dneg matrix factor cluster multipl manifold predict correct problem solv tempor collabor filter approach learn identifi student relev irrelev question micro dblog support classroom discrimin model integr document evid document dcandid associ expert search joint probabilist classif model resourc select probabilist model answer drank multilingu question danswer combin evid probabilist framework answer rank answer merg question answer discrimin graphic model faculti homepag discoveri effici privaci dpreserv similar document detect learn identifi student dtask behavior intellig tutor system learn past queri resourc select dplausibl semant preserv text sanit text document plai signific role decis make scientif research feder regul document patholog record person identifi inform share freeli properli sanit gener speak document sanit consist find hide person identifi inform task receiv attent research commun main strategi task simpli remov person identifi hard import inform diagnos person medic histori complet remov patholog record record longer readabl wors longer suffici inform research purpos observ sensit inform tuberculosi replac sensit term infecti diseas simpli remov sensit term term hidden gener semant relat term protect sensit inform unnecessarili degrad amount inform contain document base observ main contribut paper provid inform theoret approach text sanit develop effici heurist sanit text document analyz attack prevent propos model predict correct problem solv low dlevel log data intellig tutor system automat text categor mathemat word problem multipl instanc transfer learn system perform diagnosi machin learn transfer learn scalabl distribut system continu grow scale complex result increasingli involv interact compon increasingli intric failur mode hard diagnos manual increas vulner larger system increas difficulti failur diagnosi motiv machin learn approach autom diagnosi task preliminari encourag result achiev scale exist approach larg applic remain challeng increas scale current approach suffer curs dimension exacerb explod set system state measur metric paper significantli improv scalabl perform diagnosi method contribut lie intellig partit metric space coupl cooper tempor segment algorithm divid system observ time space remov multipl explos system state ii transfer learn techniqu improv accuraci leverag depend partit valid approach month product trace custom face geograph distribut tier internet servic result show signific accuraci improv averag naiv partit state space tempor segment algorithm transfer learn order magnitud reduct comput cost brute forc approach learn partit loss accuraci ic maximum margin multipl instanc cluster cluster classif regress major research topic machin learn work conduct solv multipl instanc classif multipl instanc regress problem supervis train pattern bag bag consist instanc research unsupervis multipl instanc cluster limit paper formul maximum margin multipl instanc cluster ic problem multipl instanc cluster task avoid solv nonconvex optim problem directli ic relax enabl effici optim solut combin constrain concav convex procedur cccp cut plane method paper analyz import properti propos method relationship propos method relat extens set empir result demonstr advantag propos method exist research effect effici model search respons time similar document detect limit inform disclosur discrimin probabilist model passag base retriev approach passag level evid document retriev shown mix result appli varieti test bed characterist main reason inconsist perform exist unifi framework model evid individu passag document paper propos probabilist model formal model evid set top rank passag document probabilist model retriev criterion document relev passag document relev model passag independ probabilist model step incorpor similar correl passag model train discrimin manner present combin approach combin rank list document retriev passag base retriev extens set experi conduct trec test bed show effect propos discrimin probabilist model passagebas retriev propos algorithm compar state art document retriev algorithm languag model approach passag base retriev combin approach shown provid result document retriev passagebas retriev approach bayesian framework knowledg driven regress model micro darrai data analysi bayesian framework knowledg driven regress model micro darrai data analysi effect effici result merg strategi multilingu inform retriev feder search environ combin gene sequenc similar textual inform gene function annot literatur probabilist framework answer select question answer paper describ probabilist answer select framework question answer contrast previou work individu resourc ontolog web valid answer candid work focus develop unifi framework multipl resourc valid answer candid consid evid similar answer candid order boost rank correct answer framework select answer candid gener answer extract method extens set empir result base trec factoid question demonstr effect unifi framework probabilist graphic model joint answer rank question answer graphic model appli inform retriev natur languag process task recent literatur paper appli probabilist graphic model answer rank question answer model estim joint probabl correct answer candid probabl correct individu candid infer joint predict model estim correct individu answer correl enabl list accur comprehens answer model compar logist regress model directli estim probabl correct individu answer candid extens set empir result base trec question demonstr effect joint model answer rank combin joint model logist regress model improv effici accuraci answer rank protect sourc privaci feder search feder search text search engin uncoop environ combin multipl resourc knowledg transfer opinion detect trec blog track paper describ opinion detect system develop carnegi mellon univers trec blog track system perform stage process passag retriev opinion detect due lack train data trec blog corpu onlin opinion review provid domain movi review product review train data knowledg transfer perform make cross domain learn logist regress rank sentenc