contextu dbandit approach person new articl recommend person web servic strive adapt servic advertis new articl individu user make content user inform recent advanc problem remain challeng reason web servic featur dynam chang pool content render tradit collabor filter method inapplic scale web servic practic interest call solut fast learn comput work model person recommend new articl contextu bandit problem principl approach learn algorithm sequenti select articl serv user base contextu inform user articl simultan adapt articl select strategi base user click feedback maxim total user click contribut work fold propos gener contextu bandit algorithm computation effici motiv learn theori argu bandit algorithm reliabl evalu offlin previous record random traffic final offlin evalu method successfulli appli algorithm yahoo front page todai modul dataset million event result show click lift compar standard context free bandit algorithm advantag greater data scarc maintain equilibria explor sponsor search auction introduc explor scheme aim learn advertis click rate sponsor search auction minim effect advertis incent scheme preserv current rank price polici search engin introduc paramet control rate explor paramet set explor learn advertis click rate time elimin incent advertis alter submit bid advertis inform search engin show goal achiev incent deviat made arbitrarili small appropri set explor rate advertis alter bid bound revenu loss due explor optim high probabl algorithm contextu bandit problem learn log implicit explor data learn log implicit explor data unbias problem larg unlabel sampl boost perform learn algorithm small set label exampl set descript partit distinct view motiv task learn classifi web page descript web page partit word occur page word occur hyperlink point page assum view su cient learn label data goal view inexpens unlabel data augment amuch smaller set label exampl speci calli presenc distinct view suggest strategi learn algorithm train separ view algorithm predict unlabel exampl enlarg train set goal paper provid pac style analysi set broadli pac style framework gener problem learn label unlabel data provid empir result real web page data indic unlabel exampl lead signi improv hypothes practic part analysi provid agnost activ learn constraint import weight awar gradient updat error dcorrect tournament abstract present famili adapt pairwis tournament provabl robust larg error fraction determin largest element set tournament nk pairwis comparison log depth number player robust paramet reason valu tournament give reduct multiclass binari classif machin learn yield analysi problem featur hash larg scale multitask learn empir evid suggest hash effect strategi dimension reduct practic nonparametr estim paper provid exponenti tail bound featur hash show interact random subspac neglig high probabl demonstr feasibl approach experiment result case multitask learn hundr thousand task tutori summari reduct machin learn tutori summari activ learn import weight activ learn present practic statist consist scheme activ learn binari classifi gener loss function algorithm import weight correct sampl bia control varianc give rigor label complex bound learn process learn nonlinear dynam model present approach learn nonlinear dynam model lead set tool capabl solv problem difficult provid theori show approach consist model long rang structur appli approach motion captur highdimension video data yield result superior standard altern offset tree learn partial label multi dlabel predict compress sens multi label predict problem larg output space assumpt output sparsiti target vector small support develop gener theori variant popular ecoc error correct output code scheme base idea compress sens exploit sparsiti method regard simpl reduct multilabel regress problem binari regress problem shown number subproblem logarithm total number label valu make approach radic effici state prove perform guarante method test empir featur hash larg scale multitask learn empir evid suggest hash effect strategi dimension reduct practic nonparametr estim paper provid exponenti tail bound featur hash show interact random subspac neglig high probabl demonstr feasibl approach experiment result case multitask learn hundr thousand task error dcorrect tournament abstract present famili adapt pairwis tournament provabl robust larg error fraction determin largest element set tournament nk pairwis comparison log depth number player robust paramet reason valu tournament give reduct multiclass binari classif machin learn yield analysi problem condit probabl tree estim analysi algorithm main result problem estim condit probabl label time log number label analyz natur reduct problem set binari regress problem organ tree structur prove regret bound scale depth tree motiv analysi propos onlin algorithm provabl construct logarithm depth tree set label solv problem test algorithm empir show work succesfulli dataset roughli label learn nonlinear dynam model present approach learn nonlinear dynam model lead set tool capabl solv problem difficult provid theori show approach consist model long rang structur appli approach