fast onlin learn offlin initi time dsensit recommend forecast high ddimension data ecb public featur motif banknot paper download charg flda matrix factor latent dirichlet alloc translat relev score probabl contextu advertis inform retriev system convention assess document relev bag word model relev score document retriev queri difficult compar comput disjoint set textual featur task feder search result global threshold relev score requir score global compar achiev aim propos method monoton transform relev score probabl contextu advertis select engin vector space model calibr raw score base histor click data movi recommend semant enrich unifi relev model rate predict collabor filter explor fexploit scheme web content optim regress dbase latent factor model spatio dtempor approach collabor filter paper propos spatio tempor model collabor filter applic model base low rank matrix factor spatio tempor filter approach estim user item factor spatial compon regular factor exploit correl user item model function implicit feedback rate side inform user demograph brows histori incorpor correl factor markov random field prior probabilist framework neighborhood weight function user item covari tempor compon ensur user item factor adapt process occur time implement state space framework fast estim kalman filter spatio tempor filter st kf approach singl joint model simultan incorpor spatial tempor structur rate accur method predict futur rate ensur scalabl st kf emploi field approxim infer incorpor user item covari estim neighborhood weight help deal cold start warm start problem seamlessli singl unifi model framework covari predict factor user item neighborhood illustr method simul data benchmark data data obtain recommend system applic aris context yahoo front page spatio dtempor model estim click dthrough rate onlin model content optim address email describ content publish system select articl serv user choos editori program pool frequent refresh deploi major internet portal select articl serv hundr million user visit dai significantli increas number user click origin manual approach editor period select articl displai challeng face includ dynam content pool short articl lifetim stationari click rate extrem high traffic volum fundament problem solv quickli identifi item popular user segment exploit remain current explor underli pool constantli identifi promis altern quickli discard poor perform approach base track articl perform real time onlin model describ characterist constraint applic set discuss design choic show import effect coupl onlin model simpl random procedur discuss challeng encount product onlin content publish environ highlight issu deserv care attent analysi applic suggest number futur research avenu fast comput posterior mode multi dlevel hierarch model fast comput posterior mode multi dlevel hierarch model contextu advertis combin relev click feedback contextu advertis support web ecosystem todai user experi revenu share site publish ad ad network depend relev displai ad page content document retriev system relev provid score match individu ad document content page ad shown queri paper show match improv significantli augment ad page score function extra paramet logist regress model word page ad kei properti propos model map standard cosin similar match suitabl effici scalabl implement invert index model paramet valu learnt log ad impress click shrinkag estim combat sparsiti scale comput train extrem larg train corpu consist gigabyt data parallel fit algorithm hadoop framework experiment evalu provid show improv click predict holdout set impress click event larg scale real world ad placement engin model achiev lift precis rel tradit inform retriev model base cosin similar recal click test data parsimoni explan chang hierarch data multi darm bandit problem depend arm provid framework exploit depend arm multi arm bandit problem depend form gener model cluster arm find optim mdp base polici discount reward case give approxim formal error guarante discuss lower bound regret undiscount reward scenario propos gener level bandit polici propos instanti gener polici provid theoret justif regret instanti polici depend characterist cluster final empir demonstr efficaci polici larg scale realworld synthet data show significantli outperform classic polici design bandit independ arm transform invari robust shape analysi estim rate rare event multipl resolut problem estim occurr rate rare event extrem spars data pre exist hierarchi perform infer multipl resolut focu problem estim click rate webpag advertis pair call impress page ad classifi hierarchi captur broad contextu inform level granular typic click rate low coverag hierarchi spars overcom difficulti devis sampl method analyz special chosen sampl page train set estim click rate stage model stage imput number webpag ad pair resolut hierarchi adjust sampl bia stage estim click rate resolut incorpor correl sibl node tree structur markov model model scalabl suit