level opinion object statement evalu show algorithm effect task studi mixtur model collabor filter fedlemur project feder search real world collabor imag retriev regular metric learn content base imag retriev cbir relev imag identifi base similar queri imag cbir algorithm hinder semant gap low level imag featur comput imag similar high level semant concept convei imag reduc semant gap util log data user feedback collect cbir system histori call collabor imag retriev paper present metric learn approach name regular metric learn collabor imag retriev learn distanc metric explor correl low level imag featur log data user relev judgment compar previou research regular mechan algorithm effect prevent overfit formul propos learn algorithm semi definit program problem solv effici exist softwar packag scalabl size log data extens set experi conduct show algorithm substanti improv retriev accuraci baselin cbir system euclidean distanc metric modest amount log data experi algorithm effect effici clef multilingu retriev combin multipl multilingu rank list abstract particip task multi year retriev multi result merg multi year retriev work simpl multilingu rank list built merg rank list languag gener singl type retriev algorithm algorithm propos combin simpl multilingu rank list singl rank list empir studi show approach combin multilingu retriev result substanti improv accuraci singl multilingu rank list multi result merg task primari interest task view similar result merg task feder search queri specif languag specif model propos calcul compar document score small amount document estim logist model inform document logist model estim compar score document document sort final rank list set experi demonstr advantag queri specif languag specif model altern learn weight term inform retriev categori inform adjust mixtur weight gaussian mixtur model regular probabilist latent semant analysi abstract mixtur model gaussian mixtur model wide applic model data gaussian mixtur model gmm assum data point gener set gaussian model set mixtur weight natur extens gmm probabilist latent semant analysi plsa model assign mixtur weight data point plsa flexibl gmm method tradeoff plsa suffer overfit problem paper propos regular probabilist latent semant analysi model rplsa properli adjust amount model flexibl train data fit model robust avoid overfit problem conduct empir studi applic speaker identif show effect model experi result nist speaker recognit dataset rplsa model outperform gmm plsa model substanti principl rplsa appropri adjust model flexibl natur extend applic type mixtur model model search engin effect feder search feder search link multipl search engin singl virtual search system prior research feder search focus select search engin relev content retriev effect individu search engin omiss problem feder search engin qualiti threshold strategi text classifi trec biomed triag task experi york univers trec genom track genom experi focu address major problem biomed inform retriev problem deal synonym deal frequent acronym deal homonym propos queri expans algorithm construct structur queri experi averag precis map automat run york ga algorithm manual run york gm algorithm evalu result show algorithm ectiv improv retriev perform find techniqu pseudo relev feedback extend stop word set make posit contribut retriev perform effect automat imag annot coher languag model activ learn imag annot user access larg imag databas textual queri studi automat imag annot util machin learn techniqu automat learn statist model annot imag appli gener annot unseen imag common problem share previou learn approach automat imag annot annot word predic imag independ annot word paper propos coher languag model automat imag annot take account word toword correl estim coher languag model imag approach import advantag automat determin annot length improv accuraci retriev result activ learn significantli reduc requir number annot imag exampl empir studi corel dataset present show effect coher languag model automat imag annot categori subject descriptor unifi filter combin collabor filter content dbase filter mixtur model exponenti model unifi util maxim framework resourc select paper present unifi util framework resourc select distribut text inform retriev framework show effici effect infer probabl relev document text databas estim relev inform resourc select made explicitli optim goal applic specif databas recommend select optim goal highrecal includ relev document select databas distribut document retriev select target high precis goal high precis final merg list document model solid framework distribut inform retriev empir studi show effect state art algorithm automat weight scheme collabor filter collabor filter identifi inform interest user base inform provid similar user memori base approach collabor filter pearson correl coeffici approach identifi similar user compar rate set item approach item weight equal predefin function impact rate discrep user consider item highli favor user smaller impact user similar item type user tend give rate simpl weight method varianc weight address problem empir studi shown ineffect improv perform collabor filter paper present optim algorithm automat comput weight item base rate train user specif weight scheme creat cluster distribut user vector item space bring user similar interest closer separ user interest distant empir studi dataset shown weight scheme substanti improv perform pearson correl coeffici method collabor filter effect vari number document blind feedback analysi nrrc ria workshop bf fnumdoc experi suit introduct bf numdoc experi set experi conduct reliabl inform access ria workshop host northeast region research center nrrc mitr involv research group goal ria workshop focu expans techniqu relev feedback pseudo relev feedback end mind bf numdoc suit experi design examin effect vari number document extract expans term pseudo relev feedback describ motiv set experi hypothes test overview experiment method analysi result conclus natur workshop examin system depend effect topic depend effect hypothes hypothes