learn gene regulatori network posit unlabel data dyadic predict latent featur log dlinear model technic perspect creativ help influenc predict precis predict label dyadic data account bursti topic model topic model successfulli varieti applic state art topic model suffer import flaw captur tendenc word burst fundament properti languag word document introduc topic model dirichlet compound multinomi dcm distribut model bursti phenomenon text text dataset model achiev held likelihood standard latent dirichlet alloc lda straightforward incorpor dcm extens topic model complex lda transport classif databas recent advanc advanc learn find relev biolog articl neg train exampl abstract classifi tradition learn set posit neg train exampl classifi requir train incomplet set posit exampl set unlabel exampl situat transport classif databas tcdb www tcdb org repositori inform protein involv transmembran transport paper present evalu method learn rank relev tcdb newli publish scientif articl articl referenc tcdb posit train exampl method succeed identifi articl relev tcdb fewer month major practic success gener data mine perspect contribut paper devis evalu approach solv posit problem effect ii appli support vector machin state ofth art recogn rank relev iii deploi system updat wide real world biomed databas supplementari inform includ data set publicli www cs ucsd user knoto pub ajcai learn dstage svm fcrf sequenc classifi learn sequenc classifi mean learn predict sequenc output tag base set input data item recogn handwritten word cat base imag handwritten letter gener knowledg english letter combin sequenc classif task paper describ stage approach learn sequenc classifi highli accur ii scalabl iii easi data mine applic stage approach combin support vector machin svm condit random field crf highli accur benefit maximummargin natur svm abil crf model correl neighbor output tag ii scalabl input svm small train set input crf small number featur svm output iii easi combin exist publish softwar straightforward detail experi task recogn handwritten word show stage approach accur faster scalabl lead method learn sequenc classifi includ max margin markov network ns standard crf learn classifi posit unlabel data input algorithm learn binari classifi consist set exampl set consist posit exampl concept learn set consist neg exampl case train data incomplet set posit exampl set unlabel exampl posit neg problem solv paper learn standard binari classifi nontradit train set natur assumpt label exampl select randomli posit exampl show classifi train posit unlabel exampl predict probabl differ constant factor true condit probabl posit show result wai learn classifi nontradit train set appli method solv real world problem identifi protein record includ incomplet special molecular biologi databas experi domain show model train method perform current state art bias svm method learn posit unlabel exampl make gener classifi robust select bia find transport protein gener protein databas abstract number special databas molecular biologi grow fast avail molecular data trend necessit develop automat method find relev inform includ special databas show comprehens databas swissprot sourc entri special databas tcdb transport classif databas carefulli construct keyword base queri perform poorli determin swissprot record relev tcdb show machin learn approach perform describ maximum entropi classifi train swissprot record achiev high precis recal cross valid experi classifi deploi part pipelin updat tcdb human expert examin swissprot record potenti inclus tcdb method describ flexibl gener appli easili special databas kdd cup workshop octob netflix releas dataset million anonym movi rate challeng data mine machin learn comput scienc commun develop system beat accuraci recommend system cinematch briefli describ challeng review relat work effort summar visibl progress date potenti data outlin includ applic kdd cup kdd cup workshop octob netflix releas dataset million anonym movi rate challeng data mine machin learn comput scienc commun develop system beat accuraci recommend system cinematch briefli describ challeng review relat work effort summar visibl progress date potenti data outlin includ applic kdd cup cluster document exponenti dfamili approxim dirichlet compound multinomi distribut dirichlet compound multinomi dcm distribut call multivari polya distribut model text document take account bursti fact word occur document occur repeatedli deriv famili distribut approxim dcm distribut constitut exponenti famili unlik dcm distribut call edcm distribut obtain insight properti dcm distribut deriv algorithm edcm maximum likelihood train time faster method dcm distribut investig expectationmaxim edcm compon determinist anneal cluster algorithm document experi show algorithm competit method literatur superior point view find model low perplex model word bursti dirichlet distribut multinomi distribut model text document captur phenomenon word document tend burst word appear paper propos dirichlet compound multinomi model dcm altern multinomi dcm model addit degre freedom captur bursti show experiment dcm substanti multinomi model text data measur perplex show standard document collect dcm lead classif multinomi model dcm perform compar obtain multipl heurist multinomi model deriv tf didf fisher kernel abstract dirichlet compound multinomi dcm distribut recent shown good model document captur phenomenon word bursti unlik standard model multinomi distribut paper investig dcm fisher kernel function compar document deriv dcm show dcm fisher kernel compon similar term frequenc tf invers document frequenc idf factor standard tf idf method repres document experi show dcm fisher kernel perform altern kernel