exploit site dlevel inform improv web search estim advertis tail queri sponsor search inform retriev challeng comput advertis measur reusabl test collect test collect construct time consum expens process true cost amort reus collect hundr thousand experi experi involv system retriev document judg initi construct phase system hard evalu depend judgment miss judg document retriev experiment confid evalu potenti low propos method quantifi reusabl test collect evalu system propos method provid simpl highli effect test determin exist set judgment evalu system empir evalu trec dataset confirm us propos reusabl measur show method reliabl estim confid interv indic collect reusabl automat gener bid phrase onlin advertis preval onlin advertis method textual advertis produc textual ad advertis craft short creativ text ad link land page describ product servic promot advertis associ creativ set manual chosen bid phrase repres web search queri trigger ad effici land page bid phrase chosen bid phrase creativ produc templat ad campaign larg retail involv thousand land page ten hundr thousand bid phrase entir process labori studi aim automat construct onlin ad campaign land page propos algorithm method gener bid phrase suitabl input phrase relev reflect content page form queri web search engin end phase approach candid bid phrase gener number method includ monolingu translat model capabl gener phrase contain text input previous unseen phrase candid rank probabilist framework translat model favor relev phrase bid phrase languag model favor form phrase empir evalu base real life corpu advertisercr land page bid phrase confirm approach successfulli gener human craft bid phrase perform significantli pure text extract method research conduct author summer intern yahoo research anatomi ad structur index retriev sponsor search land page sponsor search ad select compet user attent interplai organ sponsor search result 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 ad click quantifi impact land page web advertis unbeknownst user queri submit search engin distinct search perform organ algorithm search return relev web page relat data map imag sponsor search return paid advertis enorm amount work invest understand user interact organ search surprisingli research dedic ad click situat aim correct end defin studi process context transfer user transit web search context land page ad click conclud vast major case user shown type page homepag homepag advertis categori brows brows subcatalog relat origin queri search transfer search result queri execut target site show type land page accur distinguish automat text classif final automat classifi correl land page type convers data provid advertis show convers rate user respons rate ad vari consider type find understand user respons search advertis gener land page advertis improv web site search engin select suitabl ad context transfer search advertis context transfer search advertis cross dlanguag queri classif web search exogen knowledg nearest dneighbor cach content dmatch applic search dbase method forecast ad impress contextu advertis contextu advertis call content match refer placement small textual ad content gener web page signific sourc revenu publish rang individu blogger major newspap time import advertis reach intend audienc reach depend total number exposur ad impress click rate ctr view probabl end user click ad shown orthogon critic factor difficult estim individu inform plan budget advertis campaign paper address problem forecast number impress chang ad system onlin expans rare queri sponsor search sponsor search system task match queri relev advertis current state art match algorithm expand user queri varieti extern resourc web search result expans base algorithm highli effect larg ineffici appli real time practic algorithm appli offlin popular queri result expens oper cach fast access queri time paper describ effici effect approach match ad rare queri process offlin approach build expand queri represent leverag offlin process relat popular queri experiment result show approach significantli improv effect advertis rare queri neglig increas comput cost web advertis advertis import sourc incom websit attent unsuspect uninterest visitor web advertis ad tend elabor anim graphic depend specif technolog displai ad visitor screen requir vast amount cpu power present dai desktop cpu easili ad consum substanti amount energi import environment reason reduc energi consumpt increas number ad counterproduct goal paper investig power consumpt web advertis purpos energi meter measur differ pc power consumpt brows web ad enabl brows ad block simul normal web brows creat browser base tool call autobrows period open url predefin list block advertis adblock extens mozilla firefox measur power consumpt browser addit apach http server mod proxi modul act adblock proxi server measur pc browser show averag addit energi consumpt displai web advertis put number perspect calcul total amount energi displai web advertis equival total yearli electr consumpt household netherland take averag wind turbin gener amount energi classifi search queri web sourc knowledg propos methodolog build robust queri classif system identifi thousand queri class deal real time queri volum commerci web search engin pseudo relev feedback techniqu queri determin topic classifi web search result retriev queri motiv search advertis primarili focu rare queri hardest point view machin learn aggreg account consider fraction search engin traffic empir evalu confirm methodolog yield consider higher classif accuraci previous report propos methodolog lead match onlin ad rare queri user experi introduct comput advertis swing swing learn advertis web