motion captur highdimension video data yield result superior standard altern search dbase structur predict describ adapt applic search base structur predict algorithm searn unsupervis learn problem show reduc unsupervis learn supervis learn demonstr high qualiti unsupervis shift reduc pars model addition show close connect unsupervis searn expect maxim final demonstr efficaci semi supervis extens kei idea enabl applic predict idea unsupervis learn provabl secur steganographi agnost activ learn state analyz activ learn algorithm work presenc arbitrari form nois algorithm agnost activ reli assumpt sampl drawn fix distribut show achiev exponenti improv requir ln sampl find optim classifi usual sampl complex supervis learn set consid realiz case includ learn threshold classifi learn homogen linear separ respect input distribut uniform unit sphere hash kernel structur data propos hash facilit effici kernel gener previou work sampl show principl comput kernel matrix data stream spars featur space give deviat bound exact kernel matrix applic estim string graph search dbase structur predict describ adapt applic search base structur predict algorithm searn unsupervis learn problem show reduc unsupervis learn supervis learn demonstr high qualiti unsupervis shift reduc pars model addition show close connect unsupervis searn expect maxim final demonstr efficaci semi supervis extens kei idea enabl applic predict idea unsupervis learn dfinanc wager mechan forecast examin class wager mechan design elicit truth predict group peopl requir subsidi propos number desir properti wager mechan identifi mechan weight score wager satisfi properti show singl paramet gener weight score wager mechan satisfi properti explor variant core mechan base practic consider categori subject descriptor explor scaveng predict index fast search tackl comput problem queri condit search machin learn score rule queri distribut build predict index precomput list potenti result sort base expect score result futur queri predict index datastructur support anytim algorithm approxim retriev top element gener approach applic webpag rank internet advertis approxim nearest neighbor search effect set standard techniqu invert indic intract experiment find substanti improv exist method internet advertis approxim nearest neighbor spars onlin learn truncat gradient propos gener method call truncat gradient induc sparsiti weight onlin learn algorithm convex loss method essenti properti degre sparsiti continu paramet control rate sparsif sparsif total sparsif approach theoret motiv instanc regard onlin counterpart popular regular method batch set prove small rate sparsif result small addit regret respect typic onlin learn guarante final approach work empir appli dataset find dataset larg number featur substanti sparsiti discover spars onlin learn truncat gradient propos gener method call truncat gradient induc sparsiti weight onlin learn algorithm convex loss method essenti properti degre sparsiti continu paramet control rate sparsif sparsif total sparsif approach theoret motiv instanc regard onlin counterpart popular regular method batch set prove small rate sparsif result small addit regret respect typic onlin learn guarante final approach work empir appli dataset find dataset larg number featur substanti sparsiti discover offset tree learn partial label import weight activ learn present practic statist consist scheme activ learn binari classifi gener loss function algorithm import weight correct sampl bia control varianc give rigor label complex bound learn process robust reduct rank classif reduc rank measur area receiv oper characterist curv auc binari classif core theorem show binari classif regret induc binari problem impli auc regret larg improv approach order regress score regret transform nr number element robust reduct rank classif reduc rank measur area receiv oper characterist curv auc binari classif core theorem show binari classif regret induc binari problem impli auc regret larg improv approach order regress score regret transform nr number element epoch dgreedi algorithm multi darm bandit side inform maintain equilibria explor sponsor search auction introduc explor scheme aim learn advertis click rate sponsor search auction minim effect advertis incent scheme preserv current rank price polici search engin introduc paramet control rate explor paramet set explor learn advertis click rate time elimin incent advertis alter submit bid advertis inform search engin show goal achiev incent deviat made arbitrarili small appropri set explor rate advertis alter bid bound revenu loss due explor suboptim behavior bay mdl classif misspecif show form bayesian mdl infer appli classif problem inconsist mean exist learn problem amount data gener error mdl classifi bay classifi rel bayesian posterior remain bound smallest achiev gener error bayesian point view result reinterpret bayesian infer inconsist misspecif countabl infinit model extens discuss result bayesian mdl perspect agnost activ learn state analyz activ learn algorithm work presenc arbitrari form nois algorithm agnost activ reli assumpt sampl drawn fix distribut show achiev