mine billion dnode graph pattern gener tool analysi larg multi dmodal social network pattern gener surpris pattern call durat distribut mobil phone user abstract long phone call mobil user chanc call end current durat answer question studi call durat distribut cdd individu user larg mobil network analyz larg real network million user billion phone call record privat mobil phone compani larg citi span tb contribut tlac distribut fit cdd user tlac truncat version call log logist distribut skew power law distribut show tlac excel fit overwhelm major user exponenti lognorm contribut metadist model collect behavior user cdd show metadist distribut accur succinctli describ call durat behavior user larg mobil network method fast scale linearli number custom viru propag time dvari network theori immun algorithm find cluster subspac larg multi ddimension dataset fast algorithm time seri mine metric forens multi dlevel approach mine volatil graph twave high dorder analysi spatiotempor data eigenspok surpris pattern scalabl commun chip larg graph basset scalabl gatewai finder larg graph radiu plot mine tera dbyte scale graph algorithm larg multi million node graph facebook web crawl evolv time connect central node outlier graph show radiu plot pdf node radii answer question comput radiu plot prohibit expens graph reach planetari scale major contribut paper propos hadi hadoop diamet radii estim carefulli design fine tune algorithm comput diamet massiv graph run top hadoop mapreduc system excel scale number machin run hadi real world dataset includ yahooweb edg terabyt largest public graph analyz hadi report fascin pattern larg network surprisingli small effect diamet multi modal bi modal shape radiu plot palindrom motion time hcdf hybrid commun discoveri framework spex autom concis extract spatial gene express pattern fly embryo ish imag time seri analysi multipl resolut fast featur select fractal dimens ten year fast featur select fractal dimens dimensionalitycurs dimension reduct issu retain high interest data mine machin learn multimedia index cluster present fast scalabl algorithm quickli select import attribut dimens set dimension vector contrast older method method desir properti rotat attribut lead easi interpret result attribut spot attribut nonlinear correl requir constant number pass dataset good estim attribut idea fractal dimens dataset good approxim intrins dimens drop attribut affect appli method real synthet dataset gave fast good result introduct motiv whenmanag increas vo kroneck graph approach model network parsimoni linear fingerprint time seri studi problem mine summar multipl time seri effect effici propos plif method discov essenti characterist fingerprint exploit joint dynam numer sequenc fingerprint method benefit lead interpret featur versatil plif enabl numer mine task includ cluster compress visual forecast segment match top competitor task fast scalabl linear complex length sequenc experi synthet real dataset includ human motion captur data mb human motion sensor data sensor network router traffic data million raw updat year gener plif outperform top cluster method cluster top compress method compress time reconstruct error compress ratio meaning visual time enjoi linear scale spectral count triangl power dlaw network element dwise sparsif shiftr user ddirect shiftr user ddirect ipog fast interact proxim queri graph graph mine pattern gener tool rtg recurs realist graph gener random type propos recurs model gener realist graph evolv time model properti flexibl capabl gener cross product weight unweight direct undirect uni bipartit graph realist give graph obei eleven static dynam law real graph follow formal prove power law estim expon function model paramet parsimoni requir paramet fast linear number edg simpl intuit lead gener macroscop pattern empir show model mimic real world graph blognet unipartit undirect unweight node edg committe candid campaign donat bipartit direct weight node edg show handl time edg weight addit bursti similar pegasu peta dscale graph mine system abstract paper describ pegasu open sourc peta graph mine librari perform typic graph mine task comput diamet graph comput radiu node find connect compon size graph reach giga tera peta byte necess librari grow knowledg pegasu librari implement top hadoop platform open sourc version mapreduc graph mine oper pagerank spectral cluster diamet estim connect compon essenti repeat matrix vector multipl paper describ import primit pegasu call gim gener iter matrix vector multipl gim highli optim achiev good scale number machin linear run time number edg time faster perform optim version gim experi ran top supercomput world report find real graph includ largest publicli web graph yahoo billion edg keyword pegasu graph mine hadoop eigenspok surpris pattern scalabl commun chip larg graph model blog dynam blog produc post local underli mechan lead bursti tempor behavior observ blog network earlier work analyz network pattern blog found blog behavior bursti power law topolog tempor characterist intuit realist model introduc lead pattern focu work propos gener model simpl intuit principl individu blog produc tempor characterist blogospher global topolog network pattern power law degre distribut inter post time model zc cross approach base random walk combin power idea explor exploit make model simultan model topolog tempor dynam blogospher valid model experi larg collect blog million post snare link analyt system graph label risk detect classifi node network task wide rang applic anomali fraud detect resourc invest task fraud detect due high cost fraud automat detect potenti fraud quickli precis human investig work effici data analyt scheme put scheme bolster link analysi prove promis work build belief propag algorithm detect collus fraud scheme propos algorithm call snare social network analysi risk evalu allow domain knowledg link knowledg method success pinpoint misstat account sampl gener ledger data signific improv default heurist true posit rate lift factor default heurist appli snare task graph label gener publicli dataset show inform node network surprisingli high accuraci label snare applic wide varieti domain robust choic paramet highli scalabl linearli number edg graph bgp dlen pattern anomali internet rout updat border gatewai protocol bgp fundament comput commun protocol monitor mine bgp updat messag directli reveal health stabil internet rout make contribut firstli find pattern bgp updat similar power law lognorm margin pattern find anomali specif develop bgp len autom bgp updat analysi tool desir properti effect identifi phenomena unnot peculiar clotheslin behavior prolong spike long hour scalabl algorithm linear number time tick admin friendli give lead phenomenon interest showcas capabl bgp len identifi surpris phenomena verifi syadmin massiv trace bgp updat span year publicli site datapositori net larg human commun network pattern util ddriven gener dynammo mine summar coevolv sequenc miss valu multipl time sequenc miss valu propos dynammo summar compress find latent variabl idea discov hidden variabl learn dynam make algorithm function miss valu perform experi real synthet dataset span megabyt includ motion captur sequenc chlorin level drink water show propos dynammo method successfulli learn latent variabl evolut provid high compress loss reconstruct accuraci extract compact power featur segment interpret forecast complex linear durat sequenc dynammo mine summar coevolv sequenc miss valu multipl time sequenc miss valu propos dynammo summar compress find latent variabl idea discov hidden variabl learn dynam make algorithm function miss valu perform experi real synthet dataset span megabyt includ motion captur sequenc chlorin level drink water show propos dynammo method successfulli learn latent variabl evolut provid high compress loss reconstruct accuraci extract compact power featur segment interpret forecast complex linear durat sequenc bbm bayesian brows model petabyt dscale data quarter petabyt click log data estim relev url queri paper propos bayesian brows model bbm model techniqu advantag exact infer singl pass paralleliz effect present set experi test model effect effici set million search instanc million distinct queri bbm outperform state art competitor loglikelihood time faster clicklog set span quarter petabyt data showcas scalabl bbm implement commerci mapreduc cluster hour comput relev billion distinct queri url pair tangent surpris recommend algorithm doulion count triangl massiv graph coin count number triangl graph beauti algorithm problem gain import year due signific role complex network analysi metric frequent comput cluster coeffici transit ratio involv execut triangl count algorithm interest graph mine applic reli comput number triangl graph interest paper focu problem count triangl graph propos practic method triangl count algorithm potenti benefit straightforward triangl count algorithm black box perform experi real world network synthet dataset show method work high accuraci typic signific speedup result time faster perform shatterplot fast tool mine larg graph click chain model web search terabyt click log build effici effect click model commonli believ web search click log gold mine search busi reflect user prefer web document present search engin click model provid principl approach infer user perceiv relev web document leverag numer applic search busi due huge volum click data scalabl present click chain model ccm base solid bayesian framework scalabl increment perfectli meet comput challeng impos volumin click log constantli grow conduct extens experiment studi data set million queri session obtain juli commerci search engin ccm consist outperform state art competitor number metric log likelihood click perplex robust predict click posit rtg recurs realist graph gener random type propos recurs model gener realist graph evolv time model properti flexibl capabl gener cross product weight unweight direct undirect uni bipartit graph realist give graph obei eleven static dynam law real graph follow formal prove power law estim expon function model paramet parsimoni requir paramet fast linear number edg simpl intuit lead gener macroscop pattern empir show model mimic real world graph blognet unipartit undirect unweight node edg committe candid campaign donat bipartit direct weight node edg show handl time edg weight addit bursti similar optim approach effici process aggreg similar queri metric access method fast mine complex time dstamp event collect complex time stamp event find pattern anomali event meet person agenda item locat teleconfer public author unsupervis model object categori link analysi techniqu head pattern discoveri time devolv multi daspect data hierarch paramet dfree commun discoveri abstract larg bipartit graph document term userproduct graph find meaning commun quickli automat propos commun hierarchi commun commun propos method context specif cluster tree cct find commun multipl level user intervent base inform theoret principl mdl specif partit graph progress refin subgraph allow user quickli navig global coars structur graph focus local pattern fring benefit addit indic qualiti achiev compress typic hierarch method demonstr scalabl effect real larg graph util provis storag system provis storag system requir balanc cost solut benefit solut provid previou provis approach start fix set requir goal automat find minimum cost solut meet approach neglect cost benefit analysi purchas decis purchas storag system involv extens set trade off metric purchas cost perform reliabl avail power increas metric consequ fail account trade off lead poor return storag invest collect storag acquisit provis scenario show util function enabl cost benefit structur convei autom provis tool enabl tool make trade off system metric includ perform data protect purchas cost comprehens model internet auction fraud detect fraud detect common concern onlin auction websit fraudster manipul reput system commit nondeliveri fraud deal fraud group behavior network level featur user belief user paper loopi belief propag algorithm appli network level fraud detect classifi fraudster accomplic honest user method show good classif accuraci real data monitor network evolut mdl abstract public titl author evolut scientif topic commun time commun shrunk emerg split time