret human paper examin issu newli develop algorithm rule similar concis gener symbol method extract neural network data mine process neural network emphasi rule extract experiment result comparison previous publish work present critic knowledg dbase design system 