exponenti improv requir ln sampl find optim classifi usual sampl complex supervis learn set consid realiz case includ learn threshold classifi learn homogen linear separ respect input distribut uniform unit sphere outlier detect activ learn pac model dfree reinforc learn markov decis process finit state size action space size state propos algorithm delai learn prove pac achiev optim perform sa timestep sa space improv bound previou algorithm result prove effici reinforc learn learn model mdp experi learn take place singl continu thread experi reset parallel sampl smaller storag experi requir delai learn experi comput cost previou pac algorithm cover tree nearest neighbor present tree data structur fast nearest neighbor oper gener npoint metric space data set consist point data structur requir space metric structur maintain perform properti navig net krauthgam amp lee point set bound expans constant measur intrins dimension defin karger amp ruhl cover tree data structur construct log time nearest neighbor queri requir time logarithm log time experiment result show speedup brute forc search vari order magnitud natur machin learn dataset outlier detect activ learn predict condit quantil reduct classif show reduc process predict condit quantil median solv classif accompani theoret statement show regret classifi bound regret quantil regress quantil loss test reduct empir exist quantil regress method larg real world dataset discov state theart perform dgovern public dprivat partnership canada failur longer option weight dagainst dall tti chicago abstract reduct convert solver task solver task introduc notion reduct analyz reduct classif task introduct natur supervis learn task classif varieti form binari multiclass cost sensit import weight approach analyz task independ devis algorithm approach reduct convert unexplor task task analyz automat transfer exist theori algorithm understood domain domain investig approach classif problem reduct power tool number desir properti show similar task task reduc sensit error correct output code cross valid problem cross valid model test full train set mean repeat resampl maxim total number point test potenti help protect overfit improv comput power recent reduct comput cost classif algorithm develop close form solut perform cross valid class learn algorithm make test thousand million variant learn model data calcul cross valid perform larger number tune model empir show larg number model risk overfit increas perform estim cross valid longer effect estim gener paper empir remind danger cross valid close form solut make evalu cross valid problem interest addit extens experi expos discuss effect overus misus cross valid aspect includ model select featur select data dimension illustr synthet benchmark real world data set comparison tight gener error bound investig empir applic bound number true error rate learn classifi hold exampl chosen independ random fix distribut collect trick includ techniqu unlabel data tight derandom random bound tight form progress valid bound exact form test set bound bound implement semibound packag freeli relat reinforc learn perform classif perform error limit reduct classif task introduc reduct base model analyz supervis learn task model devis reduct multi class cost sensit classif binari classif guarante learn binari classifi error rate cost sensit classifi cost time expect sum cost labl cost sensit classif emb bound loss finit choic supervis learn task result show task solv binari classif oracl final present experiment result show reduct outperform exist algorithm multi class cost sensit learn covert dparti comput introduc concept covert parti comput ordinari secur parti comput guarante knowledg leak input individu parti result comput covert parti comput emploi steganographi yield addit guarante eavesdropp determin parti perform comput simpli commun learn xa xb parti run protocol protocol conclud parti determin ran protocol distinguish xa xb uniformli chosen random bit covert parti comput construct protocol return xa xb equal interest romant interest parti determin ran protocol xa xb equal interest introduc secur definit covert parti comput construct protocol provabl secur base decision di hellman assumpt tutori practic predict theori classif discuss basic predict theori impact classif success evalu implic learn algorithm design learn algorithm execut tutori meant comprehens compil result theoret rigor practic import implic suboptim behavior bay mdl classif misspecif show form bayesian mdl infer appli classif problem inconsist mean exist learn problem amount data gener error mdl classifi bay classifi rel bayesian posterior remain bound smallest achiev gener error bayesian point view result reinterpret bayesian infer inconsist misspecif countabl infinit model extens discuss result bayesian mdl perspect iter method multi dclass cost dsensit learn cost sensit learn address issu classif presenc vari cost type misclassif paper present method solv multi class cost sensit learn problem binari classif algorithm algorithm deriv kei idea iter weight expand data space gradient boost stochast ensembl establish theoret guarante