time turn point propos timefal automat answer question social network graph evolv time main novelti propos approach user defin paramet reli principl minimum descript length mdl extract commun find good cut point time commun chang abruptli cut point good lead shorter data descript illustr algorithm synthet larg real dataset show result timefal agre human intuit rtm law recurs gener weight time devolv graph real weight graph chang time pattern obei earlier studi focu unweight graph except focu static snapshot report pattern discov real weight time evolv graph report pattern detect anomali natur graph make link predict provid criteria evalu synthet graph gener propos intuit easi construct weight time evolv graph fact prove gener produc graph obei pattern law observ date provid empir evid support claim compact similar join rtm law recurs gener weight time devolv graph real weight graph chang time pattern obei earlier studi focu unweight graph except focu static snapshot report pattern discov real weight time evolv graph report pattern detect anomali natur graph make link predict provid criteria evalu synthet graph gener propos intuit easi construct weight time evolv graph fact prove gener produc graph obei pattern law observ date provid empir evid support claim graphit visual queri system larg graph present graphit system user visual construct queri pattern find exact approxim match subgraph larg attribut graph visual match social network person occup attribut user draw star queri find ceo interact secretari manag account structur similar graphit rai algorithm run queri user chosen data graph gain benefit high speed scalabl abil find exact match graphit toler indirect path ceo account direct path exist graphit fast algorithm estim node proxim find match enabl scale graph databas size demonstr graphit usag benefit dblp author public graph consist node edg demo video graphit download bayesian tensor analysi protein complex identif supervis graph local cluster graph mine law gener tool digit media pervas sens web author mobil comput scientif medic instrument physic simul virtual world deliv vast dataset relat aspect live grow fraction big data unus underexploit due overwhelm scale data involv effect share understand wealth raw inform pose great challeng todai comput research order effect comput scale research develop group purchas rack comput storag server goal tashi project develop layer util softwar turn raw rack server easili manag cloud comput remot user share explor big data big data applic typic data hungri qualiti result improv quantiti data scalabl comput technolog accommod largest dataset import fortun applic disk bandwidth limit social network ahead ghost edg classif spars label network address problem classif partial label network network classif observ class label spars techniqu statist relat learn shown perform network classif task exploit depend class label neighbor node relat classifi fail unlabel node label neighbor support learn train phase infer test phase situat aris realworld problem observ label spars paper propos approach network classif combin aspect statist relat learn semi supervis learn improv classif perform spars network approach work ad ghost edg network enabl flow inform label unlabel node experi real world data set demonstr approach perform rang condit exist approach collect classif semi supervis learn fail task approach improv area roc curv auc point exist approach demonstr approach run time proport number label node number edg weight graph disconnect compon pattern gener vast major earlier work focus graph connect typic ignor giant connect compon unweight studi numer real weight graph report surpris discoveri node join form link social network motiv question connect compon graph form chang time node join network common repeat edg studi numer divers real graph citat network network social media internet traffic make contribut observ giant connect compon stabil size observ weight edg follow power law surpris expon propos intuit gener model graph growth obei observ pattern cut dand dstitch effici parallel learn linear dynam system smp multi core processor increas number core chip preval modern parallel comput goal make multi core multi processor architectur speed data mine algorithm specif present parallel algorithm approxim learn linear dynam system ld kalman filter kf ldss wide time seri analysi motion captur model visual track propos cut stitch ca method handl data depend chain structur hidden variabl ld parallel em base paramet learn algorithm implement algorithm openmp supercomput quad core commerci desktop experiment result show parallel algorithm cut stitch achiev compar accuraci linear speedup serial version addit cut stitch gener model similar linear structur hidden markov model hmm switch kalman filter skf weight graph disconnect compon pattern gener vast major earlier work focus graph connect typic ignor giant connect compon unweight studi numer real weight graph report surpris discoveri node join form link social network motiv question connect compon graph form chang time node join network common repeat edg studi numer divers real graph citat network network social media internet traffic make contribut observ giant connect compon stabil size observ weight edg follow power law surpris expon propos intuit gener model graph growth obei observ pattern mobil call graph power dlaw lognorm distribut analyz massiv social network gather record larg mobil phone oper million user ten million call examin distribut number phone call custom total talk minut custom distinct number call partner custom find distribut skew significantli deviat expect power law lognorm distribut analyz observ distribut number call distinct call partner total talk time propos powertrack method fit lesser suitabl distribut doubl pareto lognorm dpln distribut data track paramet time powertrack find graph time consist gener process natur result dpln distribut observ show gener process lend natur appeal social wealth interpret context social network discuss applic result model forecast colibri fast mine larg static dynam graph low rank approxim adjac matrix graph essenti find pattern commun detect anomali addition desir track low rank structur graph evolv time effici limit storag real graph typic thousand million node spars standard decomposit svd preserv sparsiti led develop method cur cmd seek nonorthogon basi sampl column row spars matrix approach typic produc overcomplet base wast space time paper propos famili colibri method deal challeng version static graph colibri iter find nonredund basi prove loss accuraci compar competitor cur cmd achiev signific save space time real data colibri requir space order magnitud faster proport squar number redund column addition propos effici updat algorithm dynam time evolv graph colibri evalu larg real network traffic dataset show colibri time faster publish competitor cmd unsupervis model recognit object categori combin visual content geometr similar link graph mine law gener tool digit media pervas sens web author mobil comput scientif medic instrument physic simul virtual world deliv vast dataset relat aspect live grow fraction big data unus underexploit due overwhelm scale data involv effect share understand wealth raw inform pose great challeng todai comput research order effect comput scale research develop group purchas rack comput storag server goal tashi project develop layer util softwar turn raw rack server easili manag cloud comput remot user share explor big data big data applic typic data hungri qualiti result improv quantiti data scalabl comput technolog accommod largest dataset import fortun applic disk bandwidth limit semi dsupervis learn base semiparametr regular semi supervis learn plai import role recent literatur machin learn data mine develop semisupervis learn techniqu led data mine applic recent year paper address semi supervis learn problem develop semiparametr regular base approach attempt discov margin distribut data learn parametr function exploit geometr distribut data learn parametr function incorpor supervis learn label data prior knowledg specif contribut present semi supervis learn approach incorpor unlabel data supervis learn parametr function learn data includ label unlabel data parametr function reflect geometr structur margin distribut data propos approach natur extend sampl data induct learn method natur approach famili algorithm develop base choic origin rkh loss function provid experiment comparison show propos approach lead state art perform varieti classif task demonstr approach successfulli transduct semisupervis set effici distribut mine classif defin solv problem distribut classif gener distribut mine distribut cloud multi dimension point classifi class find pattern rule lier cloud case sale item item sold record unit price quantiti custom repres distribut cloud point item bought group similar user market segment anomali fraud detect propos mine achiev goal main contribut theorem show wavelet speed cloud similar comput extens experi synthet real multidimension data set show method achiev faster wall clock time naiv implement compar occasion classif qualiti proxim track time devolv bipartit graph author confer network evolv time confer author close relat chang time larg time evolv bipartit graph set social network citat market basket analysi collabor filter goal monitor central individu node import author ii proxim node set node import author respect confer effici increment provid time answer propos ptrack ctrack base random walk restart power matrix tool experi real data show method effect effici mine result agre intuit achiev time speed qualiti loss outlier drobust cluster independ compon graph mine influenc propag 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 head pattern discoveri time devolv multi daspect data bayesian tensor approach dd face model random walk restart fast solut applic ddem multi dmodal queri system drosophila embryo databas amount biolog data publicli experienc exponenti growth technolog advanc onlin databas plai import role inform repositori easili access platform research commun contribut recent research project imag bioinformat produc number databas imag visual spatial express pattern gene fj spars graph mine compact matrix decomposit fast monitor proxim central time devolv bipartit graph increment tensor analysi theori applic data mine fractal power law stream monitor time warp distanc data stream process recent attract increas amount interest goal paper monitor numer stream find subsequ similar queri sequenc dtw dynam time warp distanc applic includ word spot sensor pattern match monitor biomed signal ekg ecg monitor environment seismic volcan signal dtw popular distanc measur permit acceler deceler studi finit store sequenc set applic network analysi sensor monitor massiv amount data arriv continu infeas save histor data propos spring algorithm solv problem provid theoret analysi prove spring sacrific accuraci requir constant space time time tick dramat improv naiv method experi real realist data illustr spring detect qualifi subsequ correctli offer dramat improv speed time naiv implement max margin framework imag annot multimod imag retriev scalabl model real graph kroneck multipl larg real graph gener synthet graph match properti similar degre distribut similar small diamet similar spectrum propos kroneck graph natur obei properti present kronfit fast scalabl algorithm fit kroneck graph gener model real network naiv approach fit super exponenti time contrast kron fit take linear time exploit structur kroneck product sampl experi larg real synthet graph show kronfit mimic pattern found target graph fit model paramet result synthet graph anonym extrapol graph summar probabilist tensor analysi akaik bayesian inform criteria 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 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 enhanc max margin learn multimod data mine multimedia databas problem multimod data mine multimedia databas address structur predict problem learn map input structur interdepend output variabl paper built exist literatur max margin base learn develop max margin learn approach call enhanc max margin learn emml framework addit appli emml framework develop effect effici solut multimod data mine problem multimedia databas