textual advertis interpret search problem corpu ad displai context contrast convent inform retriev system return result corpu document lexic relat queri web advertis accept occasion desir show result ad relev user interest show irrelev ad avoid annoi user produc econom benefit paper pose decis problem swing show ad incom request propos method address problem simpl threshold approach machin learn approach collect analyz set candid ad augment extern knowledg experiment evalu base editori judgment show predict high accuraci swing content match sponsor search advertis search advertis web relev feedback note search base forecast ad volum contextu advertis contextu advertis estim number impress ad critic plan budget advertis campaign produc forecast larg margin error challeng attack problem simul presenc ad bid histor data involv billion impress appar enorm comput task reduc search task involv set distinct page data search made effici level search process experiment result show approach accur forecast expect number impress contextu ad real time support ddocument updat queri invert index cross dlingual queri classif preliminari studi optim relev revenu ad search queri substitut approach primari busi model web search base textual advertis contextu relev ad displai alongsid search result address problem select ad relev queri profit search engin show optim ad relev revenu equival select ad satisfi constraint natur incur high comput cost time constraint lead reduc relev profit propos stage approach conduct analysi ahead time offlin preprocess phase leverag addit knowledg impract real time rewrit frequent queri subsequ facilit fast accur onlin match empir evalu show method optim relev match state art method improv expect revenu optim revenu substanti improv expect revenu workshop target rank onlin advertis 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 relax text search taxonomi paper propos document retriev model text queri augment multi dimension taxonomi restrict restrict relax cost result qualiti model applic arena includ multifacet product local search document augment hierarch metadata topic locat present effici algorithm index queri process retriev model decompos queri process problem onlin search problem determin correct level relax cost incur gener top result budget relax search problem result relax cost produc minim cost show problem solvabl hierarch dimens np hard dimens admit effici approxim algorithm provabl guarante present experiment result evalu algorithm synthet real data show order magnitud improv baselin algorithm 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 featur select method text classif numer text document electron form dai document repres massiv amount inform easili access seek huge collect requir organ work organ document autom text classif accuraci understand system greatli influenc us paper seek advanc understand commonli text classif techniqu understand improv tool text classif begin clarifi assumpt made deriv naiv bay note basic properti propos wai extens improv investig qualiti naiv bay paramet estim impact classif analysi lead theorem explan improv found multiclass classif naiv bay error correct output code experiment evid commonli data set exhibit applic theorem final show fundament flaw commonli featur select algorithm develop statist base framework text featur select greater understand naiv bay properti text make text classif bandit taxonomi model dbase approach problem learn optim match onlin fashion featur space organ taxonomi formul multi arm bandit problem arm bandit depend due structur induc taxonomi propos multi stage hierarch alloc scheme improv explor exploit properti classic multiarm bandit polici scenario scheme taxonomi structur perform shrinkag estim bayesian framework exploit depend arm enhanc explor lose effici short term exploit prove scheme asymptot converg optim match conduct extens experi real data illustr efficaci scheme practic robust classif rare queri web knowledg semant approach contextu advertis memori requir xpath evalu xml stream import challeng evalu xpath queri xml stream spark interest past year number algorithm propos support wider fragment queri languag exhibit perform memori util algorithm date prohibit larg amount memori type queri natur question memori bottleneck inher artifact propos algorithm paper initi systemat theoret studi lower bound amount memori requir evalu xpath queri xml stream present gener lower bound techniqu queri specifi minimum amount memori algorithm evalu queri stream incur lower bound state term graph theoret properti queri proof base tool commun complex exploit insight learn lower bound obtain algorithm xpath evalu stream algorithm space close optimum algorithm deviat standard paradigm automata transduc avoid store larg transit tabl estim corpu size queri db hybrid integr nativ xml xqueri relat data sql paper present architectur design aspect nativ xml support db udb integr relat data flow engin describ xml compon db udb show xml process leverag infrastructur relat data db hybrid integr nativ xml xqueri relat data sql paper present architectur design aspect nativ xml support db udb integr relat data flow engin describ xml compon db udb show xml process leverag infrastructur relat data optim cursor movement holist twig join buffer queri evalu xml stream buffer queri evalu xml stream system rx part relat part xml statist learn techniqu cost xml queri develop cost model queri optim significantli harder xml queri tradit relat queri reason xml queri oper complex relat oper tabl scan join paper propos approach call comet model cost xml oper knowledg comet method propos address xml queri cost problem relat cost estim comet exploit set system