path match dom tree search log accur webpag cluster kroneck graph approach model network mine broad latent queri aspect search session search queri typic short mean underspecifi sens user broad latent queri aspect set keyword succinctli repres sens inform aid user reformul queri extract broad latent aspect queri reformul found histor search session log propos framework problem extract broad latent aspect reduc optim formal object function constraint total number aspect system store number aspect shown respons queri present algorithm find good set aspect pick aspect match queri empir result real world search engin log show signific gain strong baselin singl keyword reformul gain term human judg accuraci click data term consist aspect predict similar queri demonstr import broad queri aspect efficaci algorithm extract shatterplot fast tool mine larg graph quicklink select navig queri result quicklink websit navig shortcut displai websit homepag search result page user directli jump select point insid websit real estat search result page constrain valuabl pick set quicklink maxim benefit major user import problem search engin user brows trail obtain browser toolbar simpl probabilist model formul quicklink select problem combinatori optimizaton problem demonstr hard object propos algorithm provabl factor optim propos algorithm work tree find optim solut unlik previou algorithm algorithm incorpor natur constraint set chosen quicklink efficaci method demonstr empir result manual label set websit set quicklink click rate webpag obtain real world search engin gener succinct titl web url search engin automat provid titl result url link titl user persuad click url problem automat gener link titl url propos gener statist framework solv problem framework base inform divers collect sourc thought contribut candid link titl url incorpor context link titl constraint length framework applic scenario obtain succinct titl displai quicklink obtain titl url lack good titl construct succinct sitemap extens experi show method effect produc result trivial baselin mortal multi darm bandit formul studi variant arm bandit problem motiv commerc applic model arm stochast lifetim expir set algorithm continu explor arm contrast standard arm bandit model arm indefinit explor reduc optim arm identifi nearcertainti main motiv set onlin advertis ad limit lifetim due natur content campaign budget algorithm choos larg collect ad fulli explor typic ad lifetim present optim algorithm state awar determinist reward function case build techniqu obtain algorithm state oblivi stochast reward function case empir studi reward distribut includ deriv real world ad serv applic show propos algorithm significantli outperform standard multi arm bandit approach appli set graph dtheoret approach webpag segment categori subject descriptor gener term 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 epidem threshold real network viru propag real network long disinfect network valu infect rate viru death rate singl node immun answer question essenti devis network wide strategi counter virus addit viral propag similar principl spread rumor inform fad impli solut viral propag offer insight problem set answer question develop nonlinear dynam system nld accur model viral propag arbitrari network includ real synthes network graph propos gener epidem threshold condit nld system prove epidem threshold network invers largest eigenvalu adjac matrix final show epidem threshold infect die exponenti rate epidem threshold model subsum threshold special case graph erd nyi ba powerlaw homogen demonstr predict power model extens experi real synthes graph show threshold condit hold arbitrari graph final show util threshold condit practic dictat node immun assess effect throttl 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 inform surviv threshold sensor network abstract network sensor node bluetooth enabl cell phone node transmit inform link node datum piec inform report emerg condit sensor network nation tradit song mobil phone viru node transmit datum datum surviv viru case condit viru die link node fault probabl import need ascertain surviv datum propos solv problem linear dynam system fix point stabil theorem provid closedform formula surprisingli depend addit paramet largest eigenvalu connect matrix illustr accuraci analysi realist real set mote sensor network intel mit gnutella network 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 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 page dlevel templat detect isoton smooth develop framework page level templat detect problem framework built main idea automat gener train data classifi page assign templat score dom node page global smooth node classifi score solv regular isoton regress problem simpl power abstract templat page extens experi human label test data show approach detect templat effect visual larg network min dcut plot dplot dmat normal comput social network spot abnorm network internet web graph answer question vital outlier detect terrorist network illeg monei launder ring forecast simul comput viru spread heart problem find properti real graph persist multipl disciplin list pattern law includ min cut plot discov part netmin packag larg graph visual feedback pattern signific deviat expect pattern immedi flag user abnorm graph part netmin plot tool visual adjac matrix graph innov wai find outlier netmin mat recurs matrix graph gener successfulli model pattern found real world graph quickli gener realist graph captur essenc graph paramet present result multipl larg real graph show effect approach evolutionari cluster graph mine law gener algorithm web abnorm social network normal similar question import