main contribut includ develop max margin learn approach enhanc max margin learn framework effici learn faster converg rate verifi empir evalu appli emml approach develop effect effici solut multimod data mine problem highli scalabl sens queri respons time independ databas scale allow facilit multimod data mine queri larg scale multimedia databas excel exist multimod data mine method literatur scale advantag support complex analysi empir evalu state art multimod data mine method literatur emml gener framework evalu purpos appli berkelei drosophila embryo imag databas report perform comparison state art multimod data mine method cost deffect outbreak detect network network spread diseas product recommend diffus social network fundament process social network behavior cascad node node epidem new opinion rumor fad urban legend word mouth effect market rise websit free web base servic viru diseas propag chang social prioriti smoke recycl satur new coverag topic diffus blogger internet energ polit campaign cascad failur financi market local effect riot peopl walk lectureempir studi diffus experiment studi diffus long histori spread agricultur practic ryan gross adopt hybrid corn farmer iowa classic studi diffus interperson network plai import role adopt diffus social process spread medic practic coleman al studi adopt drug doctor illinoi clinic studi scientif evalu suffici convinc doctor social power peer led adoptiondiffus network initi node activ activ node spread influenc node graphscop paramet dfree mine larg time devolv graph fast deffort pattern match larg attribut graph focu larg graph node attribut social network node label person job titl set find subgraph match user queri pattern star queri find ceo strong interact manag lawyer account structur close similarli loop queri spot monei launder ring tradit sql base method recent graph index method return answer exact match exist method find exact match present user propos good order method toler indirect path ceo account sampl queri direct path exist featur scalabl gener queri nq node data graph node problem polynomi time complex nq prohibit rai graph rai method find high qualiti subgraph time linear size data graph experiment result dlbp author public graph node edg illustr effect scalabl approach result agre intuit speed excel take second averag node queri dblp graph fast direct dawar proxim graph mine paper studi asymmetr proxim measur direct graph quantifi relationship node group node measur graph mine task includ cluster link predict connect subgraph discoveri proxim measur base concept escap probabl strive summar multipl facet node proxim avoid pitfal altern proxim measur suscept uniqu featur measur account underli direct inform put special emphasi comput effici develop fast solut applic set experiment studi show us propos direct awar proxim method applic algorithm achiev signific speedup straightforward implement compact matrix decomposit larg spars graph larg spars graph find pattern anomali import applic model larg spars graph network traffic monitor research citat network analysi social network analysi regulatori network gene low rank decomposit svd cur power techniqu reveal latent hidden variabl pattern high dimension data method ignor sparsiti properti graph incur high memori comput cost practic propos method compact matrix decomposit cmd comput spars low rank approxim cmd dramat reduc comput cost space requir exist decomposit method svd cur cmd kei build block propos procedur effici construct analyz dynam graph real time applic data provid theoret guarante method present result real larg dataset network flow data gb trace host month dblp mb year show cmd order magnitud effici state art svd cur faster requir space reconstruct accuraci final demonstr cmd detect anomali monitor timeevolv graph successfulli detect worm hierarch scan pattern real network data pattern cascad behavior larg blog graph blog cite influenc link evolv popular blog post drop exponenti time question address work blog weblog import medium inform time public eas wide avail fact make headlin discuss discov evid polit event fact blog link creat publicli record inform influenc spread underli social network aggreg link blog post creat direct graph analyz discov pattern inform propag blogspac understand underli social network report surpris find blog link inform propag structur analyz largest dataset blog million blog post analysi shed light rumor virus idea propag social comput network mine larg graph stream matrix tensor tool boost dnearest neighbor queri estim suitabl queri radii parallel crawl onlin social network huge onlin social network retriev inform crawl improv crawl perform parallel crawler work independ paper present framework parallel crawler onlin social network util central queue show work practic describ implement crawler onlin auction websit crawler work independ fail crawler affect framework ensur redund crawl occur crawler built visit total approxim million auction user complet crawl netprob fast scalabl system fraud detect onlin auction network larg onlin network onlin auction user histori transact spot anomali auction fraud paper describ design implement netprob system propos solv problem netprob model auction user transact markov random field tune detect suspici pattern fraudster creat emploi belief propag mechan detect fraudster experi show netprob effici effect fraud detect report experi synthet graph node edg netprob spot fraudul node precis recal matter second report experi real dataset crawl ebai transact user netprob highli effect unearth hidden network fraudster realist respons time minut scenario underli data dynam natur propos increment netprob approxim fast variant net probe experi prove increment netprob execut doubli fast compar netprob retain accuraci fast effect method find correl attribut databas data mine compress kolmogorov complex abstract theori data mine analog relat algebra databas cluster algorithm data mine autom show answer question neg data mine close relat compress kolmogorov complex undecid data mine art goal find model pattern fit dataset tkde guidelin survei paper graph evolut densif shrink diamet real graph evolv time normal growth pattern social technolog inform network studi discov pattern static graph identifi properti singl snapshot larg network small number snapshot includ heavi tail degre distribut commun small world phenomena lack inform network evolut long period hard convert find statement trend time ric paramet dfree nois drobust cluster intellig system monitor larg cluster intemon prototyp monitor mine system larg cluster monitor host prototyp data center cmu snmp protocol store monitor data mysql databas visual time seri data jsp web base frontend interfac user set cluster monitor system abil automat analyz monitor data real time alert user potenti anomali state art stream mine method sophist definit anomali broken correl input stream pinpoint reason anomali intemon user friendli gui user perform interact mine task fulli oper omni dfamili dpurpos access method simpl effect make similar search effici effici process complex similar queri rdbm queri rewrit intellig system monitor larg cluster intemon prototyp monitor mine system larg cluster monitor host prototyp data center cmu snmp protocol store monitor data mysql databas visual time seri data jsp web base frontend interfac user set cluster monitor system abil automat analyz monitor data real time alert user potenti anomali state art stream mine method sophist definit anomali broken correl input stream pinpoint reason anomali intemon user friendli gui user perform interact mine task fulli oper fast random walk restart applic close relat node graph comput score quickli huge disk resid real graph random walk restart rwr good relev score node weight graph successfulli numer set automat caption imag gener connect subgraph person pagerank straightforward implement rwr scale larg graph requir quadrat space cubic pre comput time slow respons time queri propos fast solut problem heart approach exploit import properti share real graph linear correl blockwis commun structur exploit linear low rank matrix approxim commun structur graph partit sherman morrison lemma matrix invers experiment result corel imag dblp dabaset demonstr propos method achiev signific save straightforward implement save order magnitud pre comput storag cost achiev speed qualiti preserv supergraph visual larg social comput network visual find pattern outlier commun graph visual tool exist handl larg graph hundr thousand node possibl million edg graph bring challeng interact visual demand prohibit process power interact updat visual user overwhelm excess number graphic item cope problem propos formal innov graph hierarchi lead gmine system gmine promot scalabl hierarchi graph partit promot concomit present graph hierarchi origin graph extend analyt possibl integr graph partit interact environ stream graph dynam tensor analysi find pattern author keyword associ evolv time datacub product branchcustom sale inform matrix decomposit princip compon analysi pca variant invalu tool mine dimension reduct featur select rule identif numer set stream data text graph social network order author keyword propos envis higher order data tensor tap vast literatur topic method necessarili scale oper semi infinit stream introduc dynam tensor analysi dta method variant dta compact summari high order high dimension data reveal hidden correl algorithm design dta carefulli scalabl space effici store past fulli automat user defin paramet propos sta stream tensor analysi method fast stream approxim dta implement method appli real set anomali detect multi latent semant index real larg dataset network flow data gb month dblp mb year experi show method fast accur find interest pattern outlier real dataset center dpiec subgraph problem definit fast solut node social network authorship network find node author centerpiec direct indirect connect node common advisor start research area node belong isomorph scenario law enforc find master mind crimin connect current suspect gene regulatori network find protein particip pathwai protein viral market connect subgraph import step handl case queri node connect subgraph algorithm find intermedi node provid good connect origin queri node gener challeng multipl dimens queri node famili queri rang softand final design compar fast approxim studi qualiti speed trade present experi dblp dataset experi confirm propos method natur deal multi sourc queri result subgraph agre intuit wall clock time result dblp dataset show propos approxim achiev good accuraci speedup materi base work support sampl larg graph huge real graph deriv repres sampl algorithm comput interest measur shortest path central between impract larg graph graph sampl essenti robust inform dtheoret cluster automat mine fruit fly embryo imag present femin automat system imag base gene express analysi perform experi largest publicli collect drosophila ish situ hybrid imag show femin system achiev excel perform classif cluster content base imag retriev major innov femin automat discov latent spatial theme gene express lge embryo context oppos pattern disjoint portion embryo propos previou method distribut pattern discoveri multipl stream conclus recommend express materi author necessarili reflect view nation scienc foundat fund parti detect fraudul person network onlin auction abstract onlin auction gain immens popular creat access environ exchang good reason price surprisingli malevol auction user abus cheat paper propos method level fraud spot lf model techniqu fraudster typic carri fraudul activ detect fraudster preemptiv kei contribut mine user level featur number transact averag price good exchang initi belief spot fraudster introduc network level featur captur interact user show combin featur belief propag algorithm markov random field detect suspici pattern unnatur close nit group peopl trade algorithm scale linearli number graph edg illustr effect algorithm real dataset collect larg onlin auction site adapt dphase approach wifi locat sens evalu intrins dimens evolv data stream data mine fractal power law gmine system scalabl interact graph visual mine graph visual tool exist handl larg graph interact interest larg graph hundr thousand node graph bring challeng straightforward interact