hyperlink organ inform footprint search log support effect brows current search engin serv item search homepag find gener support exploratori search effect exploratori search user inform precis lack need knowledg formul effect queri queri support current search engin insuffici brows relat inform brows hyperlink embed web page paper propos leverag search log user brows hyperlink multi resolut topic map construct base search log specif treat search log footprint left previou user inform space build multi resolut topic map semant captur organ multipl granular topic map support user zoom zoom navig horizont inform space provid flexibl effect brows capabl end user test effect propos method support brows reli real search log commerci search engin implement propos method experiment result show propos topic map effect support brows hyperlink probabilist dyadic data analysi local global consist dyadic data aris real world applic social network analysi inform retriev order discov underli hidden structur dyadic data topic model techniqu propos typic algorithm includ probabilist latent semant analysi plsa latent dirichlet alloc lda probabl densiti function obtain algorithm support euclidean space previou studi shown natur occur data resid close underli submanifold introduc probabilist framework model topic geometr structur dyadic data explicitli take account local manifold structur specif local manifold structur model graph graph laplacian analog laplac beltrami oper manifold appli smooth probabl densiti function result obtain probabilist distribut concentr data manifold experiment result real data set demonstr effect propos approach 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 local preserv nonneg matrix factor matrix factor techniqu frequent appli inform process task neg matrix factor nmf receiv consider attent due psycholog physiolog interpret natur occur data represent part base human brain hand geometr perspect data sampl low dimension manifold embed high dimension ambient space hope find compact represent uncov hidden topic simultan respect intrins geometr structur paper propos algorithm call local preserv neg matrix factor lpnmf purpos data point kl diverg evalu similar hidden topic optim map obtain featur valu hidden topic restrict neg vari smoothli geodes data manifold empir studi show encourag result propos algorithm comparison state ofth art algorithm larg high dimension databas 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 mine term associ pattern search log effect queri reformul studi method neg relev feedback dirichletrank solv gap problem pagerank shine search heterogen interrel entiti improv retriev accuraci difficult queri neg feedback improv search accuraci difficult topic underaddress import research question paper scenario search result poor top rank document relev user queri propos exploit neg feedback improv retriev accuraci difficult queri specif propos learn number top rank relev document rerank rest unseen document propos approach penal document similar relev document languag model framework evalu propos method adapt standard trec collect construct test collect difficult queri experi result show propos approach effect improv retriev accuraci difficult queri detect track latent factor onlin nonneg matrix factor mine correl bursti topic pattern coordin text stream previou work text mine exclus focus singl stream multipl text stream index set time point call coordin text stream offer opportun text mine major event new articl publish agenc languag tend cover event period exhibit correl bursti topic pattern new articl stream gener mine correl bursti topic pattern coordin text stream reveal interest latent associ event stream paper defin studi text mine problem propos gener probabilist algorithm effect discov correl bursti pattern bursti period text stream stream complet vocabulari english chines evalu propos method new data set literatur data set show effect discov meaning topic pattern data set pattern discov new data set accur reveal major common event cover stream new articl english chines pattern discov databas public stream match major research paradigm shift databas research propos method gener requir stream share vocabulari appli coordin text stream discov correl topic pattern burst multipl stream period learn web search log organ search result effect organ search result critic improv util search engin cluster search result effect organ search result user navig relev document quickli defici approach make work cluster discov necessarili correspond interest aspect topic user perspect cluster label gener inform user identifi cluster paper propos address defici learn interest aspect topic web search log organ search result gener meaning cluster label past queri word enter user evalu propos method commerci search engin log data compar tradit method cluster search result method give result organ meaning label instant web retriev instanc dattribut queri thing chang empir studi bug characterist modern open sourc softwar softwar error major system failur effect design tool support detect recov softwar failur requir deep understand bug characterist recent softwar develop process significantli chang wai includ bug detect tool shift multi thread architectur opensourc develop paradigm increas concern secur user friendli interfac result previou studi applic present softwar aspect secur concurr open sourcerel characterist studi addition previou studi base small number bug lead repres result investig impact factor softwar error languag model inform retriev document expans languag model inform retriev depend accur estim document model paper propos document expans techniqu deal problem insuffici sampl document construct probabilist neighborhood document expand document neighborhood inform expand document accur estim document model improv retriev accuraci document expans pseudo feedback exploit corpu structur combin improv perform experi result data set demonstr effect propos document expans method languag model inform retriev document expans languag model inform retriev depend accur estim document model paper propos document expans techniqu deal problem insuffici sampl document construct probabilist neighborhood document expand document neighborhood inform expand document accur estim document model improv retriev accuraci document expans pseudo feedback exploit corpu structur combin improv perform experi result data set demonstr effect propos document expans method subject categor weblog part dof dspeech base smooth latent semant analysi multipl dtype interrel data object occurr data common real applic latent semant analysi lsa successfulli identifi semant relat data lsa handl singl occurr relationship type object practic applic case multipl type object exist pair object pairwis occurr relat occurr relat exploit allevi data spars repres object meaningfulli paper propos algorithm lsa conduct latent semant analysi incorpor pairwis occurr multipl type object base mutual reinforc principl lsa identifi salient concept occurr data repres object unifi semant space lsa gener show variant lsa special case algorithm experi result show lsa outperform lsa multipl applic includ collabor filter text cluster text categor cw compar web search system mine clickthrough data collabor web search accur languag model estim document expans dirichlet pagerank web todai includ page intend deceiv search engin content link creat attain unwarr result rank link web page calcul author rank system consider page content trust paper assum exist mechan limit gy ngyi al trustrank estim trustworthi page unlik exist work trust identifi demot spam page propos incorpor trust estim process calcul author cautiou surfer appli total forti queri larg real world dataset demonstr incorpor trust author calcul cautiou surfer improv pagerank precis averag top result qualiti uiuc fmusc trec genom track report experi result collabor particip uiuc musc trec genom track particip adhoc task categor task studi mixtur languag model task experi result show structur theme base languag model approach effect improv retriev effect ad hoc tak latent dirichlet alloc method effect dimens reduct categor task web page cluster enhanc summar cbc cluster base text classif requir minim label data semi supervis learn method construct classifi label unlabel train data sampl unlabel data sampl improv accuraci train model extent exist method face difficulti label data suffici bias underli data distribut paper present cluster base classif cbc approach approach train data includ label unlabel data cluster guidanc label data unlabel data sampl label base cluster obtain discrimin classifi subsequ train expand label dataset effect propos method justifi analyt relat issu expand label dataset interact cluster classif discuss experiment result demonstr cbc outperform exist algorithm size label dataset small 