nearest neighbor document classif tf idf represent perform fast recognit music genr rbf network bayesian network framework reject infer sourc success boost wrapper induct paper examin import recent rule base inform extract techniqu name boost wrapper induct bwi conduct experi wider varieti task previous studi includ task collect natur text document investig systemat algorithm compon bwi boost contribut success show benefit boost aris abil reweight exampl learn specif rule result high precis combin abil continu learn rule posit exampl cover result high recal quantit indic regular extract task propos measur call swi ratio show measur good predictor success tool analyz task base result analyz strength limit bwi specif explain limit inform made represent investig consequ fact confid valu return extract true probabl investig benefit includ grammat semant inform natur text document pars tree attribut inform xml learn rule improv machin translat system abstract paper show learn rule improv perform machin translat system system consist translat function languag languag train text translat back translat differ knowledg translat function identifi rule learn improv function context independ rule learn inform suggest singl translat word multipl altern translat word likelihood ratio test identifi word occur case significantli word context context depend rule appli pan american health organ corpu sentenc learn rule improv understand translat produc sdl intern engin sentenc high precis triangl inequ acceler dmean mean algorithm wide method discov cluster data show acceler dramat comput result standard algorithm acceler algorithm avoid unnecessari distanc calcul appli triangl inequ wai keep track lower upper bound distanc point center experi show algorithm effect dataset dimens effect number cluster increas gt time faster previous acceler mean method principl method advis reinforc learn agent import issu reinforc learn incorpor expert knowledg principl manner scale real world task paper present method incorpor arbitrari advic reward structur reinforc learn agent alter optim polici method extend potentialbas shape method propos ng al learn dmean cluster dataset number cluster obviou choos automat hard algorithm problem altern dmean algorithm find cluster investig behavior standard mean cluster algorithm altern harmon mean algorithm due zhang colleagu fuzzi mean gaussian expect maxim variant harmon mean aim aspect algorithm contribut nding good cluster oppos converg low qualiti local optimum describ algorithm uni ed framework introduc separ cluster membership data weight function transform classifi score accur multiclass probabl estim class membership probabl estim import applic data mine classif output combin sourc inform decis make depend misclassif cost output classifi domain knowledg previou calibr method appli class problem show obtain accur probabl estim multiclass problem combin calibr binari probabl estim propos method obtain calibr class probabl estim appli classifi produc rank exampl naiv bay support vector machin classifi give experiment result varieti class multiclass domain includ direct market text categor digit recognit share challeng data mine comput biologi abstract invit talk bayesian approach failur predict disk drive hard disk drive failur rare costli abil predict failur import consum drive manufactur comput system manufactur alik paper investig abil bayesian method predict disk drive failur base measur drive intern condit view problem anomali detect stanc introduc mixtur model naiv bay submodel cluster train expect maxim method naiv bay classifi supervis learn approach method test realworld data drive predict accuraci algorithm higher accuraci threshold method disk drive industri todai obtain calibr probabl estim decis tree naiv bayesian classifi accur calibr estim class membership probabl need supervis learn applic cost sensit decis made exampl depend cost paper present simpl success method obtain calibr probabl estim decis tree naiv bayesian classifi larg challeng kdd contest dataset testb report result detail experiment comparison ten method evalu measur conclud bin succe significantli improv naiv bayesian probabl estim improv decis tree probabl estim recommend smooth estim variant prune call curtail foundat cost dsensit learn paper revisit problem optim learn decis make misclassif error incur penalti character precis intuit cost matrix reason show avoid mistak defin cost matrix econom incoher class case prove theorem show chang proport neg exampl train set order make optim cost sensit classif decis classifi learn standard costsensit learn method argu chang balanc neg posit train exampl effect classifi produc standard bayesian decis tree learn method recommend appli method domain differ misclassif cost learn classifi train set comput optim decis learn make decis cost probabl unknown machin learn domain misclass cost dierent dierent exampl class membership probabl exampledepend domain cost probabl unknown test exampl cost estim probabl estim learn paper rst discuss make optim decis cost probabl estim present decis tree learn method obtain calibr probabl estim paper explain obtain unbias estim depend cost take account diculti gener probabl cost independ random variabl train exampl cost repres exampl problem call sampl select bia econometr solut base nobel prize win work due economist jame heckman show method propos paradox fuzzi logic revisit paper repli articl entitl elkan theoret argument reconsid prof enric trilla prof claudi alsina express dr piero bonisson invit write paper show articl trilla alsina advanc public mathemat studi fuzzi began paper entitl fuzzi set lotfi zadeh zad import distinguish fuzzi logic fuzzi set theori theorem