manipul prohibit slow sensori overload plot replot graph quickli user overwhelm vast volum inform screen clutter node edg overlap gmine system address issu summar multi resolut gmine offer multi resolut graph explor partit graph hierarchi commun commun store tree structur tree gmine offer summar implement innov subgraph extract algorithm visual output intemon intellig system monitor larg cluster intemon prototyp monitor mine system larg cluster monitor host prototyp data center cmu snmp protocol store monitor data mysql databas visual time seri data jsp web base frontend interfac user set cluster monitor system abil automat analyz monitor data real time alert user potenti anomali state art stream mine method sophist definit anomali broken correl input stream pinpoint reason anomali intemon user friendli gui user perform interact mine task fulli oper 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 spatial queri estim local uniform assumpt effici discoveri unusu pattern time seri intemon continu mine sensor data larg dscale dinfrastructur modern data center larg number compon monitor includ server switch router environment control system paper describ intemon prototyp monitor mine system data center snmp protocol monitor data center carnegi mellon store monitor data mysql databas allow visual time seri data jsp web base frontend interfac system administr set intemon cluster monitor system abil automat analyz correl monitor data real time alert administr potenti anomali effici state art stream mine method report broken correl input stream method intellig compress histor data avoid administr configur threshold base monitor band multidimension data modern disk deepli ingrain notion disk effici access dimension data current approach map multidimension data disk block effici access dimens trade effici access dimens equal penal access dimens exist technolog function readili insid disk firmwar identifi contigu logic block preserv spatial local multidimension dataset block span order hundr adjac track access minim posit cost paper detail technolog analyz trend show expos applic maintain exist abstract approach achiev access effici afford disk technolog sequenti access primari dimens access minim posit cost dimens experiment evalu prototyp implement demonstr reduct time multidimension data queri compar exist approach onlin latent variabl detect sensor network autolag automat discoveri lag correl stream data data stream multipl set environment medic socioeconom system fascin research problem set cluster summar forecast multiresolut symbol represent time seri effici accur search similar time seri discov interest pattern import trivial problem paper introduc represent time seri multiresolut vector quantiz mvq approxim distanc function novelti mvq local global inform origin time seri hierarch mechan process origin time seri multipl resolut propos represent symbol emploi kei subsequ potenti applic text base retriev techniqu similar analysi time seri propos method fast scale linearli size databas dimension contrari vast major literatur euclidean distanc mvq multi resolut hierarch distanc function perform experi real synthet data propos distanc function consist outperform major competitor euclidean dynam time warp piecewis aggreg approxim achiev precis recal cluster accuraci test dataset paramet dfree spatial data mine mdl spatial data consist set binari featur take valu collect spatial extent grid cell propos method simultan find spatial correl featur occurr pattern paramet emploi minimum descript length mdl principl coupl natur compress region defin good mean featur occurr pattern good help compress set locat featur convers spatial correl good help compress set featur region approach scalabl larg dataset number locat featur evalu method real synthet dataset 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 vivo visual vocabulari construct mine biomed imag larg collect medic imag condit treatment succinctli describ characterist set larg collect retin imag experiment condit normal detach reattach data mine biologist focu import region imag differ experiment condit imag text document find main term concept condit exist ir method tf idf lsi propos analog challeng case imag collect propos automat develop visual vocabulari break imag tile deriv kei tile vivo imag condit experi numer domain independ wai extract featur tile color histogram textur wai choos characterist tile pca ica perform experi dispar biomed dataset quantit measur success classif accuraci vivo achiev high classif accuraci class problem felin retin imag importantli qualit vivo excel job visual vocabulari term biolog mean corrobor domain expert spot characterist region imag text vocabulari term document highlight differ pair imag dbase robust outlier detect high dimension dataset detect outlier import problem applic typic possess high dimension dataset high dimension space data spars impli object regard outlier point view similar fundament issu notion object outlier typic vari user problem domain dataset paper present robust solut detect high dimension outlier base user exampl toler incorrect input studi behavior project exampl discov object outstand project exampl outli experi real synthet dataset demonstr abil propos method detect outlier user exampl graph time densif law real graph evolv time normal growth pattern social technolog inform network studi discov pattern static graph identifi properti singl snapshot larg network small number snapshot includ heavi tail degre distribut commun small world phenomena lack inform network evolut long period hard convert find statement trend time studi 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 ftw fast similar search time warp distanc mine time seri data decad wit tremend growth interest applic deal queri mine time seri data numer represent method dimension reduct similar measur gear time seri introduc individu work introduc method made specif claim occasion theoret justif provid quantit experiment observ part compar aspect experi narrowli focus demonstr benefit propos method previous introduc order provid comprehens valid conduct extens set time seri experi implement represent method similar measur variant test effect time seri data set wide varieti applic domain paper give overview techniqu present compar experiment find effect experi provid unifi valid exist achiev case suggest claim literatur unduli optimist braid stream mine group lag correl research issu protein locat imag databas stream pattern discoveri multipl time dseri paper introduc spirit stream pattern discoveri multipl timeseri numer data stream valu observ time tick spirit increment find correl hidden variabl summaris kei trend entir stream collect compar studi featur vector dbase topic detect scheme compar studi featur vector dbase topic detect scheme editori tkde topic area revis fast estim fractal dimens correl integr stream data cloud point dimension space estim intrins dimension cloud set point dimension space straight line intrins fractal dimension integ fractal dimension appear pervas natur paper give fast method estim fractal dimension point data stream algorithm estim fractal dimens exist straightforward quadrat algorithm faster nlogn algorithm call box count method algorithm requir space ill suit semi infinit stream data paper propos algorithm base tug war idea comput fractal dimens singl pass dataset constant memori experiment result synthet real world data set demonstr effect algorithm categori subject descriptor numer analysi approxim wavelet fractal linear approxim special function approxim relev search 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 relev search find node connect irrelev node anomali detect propos algorithm comput relev score node random walk restart graph partit propos algorithm identifi anomali relev score evalu qualiti relev search base semant dataset measur perform anomali detect algorithm manual inject anomali effect effici method confirm experi real dataset relat fractal dimens perform vector quantiz segment motion captur data distinct behavior motion captur data anim commerci video game carefulli segment distinct motion time captur hand captur session move collect longer motion sequenc automat segment techniqu import process result reason time frame approxim tempor aggreg tempor aggreg queri retriev summar inform record time evolv attribut exist approach shortcom incur larg space requir ii high process cost iii base complex structur commerci system paper solv problem approxim techniqu bound error propos method base multivers tree logarithm worst case queri cost techniqu shelf rtree achiev perform expect case experiment demonstr propos method consum order magnitud space competitor significantli faster case permiss error bound small mmss multi dmodal stori dorient video summar propos multi modal stori orient video summar mmss unlik previou work fine tune domain specif heurist domainindepend graph base framework mmss uncov correl inform modal meaning stori orient new video summari mmss appli video retriev give perform match tradit retriev techniqu okapi lsi fine tune heurist tf idf introduct automat imag caption paper examin problem automat imag caption train set caption imag discov correl imag featur keyword automat find good keyword imag experi multipl design altern larg dataset content style propos method achiev rel improv caption accuraci state art fast discoveri connect subgraph recov latent time dseri observ sum network tomographi particl filter hidden variabl evolv time multipl set valuabl recov typic observ sum drive applic network tomographi estim origin destin od traffic flow determin commun local area network inform network engin manag solv problem design rout configur debug monitor price direct measur od traffic difficult imposs easili measur load link sum desir od flow paper propos filter method solv problem improv state art introduc explicit time depend realist gaussian margin statist model traffic flow attempt give experi real data filter scale linearli observ perform exist solut wide varieti set specif real network traffic measur cmu amp filter reduc estim error case automat multimedia cross dmodal correl discoveri categori subject 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 storag devic perform predict cart model storag devic perform predict kei element manag storag system applic plan task data assign work explor applic machin learn tool cart model storag devic model approach predict devic perform function input workload requir knowledg devic intern propos cart model predict request respons time deriv aggreg valu predict aggreg valu directli workload characterist train experiment platform provid accur black box model rang test trace real environ experi show model predict averag percentil respons time rel error low train workload similar test workload interpol workload gcap graph dbase automat imag caption abstract imag automat assign keywordsto paper propos graph base approach gcap outperform previous reportedmethod automat imag caption fast scale train test time lin ear data set size report auto caption experi amp quot standard amp quot corel imag databas mbyte gcap outperform recent success autocapt method percentag point cap tion accuraci rel improv introduct relat work huge imag databas assign content descript keyword imag automat ob outlier abstract outlier detect larg dataset import problem recent approach emploi reason definit outlier fundament issu notion object outlier typic vari user dataset paper present solut problem bring user loop ob outlier system knowledg user give exampl outlier directli incorpor small number exampl successfulli discov hidden concept spot object exhibit outlier ness exampl describ kei design decis algorithm build system demonstr real synthet dataset ob discov outlier match user intent autosplit fast scalabl discoveri hidden variabl stream multimedia databas 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 autom assist elicit user expect peopl softwar mundan task expect depend incomplet imprecis specif everydai softwar inhibit depend enhanc techniqu requir model proper behavior failur detect offer user center approach creat model proper behavior approach base satisfi user expect softwar behavior user reli demand perfect specif util data mine templat mechan user make expect precis result