perform method show variant possess boost properti form weak learn assumpt compon binari classifi empir evalu perform propos method benchmark data set verifi method gener achiev result repres method cost sensit learn term predict perform cost minim case comput effici object evalu criterion cluster propos test object criterion evalu cluster perform cluster algorithm run unlabel data aid classif algorithm accuraci quantifi pac mdl bound semisupervis set cluster algorithm natur separ data hidden label small number cluster perform simpl extens argument lead object model select method experiment result text analysi dataset demonstr approach empir result competit bound test set perform natur dataset suboptim behaviour bay mdl classif misspecif tell human comput automat email account yahoo ll ask prove human comput singl comput program thousand free email account bad yahoo prove comput human prove human human idea ture test human judg ask bunch question decid depend answer talk human comput prove human comput matter requir test set test comput grade human pass paradox comput pass lingo requir captcha reduct classif task introduc reduct base model analyz supervis learn task model devis reduct multi class cost sensit classif binari classif guarante learn binari classifi error rate cost sensit classifi cost time expect sum cost labl cost sensit classif emb bound loss finit choic supervis learn task result show task solv binari classif oracl final present experiment result show reduct outperform exist algorithm multi class cost sensit learn comput shell decomposit bound haussler kearn seung tishbi introduc notion shell decomposit union bound mean understand empir phenomena learn curv phase transit variant idea deriv upper bound gener error hypothesi comput train error histogram train error hypothes class case bound significantli tighter tradit bound comput train error cardin class result view provid rigor foundat model select algorithm propos scheffer joachim correl equilibria graphic game examin correl equilibria recent introduc formal graphic game succinct represent multiplay game establish natur power relationship graphic structur multiplay game markov network repres distribut joint action main result establish markov network succinctli repres correl equilibria graphic game expect payoff equival main result gener algorithm comput correl equilibria graphic game base markov network special class graphic game includ tree algorithm run time polynomi graphic game represent polynomi number player exponenti graph degre pac dmdl bound point number standard sampl complex bound vc dimens pac bay relat number bit requir commun label unlabel data natur commun game motiv observ give gener sampl complex bound base game unifi dierent bound common framework captcha hard ai problem secur introduc captcha autom test human pass current comput program pass program high success captcha solv unsolv artifici intellig ai problem provid construct captcha captcha applic practic secur approach introduc class hard problem exploit secur purpos research cryptographi posit impact algorithm factor discret log hope hard ai problem secur purpos advanc field artifici intellig introduc famili ai problem construct captcha show solut problem steganograph commun captcha cost dsensit learn cost dproportion weight propos evalu famili method convert classifi learn algorithm classif theori cost sensit algorithm theori propos convers base cost proportion weight train exampl realiz feed weight classif algorithm boost care subsampl give theoret perform guarante propos method empir evid practic altern exist approach propos cost method base cost proportion reject sampl ensembl aggreg achiev excel predict perform publicli dataset drastic reduc comput requir method explor metric state space present metric provabl optim algorithm reinforc learn markov decis process natur metric state space construct accur local model algorithm gener algorithm kearn singh assum black box approxim plan unlik origin metric find optim polici amount time directli depend size state space depend cover number state space inform cover number number neighborhood requir accur local model microchoic bound bound learn algorithm major topic machin learn determin good upper bound true error rate learn hypothes base empir perform train data paper demonstr adapt bound design learn algorithm oper make sequenc choic bound call microchoic bound similar occam style bound make learn algorithm bound style freund fre show combin bound freund queri tree approach produc version freund queri tree structur implement algorithm effici approxim optim approxim reinforc learn order solv realist reinforc learn problem critic approxim algorithm paper competit analysi explor fexploit tradeoff investig explor exploit trade reinforc learn competit analysi appli abstract model state prove lower upper bound competit ratio essenti conclus analysi optim explor exploit trade easier piec extra knowledg stop time upper lower bound optim exploit polici pac dbay margin show relat thing classi er consist weight sum featur larg margin construct stochast classi er neglig larger train error rate stochast