paper paradox success fuzzi logic publish elk revis elk result standard fuzzi logic paper trilla alsina concern fuzzi set erron state theorem fals start section trilla alsina write equival formul fuzzi logic professor elkan claim refer result kdd classifi learn kdd knowledg discoveri contest data mine techniqu routin fundrais select prospect larg pool candid make financi contribut techniqu reli statist model base trial perform data trial perform data typic obtain solicit smaller sampl prospect pool collect trial data involv cost fundrais interest keep trial size small collect data build reliabl statist model evalu remaind prospect describ experiment design approach optim choos trial prospect exist larg pool prospect prospect cluster render problem practic tractabl modifi standard optim algorithm prevent repeat select prospect cluster prospect solicit assess benefit approach kdd data set compar perform model base optim trial data set model base randomli select trial data set equal size scalabl cluster algorithm revisit paper present simpl algorithm perform mean cluster scan dataset buffer point dataset fix size experi show method time faster standard mean produc cluster equal equal qualiti method simplif algorithm due bradlei fayyad reina data compress techniqu attempt improv speed cluster qualiti overhead techniqu make origin algorithm time slower standard mean materi dataset standard mean scan dataset multipl time lesion studi show compress techniqu improv cluster qualiti result hold megabyt synthet dataset dataset creat real world data kdd data mine contest algorithm implement experi design result generalis dataset gigabyt larger meme mast meta dmeme tool motif discoveri protein sequenc effici domain dindepend algorithm detect approxim duplic databas record detect databas record approxim duplic exact duplic import task databas duplic record realworld entiti data entri error unstandard abbrevi differ detail schema record multipl databas reason paper present effici algorithm recogn cluster approxim duplic record kei idea distinguish algorithm present version smith waterman algorithm comput minimum edit distanc domainindepend method recogn pair approxim duplic record union find algorithm track cluster duplic record increment pairwis duplic relationship discov algorithm prioriti queue cluster subset respond adapt size homogen cluster discov meta dmeme motif dbase hidden markov model protein famili motiv model famili relat biolog sequenc hidden markov model hmm increasingli widespread face major problem complex mathemat model requir larg train set order accur character famili famili sequenc standard linear hmm paramet train adequ result work attempt solv problem gener smaller hmm precis model conserv region famili hmm construct motif model gener em algorithm meme softwar motif base hmm paramet train smaller data set studi shortchain alcohol dehydrogenas fe ferredoxin support claim motif base hmm exhibit increas sensit select databas search train set lpmeme statist method induct logic program reason unknown static program analysi techniqu valu user input network state analysi time unknown valu treat determinist choic made program execut environ glean inform static unknown valu influenc comput give method integr determinist choic express static analysi interestingli solv result recurs constraint directli give exact method answer queri show experiment result solv form concis practic enabl appli techniqu larg program includ entir oper system field match problem algorithm applic combin inform heterogen sourc equival data multipl sourc identifi task field match problem specif task determin syntact valu altern design semant entiti address dept comput sci eng univers california san diego gilman dr dept la jolla ca ucsd comput scienc engin depart ca design depart paper describ field match algorithm evalu perform real world dataset propos method smith waterman algorithm compar dna protein sequenc applic field match knowledg discoveri briefli includ webfind softwar tool discov scientif paper publish worldwid web webfind extern inform sourc guid search author paper worldwid web tool webfind solv field match problemn order navig inform sourc exploratori analysi speedup learn data epect maxim paramem parallel implement web interfac dna protein motif discoveri tool advanc softwar tool fail reach wide audienc requir special hardwar instal expertis abund cpu cycl worldwid web offer opportun distribut system program meme discov repeat pattern call motif set dna protein sequenc tool biologist worldwid web asynchron singl program multipl data version program call paramem run intel paragon xp parallel comput san diego supercomput center paramem scale gracefulli node paragon effici larg data set worldwid web interfac paramem accept set sequenc interact user submit sequenc paragon analysi mail result back user paramem free public http www sdsc meme keyword intellig system supercomput prior knowledg discov motif meme meme tool discov motif set protein dna sequenc paper describ extens meme increas abil find motif total unsupervisedfashion benefit prior knowledg background knowledg assert meme obtain increas robust method determin motif width automat probabilist model motif absent input sequenc hand meme exploit prior knowledgeabout motif present input sequenc length motif palindrom dirichlet mixtur expect pattern individu motif posit extens experi report support claim meme benefit requir background knowledg experi previous studi dna protein sequenc famili protein famili document prosit unsupervis learn multipl motif biopolym expect maxim meme algorithm extend expect maxim em algorithm identifi motif unalignedbiopolym sequenc aim meme discov motif set biopolym sequenc advanc