field data intuit cast graph exampl rang comput network sociolog biologi relat databas terminolog repres graph lot question boil gener synthet realist graph answer understand pattern common real world graph consid mark normal realism survei give overview incred varieti work problem main contribut integr point view physic mathemat sociolog comput scienc briefli describ recent advanc relat interest graph problem neighborhood format anomali detect bipartit graph real applic model bipartit graph user file system trader stock financi trade system confer author scientif public network introduc oper bipartit graph identifi similar node neighborhood format find abnorm node anomali detect propos algorithm comput neighborhood node random walk restart graph partit propos algorithm identifi abnorm node neighborhood inform evalu qualiti neighborhood base semant dataset measur perform anomali detect algorithm manual inject anomali effect effici method confirm experi real dataset realist vast bodi theoret research focus overli simplist model parallel comput notabl pram overli specif model repres real world kind model encourag exploit formal loophol reward develop techniqu yield perform rang current futur parallel machin paper offer parallel machin model call logp reflect critic technolog trend underli parallel comput intend serv basi develop fast portabl parallel algorithm offer guidelin machin design model strike balanc detail simplic order reveal import bottleneck make analysi interest problem intract model base paramet abstractli comput bandwidth commun bandwidth commun delai effici coupl commun comput portabl parallel algorithm typic adapt machin configur term paramet util model demonstr exampl implement cm realist vast bodi theoret research focus overli simplist model parallel comput notabl pram overli specif model repres real world kind model encourag exploit formal loophol reward develop techniqu yield perform rang current futur parallel machin paper offer parallel machin model call logp reflect critic technolog trend underli parallel comput intend serv basi develop fast portabl parallel algorithm offer guidelin machin design model strike balanc detail simplic order reveal import bottleneck make analysi interest problem intract model base paramet abstractli comput bandwidth commun bandwidth commun delai effici coupl commun comput portabl parallel algorithm typic adapt machin configur term paramet util model demonstr exampl implement cm fulli automat cross dassoci larg spars binari matric aris numer data mine applic analysi market basket web graph social network citat inform retriev collabor filter spars matrix reorder virtual popular method analysi matric mean cluster meti graph partit svd pca frequent itemset mine requir user paramet number cluster number princip compon number partit support choos suitabl valu paramet challeng problem cross associ joint decomposit binari matrix disjoint row column group rectangular intersect group homogen start principl furnish clear informationtheoret criterion choos good cross associ paramet number row column group provid scalabl algorithm approach optim algorithm paramet free requir user intervent practic scale linearli problem size applic larg matric final present experi multipl synthet real life dataset method high qualiti intuit result materi base work support nation autopart paramet dfree graph partit outlier detect introduct motiv larg spars graph aris applic guis cons quentli import preval problem discov structur wide studi domain socialnetwork citat network ecolog food web protein interact graph structur insight graph detect amp quot commun amp quot problem descript graph set edg connect anypair node set dmat recurs model graph mine normal comput social network spot abnorm network internet web graph answer question vital outlier detect terrorist network illeg monei launder ring forecast simul comput viru spread heart problem find properti real graph persist multipl disciplin list law importantli propos simpl parsimoni model recurs matrix mat model quickli gener realist graph captur essenc graph paramet contrari exist gener model trivial gener weight direct bipartit graph subsum celebr erd nyi model special case match power law behavior deviat winner model pennock al present result multipl larg real graph show paramet fit algorithm automat fast fit real dtime expect dmaxim algorithm acquir multiplanar map indoor environ mobil robot epidem spread real network eigenvalu viewpoint abstract viru propag real network epidem threshold exist finit powerlaw graph finit graph long ittak disinfect network valu infect rate viru death rate answer question provid equa tion accur model viru propag network includ real synthes networkgraph propos gener epidem threshold condit appli arbitrari graph weprov reason approxim epidem threshold network close relatedto largest eigenvalu adjac matrix final question show infec tion tend exponenti epidem threshold show epidem threshold modelsubsum threshold special case graph erd amp quot os enyi ba power law homog neou show threshold infinit power law graph final illustr thepredict power model extens experi real synthes graph show thatour threshold condit hold arbitrari graph em learn model indoor environ mobil robot paper describ algorithm gener compact model indoor environ mobil robot algorithm emploi expect maxim algorithm fit lowcomplex planar model data collect rang finder panoram camera complex model determin model fit increment ad remov surfac final post process step measur convert polygon project surfac model empir result obtain mobil robot illustr high resolut model acquir reason time real dtime algorithm acquir multi dplanar volumetr model mobil robot 