precis expect serv proxi miss specif detect unexpect data behavior concentr data feed continu stream data challeng everydai softwar method real world data feed hour detect problem data provid month detect independ problem surpris user domain expert previous analyz data feed systemat analysi support us method storag devic perform predict cart model storag devic perform predict kei element manag storag system applic plan task data assign work explor applic machin learn tool cart model storag devic model approach predict devic perform function input workload requir knowledg devic intern propos cart model predict request respons time deriv aggreg valu predict aggreg valu directli workload characterist train experiment platform provid accur black box model rang test trace real environ experi show model predict averag percentil respons time rel error low train workload similar test workload interpol workload predict index move object unknown motion pattern exist method predict spatio tempor databas assum object move linear function sever limit applic practic movement complex individu object follow drastic motion pattern order overcom problem introduc gener framework monitor index move object object comput individu function accur captur movement ii server index object locat coars level process queri filter refin mechan contribut recurs motion function support broad class linear motion pattern function presum priori movement postul motion object examin locat recent timestamp final propos effici index scheme facilit process predict queri fals miss index mine stream audit complianc hippocrat databas introduc audit framework determin databas system adher data disclosur polici user formul audit express sensit data subject disclosur review audit compon accept audit express return queri deem suspici access data execut overhead approach queri process small involv primarili log queri string minor annot databas trigger captur updat backlog databas time audit static analysi phase select subset log queri analysi queri combin transform sql audit queri run backlog databas identifi suspici queri effici precis describ algorithm data structur db base implement framework experiment result reinforc design choic show practic approach fractal dimens vector quantiz krishna kumaraswami center autom learn discoveri carnegi mellon univers skkumar cs cmu vasileio megalooikonom depart comput inform scienc templ univers vasili ci templ abstract determin perform vector quantiz method fast method determin perform vector quantiz method paper show relationship concept vector quantiz intrins fractal dimens deriv formula predict error rate vector quantiz method fractal dimens data set show result true synthet real data set discuss result faster data mine task adapt unsupervis stream mine proceed ninth acm sigkdd intern confer knowledg discoveri data mine problem identifi approxim duplic record databas essenti step data clean data integr process exist approach reli gener manual tune distanc metric estim similar potenti duplic paper present framework improv duplic detect trainabl measur textual similar propos emploi learnabl text distanc function databas field show measur capabl adapt specif notion similar field domain present learnabl text similar measur suitabl task extend variant learnabl string edit distanc vector space base measur emploi support vector machin svm train experiment result rang dataset show framework improv duplic detect accuraci tradit techniqu power dmethod comprehens estim techniqu multi ddimension queri exist estim approach multi dimension databas reli assumpt data distribut small region uniform seldom hold practic applic limit specif estim task distanc metric paper develop power method comprehens techniqu applic wide rang queri optim problem metric powermethod elimin local uniform assumpt accur scenario exist approach complet fail perform estim evalu simpl formula minim comput overhead extens experi confirm powermethod outperform previou techniqu term accuraci applic optim scenario gener data mine tool power law dsimilar graph loci fast outlier detect local correl integr outlier detect integr part data mine attract attent recent paper propos method evalu outlier call local correl integr loci previou method loci highli effect detect outlier group outlier microclust addit offer advantag novelti automat data dictat cutoff determin point outlier contrast previou method forc user pick cut off hint cut dataset provid loci plot point plot summar wealth inform data vicin point determin cluster micro cluster diamet inter cluster distanc exist outlier detect method match featur output singl number point outlier ness score loci method comput quickli previou method loci lead practic linear approxim method aloci approxim loci fast highli accur outlier detect knowledg work approxim comput speed outlier detect electr base extern similar categor attribut abstract similar distanc measur fundament critic properti data mine tool categor attribut abound databas car make gender occup field automobil insur databas inform sadli categor data easili amen similar comput domain expert manual similar relationship error prone feasibl attribut larg domain cross attribut similar gender occup extern similar function defin similar car make occur categor attribut exploit rich dualiti random walk graph electr circuit develop rep extern similar function rep theoret ground prior work ad hoc us rep shown experi cluster categor attribut valu show improv infer relationship rep effect nearest neighbour classifi gener data mine tool power law dsimilar graph 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 cross doutlier detect abstract problem outlier detect studi context domain receiv attent databas research commun knowledg work date focus exclus problem singl set observ space find deviat arous suspicion gener mechan domain set observ equival singl set class label assign observ astronom data label involv type galaxi spiral galaxi abnorm concentr ellipt galaxi neighborhood biodivers data label involv popul type patch speci popul food type diseas singl observ normal class entir set observ examin respect class arous suspicion paper problem set observ class label find arous suspicion take account class label variant signific practic import exist outlier detect approach extend case present practic approach deal problem demonstr perform real synthet dataset adapt hand doff stream mine sensor devic embed processor ubiquit measur monitor applic automat discoveri pattern trend larg volum data paramount import combin limit resourc cpu memori commun bandwidth power pose interest challeng power concis languag repres import featur data adapt handl arbitrari period compon includ burst requir memori singl pass data sensor automat discov interest pattern trend data perform outlier detect alert user sensor discov hourli phone call volum daili weekli period burst roughli year human recogn mother dai surg desir user issu explicit queri investig report pattern work propos awsom arbitrari window stream model method sensor oper remot hostil environ discov pattern effici power law dlevel internet topolog paper studi character topolog internet autonom system level show topolog effici power law eleg simplic powerlaw provid perspect seemingli uncontrol internet structur show power law consist year observ power law hold recent complet topolog correl coeffici degre power law addit studi evolut power law expon year interv observ variat degre base power law provid relationship expon topolog metric environment sensor network determin drink water qualiti secur larg dscale autom forecast fractal forecast attract lot research interest success method period time seri propos fast autom method linear forecast period chaotic time seri techniqu delai coordin embed paramet contribut autom set paramet concept intrins dimension oper system fast scalabl algorithm preprocess tree fast method forecast result work blackbox time seri input find paramet set gener predict system test real synthet data show system achiev low error handl arbitrarili larg dataset categori subject descriptor databas applic data mine time seri forecast geoplot spatial data mine video librari futur direct data mine stream index valu continu field databas videocub tool video mine classif videocub tool video mine classif data mine meet perform evalu fast algorithm model bursti traffic network web disk traffic bursti similar model adequ poisson arriv model type traffic gener realist trace obviou applic disk schedul network manag web server design previou model fraction brownian motion arfima captur bursti propos model requir paramet fit requir prohibit larg quadrat time gener larg trace propos simpl parsimoni method model solv problem requir paramet easili gener larg trace addit attract properti propos estim algorithm requir singl pass actual trace estim dai long disk trace millisecond mb data point requir minut model fit minut gener result synthet trace realist experi real disk web trace show synthet trace match real term queu behavior data visual rdbm make bit count fast nonlinear axi scale anf fast scalabl tool data mine massiv graph graph increasingli import data sourc import graph internet web familiar graph includ cad circuit phone record gene sequenc citi street social network academ citat kind relationship actor appear movi repres graph work present data mine tool call anf quickli answer number interest question graph repres data robust internet failur influenti databas paper gender differ movi appear pattern core anf base fast memori effici approach approxim complet neighbourhood function graph internet graph node anf highli accur approxim time faster exact comput reduc run time dai matter minut allow user perform ad hoc drill task repeatedli answer question chang data sourc enabl drill anf emploi techniqu approxim neighbourhood type function graph distinguish node edg compar exist approxim anf approach faster accur resourc addition unlik previou approach anf scale gracefulli handl disk resid graph final present result mine larg graph anf sensor data mine similar search pattern analysi data ddriven evolut data mine algorithm fast index visual metric data set slim dtree imagemap imag index method base spatial similar introduc imagemap method index similar search imag databas idb imagemap answer queri involv number object region take account inter relationship adopt gener imag content represent attribut relat graph arg conjunct accept arg edit distanc arg test imagemap real realist medic imag method visual dataset cluster data mine achiev fold speed search sequenti scan fals dismiss fundament schedul perform video tape librari robot tape librari popular applic high storag requir video server studi throughput tape librari system design schedul algorithm call relief compar older straightforward fcf maximum queue length mql unfair bypass roughli equival shortest job propos algorithm incorpor ag mechan order attain fair prove assumpt minim averag start latenc extens simul experi show relief outperform competitor fair unfair alik improv throughput reject ratio captur spatio dtempor behavior real traffic data traffic disk memori access typic exhibit bursti tempor local spatial local recent ground break work tempor model similar disk web traffic statist model gener realist seri time stamp work gener realist trace time locat block id kdd workshop report fractal dsimilar data mine issu approach similar search tear omni famili dpurpos access method videograph tool video mine classif paper introduc videograph tool video mine visual structur plot video sequenc main idea stitch similar scene time give fast algorithm stitch show case studi approach good featur classic accuraci result videograph reveal logic structur plot video clip dgx distribut mine massiv skew data skew distribut practic tradit zipf distribut fail model paper propos probabl distribut discret gaussian exponenti dgx achiev excel fit wide varieti set distribut includ zipf distribut special case present statist sound method estim dgx paramet base maximum likelihood estim mle appli dgx wide varieti real world data set sale data larg retail chain usag data amp internet clickstream data case dgx fit distribut correl coeffici quantil quantil plot algorithm scale requir singl pass data final illustr power dgx tool data mine task