motif present meme innov expand rang problem solv em increas chanc find good solut subsequ occur biopolym sequenc start point em algorithm increas probabl find global optim motif assumpt sequenc occurr share motif remov multipl appear motif occur sequenc permit algorithm ignor sequenc appear share motif increas resist noisi data method probabilist eras share motif found incorpor tha fit mixtur model expect maxim discov motif biopolym abstract algorithm paper discov motif collect dna protein sequenc techniqu expect maxim fit compon finit mixtur model set sequenc multipl motif found fit compon finit mixtur model data probabilist eras occurr motif found repeat process find success motif algorithm requir set sequenc number width motif input return model motif threshold bay optim classifi search occurr motif databas algorithm estim time motif occur input dataset output align occurr motif algorithm capabl discov motif differ number occurr singl dataset motif discov treat set sequenc creat stochast process model mixtur densiti gener occurr motif rest posit sequenc expect maxim estim paramet densiti mix paradox success fuzzi logic applic fuzzi logic heurist control highli success aspect fuzzi logic essenti practic us paper show appar reason version fuzzi logic collaps mathemat valu logic publish report expert system real world reason uncertainti fuzzi logic appear limit fuzzi logic detriment control applic current fuzzi control simpler knowledg base system futur technic limit fuzzi logic expect import practic work fuzzi control encount problem scale knowledg base system elkan repli paradox controversi fuzzi logic paradox success fuzzi logic applic fuzzi logic heurist control highli success aspect fuzzi logic essenti practic us paper show appar reason version fuzzi logic collaps mathemat valu logic publish report expert system real world reason uncertainti fuzzi logic appear limit fuzzi logic detriment control applic current fuzzi control simpler knowledg base system futur technic limit fuzzi logic expect import practic work fuzzi control encount problem scale knowledg base system lenat guha lenat guha critic experiment evalu ebl ration reconstruct nonmonoton truth mainten system properti lui moniz pereira anil nerod editor logic program monoton reason proceed intern workshop page lisbon mit press juli min jack minker indefinit databas close world assumpt proceed confer autom deduct york page springer min jack minker foundat deduct databas morgan kaufmann street lo alto ca st edit min jack minker perspect deduct databas journal logic program min jack minker foundat disjunct logic program ew lusk ross overbeek editor proceed north american confer logic program page mit press min jack minker overview nonmonoton reason logic program cs tr comput scienc dept univ maryland univers marylan increment approxim plan independ logic databas queri updat ration reconstruct nonmonoton truth mainten system properti lui moniz pereira anil nerod editor logic program monoton reason proceed intern workshop page lisbon mit press juli min jack minker indefinit databas close world assumpt proceed confer autom deduct york page springer min jack minker foundat deduct databas morgan kaufmann street lo alto ca st edit min jack minker perspect deduct databas journal logic program min jack minker foundat disjunct logic program ew lusk ross overbeek editor proceed north american confer logic program page mit press min jack minker overview nonmonoton reason logic program cs tr comput scienc dept univ maryland univers marylan conspiraci number cach search tree theorem dprove paper made import contribut demonstr practic feasibl gener abduct approach explan successfulli build domain independ system accel gener perform plan recognit diagnosi effici practic util support claim extens evalu system narr text plan recognit domain real patient case set cover diagnosi domain model base diagnosi scenario adder temperatur control water balanc system human kidnei adder circuit knowledg base hundr horn claus rule develop evalu criterion explanatori coher evalu qualiti explan plan recognit domain present empir result indic coher metric outperform simplic metric select explan plan recognit domain coher base approach perform probabilist approach plan recognit engin numer prior posterior probabl present empir evid show cach previous comput explan critic effici abduct algorithm specif speedup order magnitud obtain test problem summari paper demonstr implement system gener effici abduct task plan recognit diagnosi futur hold promis gener abduct approach explan acknowledg conspiraci number cach search tree theorem dprove paper made import contribut demonstr practic feasibl gener abduct approach explan successfulli build domain independ system accel gener perform plan recognit diagnosi effici practic util support claim extens evalu system narr text plan recognit domain real patient case set cover diagnosi domain model base diagnosi scenario adder temperatur control water balanc system human kidnei adder circuit knowledg base hundr horn claus rule develop evalu criterion explanatori coher evalu qualiti explan plan recognit domain present empir result indic coher metric outperform simplic metric select explan plan recognit domain coher base approach perform probabilist approach plan recognit engin numer prior posterior probabl present empir evid show cach previous comput explan critic effici abduct algorithm specif speedup order magnitud obtain test problem summari paper demonstr implement system gener effici abduct task plan recognit diagnosi futur hold promis gener abduct approach explan acknowledg autom induct reason logic program decis procedur conjunct queri disjoint autom induct reason logic program 