larg scale data mine applic real world dataset consist billion impress demonstr neg click event train set method effect discrimin extrem rare event term click propens predict discret latent factor model larg scale dyadic data effici effect explan chang hierarch summari dimens attribut data warehous typic hierarch geograph locat sale data url web traffic log olap tool summar measur attribut total sale dimens hierarchi character trend anomali hierarch summari time number identifi larg total sale store differ expect valu parsimoni explan signific desir paper propos natur model parsimoni explan composit node weight root leaf path dimens hierarchi permit aggreg maxim gener dimens hierarchi formal model explain hierarch summari investig problem identifi optim parsimoni explan arbitrari root dimension tree hierarchi show explan comput effici time essenti proport number leav depth hierarchi method produc parsimoni explan output statist model predict confid interv make wide applic experi real data set demonstr util robust propos model explain signific superior parsimoni compar altern detect anomali cross dclassifi stream bayesian approach spatial scan statist approxim perform studi spatial scan statist determin hotspot spatial data wide epidemiolog biosurveil recent year effort invest design effici algorithm find high discrep region method rang fast heurist special case gener grid base method effici approxim algorithm provabl guarante perform qualiti paper make number contribut comput studi spatial scan statist describ simpl exact algorithm find largest discrep region domain propos approxim algorithm larg class discrep function includ kulldorff scan statist improv approxim versu runtim trade prior method extend simpl exact approxim algorithm data set lie natur grid accumul grid fourth conduct detail experiment comparison method number method demonstr approxim algorithm superior perform practic prior method exhibit good perform accuraci trade extant method includ paper suitabl data set modestli size data set order million data point method scale massiv data set natur examin small space stream algorithm yield accur answer provid neg result show stream algorithm provid approxim optim answer discrep maxim problem space linear input hunt bump maxim statist discrep anomali detect import applic biosurveil environment monitor compar measur data data drawn baselin distribut find cluster measur data repres true anomali cluster cluster baselin distribut discrep function examin measur data baselin data region anomal region defin high discrep paper present algorithm maxim statist discrep function space axi parallel rectangl give provabl approxim guarante addit rel method appli convex discrep function algorithm work connect statist discrep combinatori discrep roughli speak show order maxim convex discrep function class shape maxim linear discrep function set shape deriv gener discrep function data gener paramet exponenti famili gener wide kulldorff scan statist data poisson distribut present algorithm run ning log comput maximum discrep rectangl addit error kulldorff scan statist similar result hold rel error discrep function data come gaussian bernoulli gamma distribut prior work algorithm exact ran time hunt bump maxim statist discrep anomali detect import applic biosurveil environment monitor compar measur data data drawn baselin distribut find cluster measur data repres true anomali cluster cluster baselin distribut discrep function examin measur data baselin data region anomal region defin high discrep paper present algorithm maxim statist discrep function space axi parallel rectangl give provabl approxim guarante addit rel method appli convex discrep function algorithm work connect statist discrep combinatori discrep roughli speak show order maxim convex discrep function class shape maxim linear discrep function set shape deriv gener discrep function data gener paramet exponenti famili gener wide kulldorff scan statist data poisson distribut present algorithm run ning log comput maximum discrep rectangl addit error kulldorff scan statist similar result hold rel error discrep function data come gaussian bernoulli gamma distribut prior work algorithm exact ran time mine custom care dialog daili new network arriv model poisson process analyt simplic number traffic studi shown packet interarriv exponenti distribut evalu wide area trace investig number wide area tcp arriv process session connect arriv ftp data connect arriv ftp session telnet packet arriv determin error introduc model poisson process find user initi tcp session arriv remotelogin file transfer model poisson process fix hourli rate connect arriv deviat consider poisson model telnet packet interarriv exponenti grievous underestim bursti telnet traffic empir tcplib danzig al interarriv preserv bursti time scale ftp data connect arriv ftp session bunch connect burst largest larg complet domin ftp data traffic final offer result find relat similar widearea traffic 