outlier detect keyword dgx zipf law rank frequenc plot frequenc count plot maximum likelihood estim lognorm distribut outlier detect tri dplot scalabl tool multidimension data mine focu problem find pattern larg multidimension dataset featur vector healthi healthi patient answer question cloud point separ smallest largest pair wise distanc dataset cloud point featur vector propos tool tri plot gener pq plot answer question provid set rule interpret tri plot appli rule synthet real dataset show tool classif tradit method nearest neighbor classif tree fail netcub scalabl tool fast data mine compress propos method comput store datacub idea bayesian network gener approxim count queri combin attribut valu don care bayesian network repres underli joint probabl distribut data gener mean network propos method netcub exploit correl attribut propos preprocess algorithm scale linearli size databas scalabl paralleliz straightforward parallel implement give algorithm estim count arbitrari queri fast constant databas size experiment result show netcub fast gener minut preprocess time record queri time achiev excel compress compress ratio real data low reconstruct error averag method natur visual data mine extra cost keyword datacub approxim count queri bayesian network bayesian network structur learn machin learn problem problem comput count record desir characterist databas common area decis support system data mine typic scenario custom analyst discov group custom exhibit interest unusu behavior lead possibl profit insight compani custom behavior word compani model custom base insight obtain model activ disk larg dscale data process leav suffici area includ mhz arm core similar embed microprocessor disk drive chip manufactur pursu processor asic technolog infineon siemen microelectron market chip call tricor includ mhz bit microcontrol mbyte chip ram custom specif logic disk function figur upper micron process cirru logic offer integr system chip hard disk drive control call ci includ mhz arm core gener promis mhz gener take larger system view tabl show detail larg databas system manag transact data mine workload trend ratio cpu versu aggreg process power remain roughli steadi compil data inform transact process perform council identifi web brows trend pattern multimedia queri relev feedback describ falcon system handl multimedia queri relev feedback falcon distinguish abil handl disjunct queri metric space experi show perform real synthet data term precis recal speed converg individu queri speed scalabl easili advantag shelf spatial metric access method analysi cluster properti hilbert space dfill curv abstract scheme linear map multidimension space propos applic access method spatio tempor databas imag compress applic desir properti linear map cluster mean local object multidimension space preserv linear space wide believ hilbert space fill curv achiev cluster paper analyz cluster properti hilbert space fill curv deriv close form formula number cluster queri region arbitrari shape polygon polyhedra asymptot solut gener case exact solut special case gener previou work agre empir result number cluster depend hypersurfac area queri region hypervolum show hilbert curv achiev cluster curv practic point view formula paper provid simpl measur predict requir disk access behavior total access time dimension curs dsimilar bless accur model region data spatial data numer applic gi multimedia tradit databas analysi focus point data typic uniform assumpt accur fractal distribut result exist point spatial data region island volum physic object real world problem solv paper base experiment evid real area volum follow power law name regal region area law show theoret implic model connect ubiquit fractal practic select estim rang queri experi varieti real dataset island lake human inhabit area show method extrem accur enjoi maximum rel error rang versu naiv model slim dtree high perform metric tree minim overlap node abstract paper present slim tree dynam tree organ metric dataset page fix size slim tree fat factor simpl quantifi degre overlap node metric tree degre overlap directli affect queri perform index structur suggest reduc overlap multidimension index structur slim tree metric structur explicitli design reduc degre overlap present algorithm insert object split node insert algorithm lead tree high storag util improv queri perform split algorithm run consider faster previou gener sacrif search perform result obtain experi real world data set show algorithm slim tree consist lead perform improv onlin data mine devolv time sequenc applic data interest compris multipl sequenc evolv time exampl includ currenc exchang rate network traffic data develop fast method analyz evolv time sequenc jointli estim forecast miss delai futur valu quantit data mine outlier detect method muscl adapt chang correl time sequenc handl indefinit long sequenc effici increment algorithm requir small amount storag oper make scale larg number sequenc present variat select muscl method propos effici algorithm reduc problem size experi real dataset show muscl outperform popular competitor predict accuraci time discov interest correl select muscl scale larg number sequenc reduc respons time time muscl improv predict qualiti deflat dimension curs multipl fractal dimens fast featur select fractal dimens dimensionalitycurs dimension reduct issu retain high interest data mine machin learn multimedia index cluster present fast scalabl algorithm quickli select import attribut dimens set dimension vector contrast older method method desir properti rotat attribut lead easi interpret result attribut spot attribut nonlinear correl requir constant number pass dataset good estim attribut idea fractal dimens dataset good approxim intrins dimens drop attribut affect appli method real synthet dataset gave fast good result introduct motiv whenmanag increas vo data mine oltp system free paper propos scheme schedul disk request take advantag abil high level function oper directli individu disk drive show scheme make support data mine workload oltp system free small impact throughput respons time exist workload specif show oltp system disk resourc consist provid sequenti bandwidth background data mine task close impact oltp throughput respons time high transact load low transact load show lower impact observ previou work mean product oltp system data mine task expens dedic system scheme take advantag close interact disk schedul read block data mine workload disk head pass satisfi spatial join select power law discov surpris law govern spatial join select set point spatial join find librari mile school law dictat number qualifi pair power law expon call pair count expon pc show law hold spatial join find school mile school addit gener case point set distinct law hold real dataset includ divers environ geograph dataset featur vector biologi data galaxi data astronomi addit introduc concept box occup product sum bop plot show comput pair count expon time manner reduc run time order magnitud quadrat linear due pair count expon analysi law achiev accur select estim constant densiti bias sampl improv method data mine cluster data mine larg data set requir sampl summar step form core represent data process effici uniform random sampl frequent practic frequent critic miss small cluster natur phenomena follow zipf distribut inabl uniform sampl find small cluster practic concern densiti bias sampl propos probabilist sampl dens region oversampl light region weight sampl preserv densiti origin data densiti bias sampl natur includ uniform sampl special case memori effici algorithm propos approxim densiti bias sampl singl scan data empir evalu densiti bias sampl synthet data set exhibit vari cluster size distribut find factor improv uniform search data mine visual multimedia data falcon feedback adapt loop content dbase retriev method exist perform simpl queri driven relev feedback larg multimedia databas method work vector space requir object repres vector featur space impli queri region typic convex research paper explain solut propos method design handl disjunct queri metric space user weight posit exampl system learn impli concept return similar object method differ exist relevancefeedback method base euclidean mahalanobi metric facilit learn disjunct concav model vector space arbitrari metric space addit method complet driven impos requir user aspect featur select main fast time sequenc index arbitrari lp norm analysi rang queri dspatial join queri real region dataset store dtree reminisc influenti paper quantifi data mine ratio rule associ rule mine algorithm oper data matrix custom theta product deriv associ rule agraw imielinski amp swami srikant amp agraw propos paradigm ratio rule quantifi measur good set discov rule propos guess error measur good root squar error reconstruct valu cell matrix pretend unknown contribut method guess miss hiddenvalu ratio rule method deriv bought milk bread rule guess amount spent butter unlik associ rule ratio rule perform varieti import task forecast answer scenario detect outlier visual data show comput ratio rule sing sigmod objectrank system perform author base keyword search databas inspir pagerank pagerank excel tool rank global import page web proven success googl googl fo complex rang queri region data store dtree paper studi node distribut tree store region data instanc island lake human inhabit area show real region dataset pack minimum bound rectangl mbr area distribut power law name regal region area law region mbr pack turn mbr law iter root tree base observ accur estim search effort rang queri promin spatial oper small number easi retriev paramet experi varieti real dataset island lake human inhabit area show estim accur enjoi maximum geometr averag rel error power dlaw relationship internet topolog appar random internet discov surprisingli simpl power law internet topolog power law hold snapshot internet novemb decemb growth size period show power law fit real data result correl coeffici higher observ provid perspect structur internet power law describ concis skew distribut graph properti node outdegre addit power law estim import paramet averag neighborhood size facilit design perform analysi protocol gener select realist topolog simul purpos introduct internet topolog properti don chang time year autom learn discoveri state dof dthe dart research topic rapidli grow field report summar conald meet place june carnegi mellon univers conald brought interdisciplinari group scientist concern decis make base data report organ part part page summar conald meet highlight main outcom individu workshop level part page summar result obtain individu workshop discuss depth promis research topic report web http www cs cmu conald conald support financi nation scienc foundat grant number ii gratefulli acknowledg financi support cald corpor member gratefulli acknowledg caterpillar exxon paramin giant eagl keybank mellon bank schlumberg cambridg research siemen corpor research vsi satech westinghous scie select estim window queri fact larg line segment dataset appear frequent numer applic involv spatial data gi multimedia tradit databas analysi estim select window queri pose spatial data import paramet queri optim focus point region data paper move signific step forward line segment dataset theoret analysi discov real line close follow distribut law name sled law segment length distribut sled law accur estim select window queri experi varieti real line segment dataset hydrograph system roadmap railroad util network show law hold formula extrem accur enjoi maximum rel error estim select leav dipartimento di imag index retriev base human perceptu color cluster propos imag retriev method base human perceptu cluster color imag color cluster produc imag small set repres color captur color properti imag small set sizabl contigu region captur spatial geometr properti imag propos method outperform tradit histogram improv method richer imag retriev capabl cover wider spectrum user requir power index scheme essenti cater larg scale imag databas effici retriev similar time sequenc time warp fast similar search larg time sequenc databas attract lot research interest euclidean distanc variat metric metric lead effici index featur extract keep dft coeffici subsequ fast spatial access method point featur space work examin