examin bf numdoc experi system demonstr tradeoff studi method normal user rate collabor filter goal collabor filter make recommend test user util rate inform user share interest similar test user rate determin user interest rate habit user import normal rate user scale paper compar normal strategi user rate gaussian normal method decoupl normal method incorpor rate normal method collabor filter algorithm evalu effect eachmovi dataset experi result shown decoupl method rate normal effect gaussian normal method improv perform collabor filter algorithm bayesian approach activ learn collabor filter collabor filter techniqu exploit prefer pattern group user predict util item activ user gener perform collabor filter depend number rate exampl activ user number rate exampl activ user accur predict rate activ learn effect acquir inform rate exampl activ user previou work activ learn collabor filter consid expect loss function base estim model mislead estim model inaccur paper take step take account posterior distribut estim model result robust activ learn algorithm empir studi dataset movi rate show number rate activ user restrict small activ learn method base estim model don perform activ learn method model distribut achiev substanti perform collabor filter decoupl model prefer rate paper describ model collabor filter motiv work fact user similar prefer item rate scheme user tend assign higher rate item user unlik previou model collabor filter determin similar user base rate perform model treat user prefer item separ user rate scheme specif user build separ model prefer model captur item favor user rate model captur user rate item prefer inform similar user comput base underli prefer model surfac rate compar model repres previou approach data set experi result show model outperform previou approach test consist data set effect databas size distribut resourc select algorithm resourc select import topic distribut inform retriev research compon distribut inform retriev task serv independ applic databas recommend system resourc represent part larg bodi valuabl prior research resourc select studi effect databas size distribut resourc select paper propos extend version resourc select algorithm cori kl diverg order factor databas size distribut compar propos relev document distribut estim redd resourc select algorithm experi testb characterist redd extend kl diverg resourc select algorithm shown robust environ flexibl mixtur model collabor filter paper present flexibl mixtur model fmm collabor filter fmm extend exist partit cluster algorithm collabor filter cluster user item simultan assum user item belong singl cluster introduct prefer node propos framework explicitli model user rate item vari dramat user similar tast item relev document distribut estim method resourc select prior research varieti condit shown cori algorithm effect resourc select algorithm rang databas size studi larg paper show cori algorithm environ mix small larg databas resourc select algorithm propos inform databas size databas content show acquir databas size estim uncoop environ extens queri base sampl acquir resourc descript experi demonstr databas size estim accur larg databas estim produc compet method resourc rank algorithm effect cori algorithm algorithm result document rank cori algorithm prefer dbase graphic model collabor filter collabor filter gener techniqu exploit prefer pattern group user predict util item user previou research studi probabilist graphic model collabor filter promis result model succeed captur similar user item consid fact user similar interest item rate pattern user tend assign higher rate item user paper propos studi graphic model address distinct user prefer rate model call decoupl model introduc variabl decoupl user prefer rate call prefer model model order item prefer user user numer rate item empir studi dataset movi rate show due model distinct user prefer rate propos decoupl model significantli outperform exist approach compar prefer model perform wors decoupl model suggest explicit model underli user prefer import collabor filter afford ignor rate inform complet semisupervis learn method merg search engin result articl present semisupervis learn solut result merg problem kei contribut observ inform creat resourc descript resourc select creat central sampl databas guid normal document score return databas retriev time queri select databas return databas specif document score central sampl databas return databas independ document score document databas specif score databas independ score serv train data learn normal score document extens set experi demonstr method effect cori result merg algorithm varieti condit languag model framework resourc select result merg statist languag model propos recent inform retriev task includ resourc select task distribut inform retriev paper extend languag model approach integr resourc select ad hoc search merg result text databas singl probabilist retriev model approach design primarili intranet environ reason assum resourc provid homogen adopt kind search engin experi demonstr integr approach effect prior state art distribut ir sampl data regress merg search engin result paper address problem merg result obtain databas search engin distribut inform retriev environ prior research problem assum exchang statist normal score cooper solut heurist approach disadvantag languag model ir collect inform earli ag ir document repres control vocabulari paper explor usag meta data inform framework languag model present languag model advantag categori inform document improv retriev accuraci compar languag model tradit languag model trec dataset collect inform document obtain kmean cluster method languag model outperform tradit languag model verifi statement statist model scientif readabl paper present method statist model estim read difficulti web page languag model repres content typic readabl level read level classifi creat linear combin languag model surfac linguist featur experi show method accur wide flesch kincaid readabl formula keyword readabl flesch kincaid unigram languag model em 