classi er futur error rate bound depend margin distribut independ size base hypothesi class improv predict accuraci bound averag classifi present improv bound differ train test error vote classier improv averag bound theoret justic popular averag techniqu bayesian classic maximum entropi discrimin winnow bay point machin implic learn algorithm design bound true error present approach bound true error rate continu valu classifi base pac bay bound method construct distribut classifi determin sensit paramet model nois true error rate stochast classifi found sensit analysi tightli bound pac bay bound paper demonstr method artifici neural network result order magnitud improv determinist neural net bound risk sensit particl filter propos particl filter incorpor model cost gener particl approach motiv observ cost accident track hypothes signific area state space irrelev incorpor cost model particl filter state critic system perform track automat calcul cost model implement mdp function calcul estim track state experi mobil robot domain illustr appropri approach comput shell decomposit bound haussler kearn seung tishbi introduc notion shell decomposit union bound mean understand empir phenomena learn curv phase transit variant idea deriv upper bound gener error hypothesi comput train error histogram train error hypothes class case bound significantli tighter tradit bound comput train error cardin class result view provid rigor foundat model select algorithm propos scheffer joachim featureboost meta dlearn algorithm improv model robust machin learn algorithm lazi extract train set minimum inform need predict label lead model robust featur remov obscur futur test data backprop net train steer car typic learn recogn edg road learn recogn featur stripe paint road road edg disappear tunnel obscur pass truck net learn minimum steer train set contrast human drive remark robust featur obscur motiv propos framework robust learn bias induct learn model input present meta algorithm robust learn call featureboost demonstr problem backprop net nearest neighbor decis tree beat hold dout bound dfold progress cross dvalid empir error test set hold estim reliabl estim gener error observ error train set train estim fold cross valid practic hope accur hold estim reduc number train exampl argu fold estim fact achiev goal specif show nontrivi learn problem learn algorithm insensit order fold estim strictli accur singl hold estim data lt lt leav base varianc higher moment previou bound term sanitycheck compar fold estim train estim restrict vc dimens requir notion hypothesi stabil order avoid depend fold hypothesi random combin averag individu hypothes introduc progress validationa improv hold estim estim gener error wai good singl hold averag half exampl test procedur involv hold set test ad train set learn algorithm rerun microchoic bound bound learn algorithm major topic machin learn determin good upper bound true error rate learn hypothes base empir perform train data paper demonstr adapt bound design learn algorithm oper make sequenc choic bound call microchoic bound similar occam style bound make learn algorithm bound style freund fre show combin bound freund queri tree approach produc version freund queri tree structur implement algorithm effici probabilist plan graphplan framework graphplan planner enjoi consider success plan algorithm classic strip domain paper explor extent represent probabilist plan mdp style framework state world action probabilist object produc finit horizon conting plan highest probabl success horizon describ extens graphplan direct pgraphplan produc optim conting plan typic suffer perform hit compar graphplan appear fast compar approach probabilist plan problem tgraphplan run essenti speed graphplan produc potenti optim polici tgraphplan polici select action highest probabl trajectori current state goal ideal mont carlo hidden markov model learn dparametr model partial observ stochast process present learn algorithm parametr hidden markov model continu state observ space probabl densiti approxim sampl densiti tree gener sampl amont carlo version baum welch em emploi learn model data regular learn achiev exponenti shrink techniqu shrinkag factor determin effect capac learn algorithm anneal multipl iter baumwelch earli stop appli select model train mont carlo hmm run time fashion prove mild assumpt mont carlo hidden markov model converg local maximum likelihood space convent hmm addit provid empir result obtain gestur recognit domain introduct hidden markov model hmm appli successfulli larg rang learn monoton boolean function problem learn monoton boolean function uniform distribut specif polynomi number uniform random sampl unknown monoton boolean function polynomi comput time approxim describ simpl algorithm prove achiev error gamma omega gamma improv previou bound gamma omega gamma prove algorithm polynomi number sampl guarante error gamma log improv previou hard bound lower bound hold learn algorithm allow membership queri paper settl log factor question achiev error learn class monoton boolean function respect uniform distribut introduct monoton boolean function map 