popular field test dissimilar function time warp distanc function permit local acceler deceler rate signal sequenc function natur suitabl applic match voic audio medic signal electrocardiogram index viewpoint present major challeng lead natur featur preclud spatial access method quadrat len len length sequenc involv develop high dlevel represent video clip videotrail high level represent video clip compris inform physic semant structur provid process index retriev capabl video databas describ techniqu reduc sequenc mpeg encod video frame trail point low dimension space earlier work present techniqu applic paper describ techniqu extend higher dimens improv perform expect low dimension space cluster frame analyz transit cluster comput properti result trail portion trail classifi stationari transit lead high level descript video track interact cluster time lai groundwork complet analysi represent video physic semant structur keyword video represent mindread queri databas multipl exampl user easili express queri multimedia imag content set user photograph sunset current system qbic user give sampl queri rel import color shape textur wors user correl attribut tradit medic record databas medic research find mildli overweight patient impli queri weight height amp asymp lb inch goal provid user friendli theoret solid method handl queri user give exampl option good score propos method guess attribut import correl import weight contribut twofold formal problem minim problem show solv optim solut complet av ratio rule paradigm fast quantifi data mine associ rule mine algorithm oper data matrix custom theta product deriv associ rule propos paradigm ratio rule quantifi measur good set discov rule propos guess error measur good rootmean squar error reconstruct valu cell matrix pretend unknown contribut method guess miss hidden valu ratio rule method deriv bought milk bread rule guess amount spent butter perform varieti import task forecast answer scenario detect outlier visual data show comput ratio rule singl pass dataset small memori requir small matric activ storag larg dscale data mine multimedia increas perform decreas cost processor memori caus system intellig move peripher cpu storag system design trend excess comput power perform complex process optim insid storag devic date optim low level storag protocol time trend storag densiti mechan electron elimin bottleneck move data media put pressur interconnect host processor move data effici propos system call activ disk take advantag process power individu disk drive run applic level code move portion applic process execut directli disk drive dramat reduc data traffic advantag storag parallel present larg system todai discuss type applic dspacial join select estim fractal concept design access method bitempor databas fast effect retriev medic tumor shape abstract investig problem retriev similar shape larg databas focu medic tumor shape find tumor similar pattern natur similar function shape match base concept mathemat morpholog show lower bound set shape featur safe prune candid give fast correct output featur organ spatial access method lead fast index rang queri nearest neighbor queri addit lower bound contribut design fast algorithm nearest neighbor search achiev signific speedup provabl guarante correct experi demonstr roughli percent candid prune techniqu result time perform compar sequenti scan index term content base retriev multimedia index mathemat morpholog pattern spectrum overlai stripe optim parallel fo modern applic disk arrai system rapidli secondari storag media choic emerg applic larg storag high bandwidth requir stripe data disk disk arrai introduc signific perform benefit effect transfer rate secondari storag increas factor equal stripe width choic optim stripe width open problem gener formal analysi report intuit fail provid good guidelin result find occasion contradictori recommend literatur work contribut analyt calcul optim stripe width recogn optim stripe width sensit multiprogram level priori fluctuat time calcul optim stripe width practic reason propos str advanc databas system monet customiz databas system develop cwi univers amsterdam intend databas backend wide vari applic domain design maximum databas perform todai workstat multiprocessor system achiev consider success support data mine applic work project high end gi applic monet type algebra extens databas system emploi share memori parallel paper give goal motiv monet outlin architectur featur includ decompos storag model dsm emphasi bulk oper main virtual memori server custom case discuss issu build gi top monet monet handl larg data volum involv part work support sion grant videotrail repres visual structur video sequenc problem determin physic semant structur extend video sequenc essenti provid process index retriev capabl video databas paper describ techniqu reduc sequenc mpeg encod video frame trail point low dimension space space cluster frame analyz transit cluster comput properti result trail classifi portion trail stationari transit detect gradual edit shot track interact cluster time lai groundwork complet analysi represent video physic semant structur introduct recent advanc digit storag technolog comput perform led wide spread distribut video promot video valuabl inform resourc obtain real time effici support ad hoc queri larg dataset time sequenc ad hoc queri difficult larg dataset entir dataset disk compress decreas size dataset compress data notori difficult index access paper larg dataset compris multipl distinct time sequenc point sequenc numer show compress dataset format support ad hoc queri provid small error toler data uncompress experi larg real world dataset amp custom call pattern show propos method achiev averag error data compress mere origin space compress ratio number sensit dataset size experi aggreg queri achiev reconstruct error space requir introduct bulk data compress ddomain video index techniqu dct motion vector inform mpeg video develop multimedia applic hing avail fast effici storag brows index retriev techniqu video typic store effici compress format analyz compress represent directli avoid costli overhead decompress oper pixel level compress domain pars video present earlier work video clip divid shot subshot scene paper describ kei frame select featur extract index retriev techniqu directli applic mpeg compress video develop frame type independ represent type frame present mpeg video frame consid equival featur deriv dct macroblock motion vector inform map low dimension space access standard databas multimedia index web work databas system retriev multimedia object content extend work world wide web sourc storag multimedia object current text search engin textual inform system access type multimedia object content formid task improv constantli made index techniqu import step demonstr viabil techniqu lai groundwork larger capabl system implement simpl index scheme concentr build infrastructur support system system retriev refer imag www index imag store imag spatial access method queri find set imag www resembl queri imag due design easi includ addit context featur substitut multidimension access method tree grown recov inform summari data data store summar form histogram aggreg count sum averag rang studi estim origin detail data store summari formul task invers problem defin cost function optim constraint show formul includ uniform independ assumpt special case achiev reconstruct result maxim smooth oppos uniform experi real synthet dataset propos method consist outperform competitor improv root squar error cent stock price data cent smoother data set final show appli theori varieti databas problem involv partial inform olap data wareh histogram queri optim index multimedia data jersei data reduct report quick approxim answer larg databas lead data reduct di erent approach problem tradition pose solut data reduct problem paper describ evalu popular techniqu data reduct analysi ddimension quadtre decomposit arbitrari hyperectangl similar search medic imag databas relax uniform independ assumpt concept fractal dimens fast nearest neighbor search medic imag databas examin problem find similar tumor shape start natur similar function call max morpholog distanc show lower bound search nearest neighbor larg collect tumor shape specif state art concept morpholog pattern spectrum shape map shape point dimension space organ point tree show max norm space lower bound actual distanc guarante fals dismiss rang queri addit present nearest neighbor algorithm guarante fals dismiss final implement method test testb realist tumor shape establish tumor growth model murrai eden experi permiss copi fee part materi grant provid copi model skew distribut multifract law law analysi ddimension quadtre hausdorff fractal dimens mount evid man sch real dataset statist similar fractal import insight permit compact statist descript spatial dataset subsequ show form basi theoret analysi spatial access method typic unrealist uniform assumpt paper focu estim number quadtre block real spatial dataset requir hausdorff fractal dimens deriv close formula predict number quadtre block paramet formula predict space overhead respons time linear quadtre order om wide practic order verifi analyt model perform extens experiment investig real dataset come domain design implement qbism medic imag databas system experi pattern dmatch algorithm fastmap fast algorithm index promis idea fast search tradit multimedia databas map object point space featur extract function provid domain expert jag subsequ highli fine tune spatial access method sam answer type queri includ queri type translat rang queri pair queri translat spatial join bkss nearest neighbor match queri design featur extract function hard easier domain expert assess similar distanc object distanc inform obviou map object point topic paper describ fast algorithm map object point dimension space user defin di similar preserv benefit map index multimedia databas tutori access method bi dtempor databas estim select spatial queri correl fractal dimens examin estim select rang spatial join queri real spatial databas shown earlier fk real point set violat consist uniform independ assumpt fractal integ fractal dimens paper show infinit famili fractal dimens call correl dimens predict select spatial join main contribut real synthet point set averag number neighbor point point set power law expon immedi solv select estim spatial join bias rang queri queri center prefer area high point densiti present formula estim select bias queri includ integr constant shape ea fast search content multimedia databas grow number applic access video data store digit form secondari storag devic video demand multimedia messag result video server respons storag retriev fix rate hundr video disk increasingli import video data volumin disk order store video challeng devis scheme storag retriev video distribut workload evenli disk reduc cost server time provid good respons time client request video data paper present scheme base stripe video fine grain coars grain disk order effect util disk bandwidth scheme show optim cost server architectur determin data pre number video concurr retriev flexibl adapt buffer manag techniqu databas manag system problem buffer manag databas manag system concern effici main memori alloc manag answer databas queri previou work buffer alloc base exclus avail buffer runtim access pattern queri paper propos unifi approach buffer alloc consider account approach base notion margin gain expect reduct page fault alloc extra buffer queri extend approach support adapt buffer alloc adapt buffer alloc algorithm automat optim specif queri workload achiev adapt propos run time inform load system buffer alloc decis approach simpl queue model predict buffer alloc improv qbism extend dbm support medic imag paper refer adali subrahmanian amalgam knowledg base ii distribut mediat intern journal intellig cooper inform system aren knoblock shen queri reformul dynam inform integr wiederhold ed intellig integr inform kluwer academ publish arya codi faloutso richardson toga qbism extend dbm support medic imag proceed loth intern ieee confer data engin houston texa batini lenzerini navath compar analysi methodolog databas schema integr acm comput survei bayardo bohrer brice cichocki fowler helal kashyap