catalog statist summar xml data set simpl path statist propos suit xml set queri xml stream stream algorithm xml queri answer call effici memori manag tractabl queri distinguish bound number concurr aliv answer candid time point bound delai node select candid show properti decid polynomi time queri defin determinist tree automata result obtain reduct bound valued problem recogniz relat unrank tree problem show decid polynomi time kei word stream tree automata xml databas memori requir xpath evalu xml stream import challeng evalu xpath queri xml stream spark interest past year number algorithm propos support wider fragment queri languag exhibit perform memori util algorithm date prohibit larg amount memori type queri natur question memori bottleneck inher artifact propos algorithm paper initi systemat theoret studi lower bound amount memori requir evalu xpath queri xml stream present gener lower bound techniqu queri specifi minimum amount memori algorithm evalu queri stream incur lower bound state term graph theoret properti queri proof base tool commun complex exploit insight learn lower bound obtain algorithm xpath evalu stream algorithm space close optimum algorithm deviat standard paradigm automata transduc avoid store larg transit tabl rox relat xml increas percentag data need busi applic gener xml format store xml nativ format facilit applic exchang busi object xml format queri portion xml document xqueri paper explor feasibl access nativ store xml data tradit sql interfac call relat xml rox order avoid costli convers legaci applic xqueri describ forc drive industri evolv rox scenario issu rais rox impact denorm data xml document discuss semant perform perspect weigh implic rox manag queri optim evalu binari xml encod optim fast stream base xml process paper object evalu perform impact binari xml encod fast stream base xqueri processor repres applic propos binari format compar standard xml parser investig individu effect binari encod techniqu share propos goal provid deeper understand perform impact binari xml encod order clarifi ongo contenti debat merit domain high perform xml stream process scalabl view expans peer mediat system integr data sourc peer mediat framework view defin peer logic compos term common approach execut queri mediat treat view data sourc black box mediat local decompos queri queri fragment submit data sourc process approach distribut dbmss treat view transpar box import fulli expand view merg queri black box approach lead ineffici queri plan peer mediat framework full view expans ve lead prohibit long queri compil time peer involv limit peer autonomi peer reveal view definit investig peer mediat framework tradeoff partial full ve distribut view composit scenario show favor respect queri execut respect queri compil time expand view common hidden peer subview case black box approach peer autonomi prohibit view import base hybrid strategi ve peer mediat propos extract relat xml document xml preval format data exchang xml document complex schema vari wide inform sourc applic contrast databas applic reli flat relat model propos partial supervis approach extract userdefin relat xml document unknown schema extract relat directli rdbm util inform integr data mine task method attempt automat captur lexic structur featur relev portion input document base user annot exampl inform extract relat interest document schema potenti di erent train exampl present preliminari experi show method capabl extract target relat xml document presenc signific variat document schema stream xpath process forward backward ax present stream algorithm evalu xpath express backward ax parent ancestor forward ax singl document order travers xml document stream xpath processor handl forward ax show experi algorithm significantli outperform factor tradit stream xpath engin algorithm scale retain relev portion input document memori engin successfulli process document gb size tradit xpath engin degrad consider perform document mb size fail complet document size mb super dfast xml wrapper gener db demonstr queri xml data sourc db xml wrapper garlic flavor feder queri process db queri decomposit distribut object dorient mediat system mediat wrapper approachtointegr data heterogen data sourc central sens singl mediat system number data sourc applic number data sourc increas central mediat architectur administr perform bottleneck paper present queri decomposit algorithm distribut mediat architectur commun mediat higher level commun mediat data sourc salient featur propos approach explor queri execut schedul data ow sourc integr object orient sourc provid servic program data ii handl function multipl implement mediat sourc iii multi phase queri decomposit combin heurist cost base strategi iv tree rebalanc distribut queri recompil keyword queri decomposit middlewar data integr distribut databas sql fmed statu report report edbt panel scientif data integr evalu join strategi distribut mediat abstract join algorithm evalu environ distribut main memori base mediat data sourc stream ship join ship bulk tupl mediat data sourc post process client extend stream semi join addit build main memori hash index client mediat ship algorithm materi join data client mediat algorithm suitabl sourc requir paramet execut queri web search engin comput softwar suitabl compar execut time obtain tupl analyz percentag time spent subsystem vari network commun speed bulk size data duplic join algorithm lead order magnitud perform differ mediat environ distribut data integr object dorient mediat server paper overview data integr system amo ii base wrapper mediat approach amo ii consist mediat databas engin process execut queri data store local extern data sourc ii object orient oo multi databas view reconcili data schema heterogen sourc capabl data store type data sourc translat integr oo mediat primit provid user consist view data sourc multi databas facil distribut amo ii system interoper feder data resid data sourc achiev high perform core system main memori dbm storag manag queri optim transact client server interfac disk backup amo ii data manag optim main memori access extens sql manag extern data distribut view expans compos mediat optim queri distribut compos mediat distribut mediat light dweight oodbm overview light weight object orient oo multi databas system name amo ii object orient multi databas queri view defin extern data sourc kind translat amo ii integr oo mediat primit multi databas facil distribut amo ii system interoper data resid data sourc achiev good perform system design main memori dbm storag manag queri optim transact client server interfac amo ii data manag optim main memori extens data type queri oper ad implement extern program languag extens essenti data integr introduct amo ii system root workstat version iri system ws iri daplex function data model core amo ii distribut mediat light dweight oodbm overview light weight object orient oo multi databas system name amo ii object orient multi databas queri view defin extern data sourc kind translat amo ii integr oo mediat primit multi databas facil distribut amo ii system interoper data resid data sourc achiev good perform system design main memori dbm storag manag queri optim transact client server interfac amo ii data manag optim main memori extens data type queri oper ad implement extern program languag extens essenti data integr introduct amo ii system root workstat version iri system ws iri daplex function data model core amo ii function queri optim object dorient view data integr abstract amo mediat system support passiv intrus integr data heterogen autonom data sourc base function data model declar function queri languag amosql foreign data sourc relat databas text file type data sourc wrap amo mediat make access amosql amo mediat commun multi databas construct amosql definit function queri oo view access amo server integr view function type deriv data sourc local data view definit store mediat databas paper describ amo multi databas queri facil optim techniqu calculu base function transform gener minim queri express queri decomposit cost base algebra optim step place object identifi oid gener correctli repres deriv object mediat select oid gener mechan avoid overhead gener mediat oid deriv object need process queri local data mediat databas valid deriv object assign oid complet queri view guarante system gener predic ad queri keyword heterogen data integr object orient view queri optim integr heterogen overlap databas object dorient transform integr data autonom heterogen data sourc requir mean mediat reconcil overlap conflict integr data desir mediat system store local data data sourc achiev accept queri respons time mediat system research problem work present mediat queri process framework base represent data mediat reconcili number auxiliari system defin object orient oo view overload function queri framework support defin overload late bind mechan oo view declar queri queri mediat oo view late bound subqueri invoc transform disjunct queri express consist complet queri guarante expand queri valid subqueri perform improv bytyp awar queri rewrit select oid gener mediat experi show propos queri optim dramat improv queri execut time compar naiv instanc orient queri strategi partial strategi function queri optim object dorient view data integr abstract amo mediat system support passiv intrus integr data heterogen autonom data sourc base function data model declar function queri languag amosql foreign data sourc relat databas text file type data sourc wrap amo mediat make access amosql amo mediat commun multi databas construct amosql definit function queri oo view access amo server integr view function type deriv data sourc local data view definit store mediat databas paper describ amo multi databas queri facil optim techniqu calculu base function transform gener minim queri express queri decomposit cost base algebra optim step place object identifi oid gener correctli repres deriv object mediat select oid gener mechan avoid overhead gener mediat oid deriv object need process queri local data mediat databas valid deriv object assign oid complet queri view guarante system gener predic ad queri keyword heterogen data integr object orient view queri optim calculu dbase transform queri object dorient view databas mediat system concept object orient oo view popular approach data integr report result optim queri integr oo view work develop oo view system data integr base amo databas mediat system paper describ system architectur implement take advantag queri optim techniqu improv perform queri integr oo view main featur system passiv mediat framework preserv autonomi data sourc select materi mechan minim number materi view object predic base mechan guarante valid materi view object complet queri view order reduc overhead passiv view integr inexpens calculu base transform gener minim queri express queri decomposit cost base algebra optim place incorpor associ pattern oper specif odmg oql extend databas program languag declar queri facil paper work address imped mismatch problem syntax semant program queri languag queri languag oval base function queri languag fql extend manipul complex object salient featur oval queri languag function natur queri languag make languag suitabl integr procedur program languag modular style queri definit ii schema inform express queri iii recurs evalu algebra oper set structur complex object 