ksiezyk martin nodin rashid rusinkiewicz shea unnikrishnan unmh woelk semant integr inform open dynam environ proceed acm intern confer manag data sigmod tucson uniform independ analysi dtree concept fractal dimens propos concept fractal dimens set point order quantifi deviat uniform distribut measur real data set road intersect counti star coordin nasa infrar ultraviolet explor provid evid real data skew show behav mathemat fractal measur integ fractal dimens arm tool show practic predict perform spatial access method specif tree provid analysi tree skew distribut point develop formula estim number disk access rang queri fractal dimens point set count experi real data set show formula accur rel error rare exce fractal dimens rep fast subsequ match time dseri databas present effici index method locat dimension subsequ witbin collect sequenc subsequ match queri pattern toler idea map data sequenc small set multidimension rectangl featur space rectangl readili index tradit spatial access method tree deteil slide window data sequenc extract featur result trail featur space propos effici effect algorithm divid trail trail subsequ repres minimum bound rectangl mbr examin queri vari length show handl case effici implement method carri experi synthet real data stock price movement compar method sequenti scan obviou competitor result excel method acceler search time time time quest project databas mine hilbert dtree improv dtree fractal propos tree structur outperform older heart idea facilit defer split approach tree propos order tree node order good sens group similar data rectangl minim area perimet result minimum bound rectangl mbr chosen call method sort rectangl hilbert center rectangl order node welldefin set sibl node defer split adjust split polici hilbert tree achiev high util desir contrari tree control space util typic achiev design manipul algorithm detail full implement hilbert tree experi show effici effect queri imag content qbic queri imag content project studi method queri larg line imag databas imag content basi queri exampl content includ color textur shape imag object region potenti applic includ medic give imag tumor textur photo journal give imag blue top red bottom art fashion catalog retail industri describ set featur similar measur allow queri color textur shape imag object demonstr effect qbic system normal precis recal experi test databas imag object popul commerci photo clip art imag imag airplan silhouett effici index featur specif addr tv dtree index structur high ddimension data propos file structur index high dimension data typic point featur space idea featur util addit featur addit discriminatori power absolut present detail design tree structur algorithm handl vari length featur vector final report simul result compar propos structur tree success method low dimension space result illustr superior method save disk access type contribut index structur high dimension featur space algorithm perform measur keyword spatial index similar retriev queri content introduct applic requir enhanc index capabl perform similar search tradit exot data type targ pack dtree propos tree pack techniqu static databas collect rectangl sort build tree bottom wai sort rectangl innov work fractal specif hilbert curv achiev order rectangl eventu pack propos implement variat perform experi synthet real data tiger file bureau censu win variat sort rectangl hilbert center variat consist outperform pack method roussopoulo leifker tree variant perform gain method increas skew data distribut specif highli skew tiger dataset achiev improv respons time older pack algorithm ov effici similar search sequenc databas propos index method time sequenc process similar queri discret fourier transform dft map time sequenc frequenc domain crucial observ sequenc practic interest frequenc strong import observ parsev theorem specifi fourier transform preserv euclidean distanc time frequenc domain map sequenc lowerdimension space fourier coeffici tree index sequenc effici answer similar queri provid experiment result show method superior search base sequenti scan experi show coeffici adequ provid good perform perform gain method increas number length sequenc sabbat dept comput scienc declust fractal propos method achiev declust cartesian product file unit focu rang queri oppos partial match queri older declust method examin method distanc preserv map hilbert curv impos linear order multidimension point bucket travers bucket order assign bucket disk round robin fashion good distanc preserv properti hilbert curv end result disk bucket linear order address space goal declust experi show intuit argument lead good perform propos method perform older declust scheme categori subject descriptor data structur file data base manag physic de qbic project queri imag content color textur shape bonel mitchel editor machin learn volum margan kaufmann publish niblack wayn barber equitz flickner glasman petkov yanker faloutso taubin qbic project queri imag content color textur shape storag retriev imag video databas spie volum shannon weaver mathemat theori commun univers illinoi press urbana ill refer qbism prototyp dd medic imag databas system paper automat semi automat warp algorithm extrem import applic precis technolog permit anatom structur base access acquir medic imag comparison studi patient long warp atla enabl databas grow queryabl time consum manual segment data high perform dtree hybrid index organ text databas parallel dtree problem exploit parallel acceler perform spatial access method specif tree goal design server spatial data maxim throughput rang queri achiev maxim parallel larg rang queri engag disk point queri propos simpl hardwar architectur consist processor disk attach architectur propos distribut node tradit tree cross disk pointer multiplex tree tree code ident singl disk tree addit decid disk newli creat tree node store propos examin criteria choos disk node success term proxim index pi estim similar node tree node dtree indic skew distribut case set valu tree construct skew distribut present geometr growth techniqu manag post record case show perform techniqu straightforward fix length post list guarante disk access search take fraction space competitor requir experi signatur file paper deal estim signatur weight gener superimpos code techniqu adopt multipl signatur file estim need perform evalu organ inform retriev applic simpl formula probabl densiti function expect varianc signatur weight present present formula deriv gener methodolog call transform approach previou work result simpl method author deriv densiti function equival result shown frame dslice signatur file paper provid case purpos studi perform ifd convent sequenti signatur file ssf method vertic partit environ multifram signatur file mfsf method introduc recent studi purpos develop test environ implement ssf mfsf method extend method ifd test perform real data experi show ifd improv perform inspect method reduc observ fd queri respons time experi similar result involv gener frame slice signatur file approach report kocberb organ paper section convent fd estim method propos fd estim method ifd explain section explain test environ experi section appli ifd ssf mfsf method measur perform improv obtain ifd experiment real data section conclus appendix provid formal proof show condit number fals drop record fd estim averag number term record equal fd estim individu valu record analyt result quadtre decomposit arbitrari rectangl dot spatial access method fractal exist databas manag system dbmss handl effici multi dimension data box polygon point multi dimension space examin access method data design goal mind effici term search speed space overhead abil integr dbm easili propos method map multidimension object point dimension space tradit primari kei access method appli extens part dbm propos map base fractal implement method top tree map simul experi distribut input data show flexibl buffer alloc base margin gain predict load control flexibl buffer alloc disk alloc method error correct code architectur high perform engin inform system incorpor string search hypertext system user interfac signatur file design issu hypertext system provid appeal mechan inform brows databas travers select link fact find situat string search effect complement brows paper describ applic signatur file method achiev rapid conveni string search small person comput hypertext environ method implement prototyp commerci product perform data search time storag space present commerci hypertext databas user interfac issu discuss experi string search interfac sucessfulli novic user address correspond christo faloutso comput scienc depart univers maryland colleg park md usa tel fax email christo cs umd introduct earli exploratori hypertext system give numer commerci analysi end block wast space incorpor string search hypertext system user interfac signatur file design issu hypertext system provid appeal mechan inform brows databas travers select link fact find situat string search effect complement brows paper describ applic signatur file method achiev rapid conveni string search small person comput hypertext environ method implement prototyp commerci product perform data search time storag space present commerci hypertext databas user interfac issu discuss experi string search interfac sucessfulli novic user address correspond christo faloutso comput scienc depart univers maryland colleg park md usa tel fax email christo cs umd introduct earli exploratori hypertext system give numer commerci signatur dbase text retriev method survei gure interpret method int confer fractal secondari kei retriev paper propos fractal hilbert curv order design good distanc preserv map map improv perform secondari kei spatial access method multi dimension point store dimension medium disk good cluster reduc number disk access retriev improv respons time experi rang queri nearest neighbor queri show propos hilbert curv achiev cluster older method bit shuffl peano curv situat categori subject descriptor databas manag physic design access method inform storag retriev content analysi index index method gener term algorithm design perform addit keyword distanc preserv map geometr data introduct work propos space fill curv achiev declust error correct code perform issu binari relationship model tri dcell da data structur spatial object fast text access method optic larg magnet disk design perform comparison signatur file integr access method text attribut expert databas system effici support engin environ grai code partial match rang queri effici pictori databas system psql comput vision offer number techniqu context pictori inform system motiv point view present basic issu involv comput vision multimedia inform system emphas differ domain outlin contribut comput vision make develop effici multimedia inform system handl queri visual qve part paper concentr matter archiv retriev imag larg databas case qve approach propos reli concept relev focusof attent mechan locat select pertin inform imag imag archiv stage automat determin part imag provid meaning indic compil imag access tabl relev focu analysi object orient spatial access method dtree dynam index multi ddimension object problem index multidimension object consid classif exist method discuss major issu involv multidimension data index variat guttman tree tree avoid overlap rectangl intermedi node tree introduc algorithm search updat initi pack reorgan structur discuss detail final provid analyt result indic tree achiev save disk access compar tree search file thousand rectangl univers maryland system research center univers maryland institut advanc comput studi umiac research sponsor partiali nation scienc foundat grant cdr introduct recogn past exist databas manag system dbmss optim signatur extract inform loss descript perform analysi signatur file method offic file multiattribut hash grai code timo selli nick roussopoulo christo faloutso tree dynam index design perform consider optic disk dbase multimedia object server signatur file design perform comparison signatur extract method design signatur file method account duniform occurr queri frequenc access method text signatur file access method document analyt perform evalu design consider messag file server multimedia offic file system perform consider messag file 