sparsiti dcogniz overlap dcluster behavior infer social network residu bayesian dcluster matrix approxim simpl time reexamin probabilist topic detect model semi dsupervis learn user dprefer travel schedul present graph base semi supervis approach learn user prefer travel schedul assum set user small number label travel schedul classifi schedul desir desir task trivial small number label point complic fact schedul compris multipl compon aspect natur instanc case arriv time model probabl distribut account uncertainti aspect wait time featur vector aspect thought type observ schedul exist label propag approach exploit vast amount unlabel data handl multi aspect data propos multi aspect label propag malp approach extend label propag handl multipl type observ approxim algorithm tensor cluster present knowledg approxim algorithm tensor cluster power gener basic cluster tensor increasingli common modern applic deal complex heterogen data cluster fundament tool data analysi pattern discoveri akin cousin common tensor cluster formul np hard optim unlik case approxim algorithm address imbal build recent cluster work deriv tensor cluster algorithm approxim guarante allow metric diverg bregman object function therewith answer open question anagnostopoulo al analysi yield constant approxim factor independ data size worst case show factor tight euclidean cluster empir approxim factor observ conserv method practic discrimin mix dmembership model bayesian overlap subspac cluster bayesian cluster ensembl anomali detect survei anomali detect import problem research divers research area applic domain anomali detect techniqu specif develop applic domain gener survei provid structur comprehens overview research anomali detect group exist techniqu categori base underli approach adopt techniqu categori identifi kei assumpt techniqu differenti normal anomal behavior appli techniqu domain assumpt guidelin assess effect techniqu domain categori provid basic anomali detect techniqu show exist techniqu categori variant basic techniqu templat easier succinct understand techniqu belong categori categori identifi advantag disadvantag techniqu categori provid discuss comput complex techniqu import issu real applic domain hope survei provid understand di erent direct research topic techniqu develop area appli domain intend begin anomali detect manifold embed applic transport corridor bayesian dcluster multipl mixtur model overlap cluster fo scalabl bregman dcluster approxim algorithm bregman dcluster tensor cluster latent dirichlet condit naiv dbay model analysi logist model exponenti famili connect onlin perform multi dwai cluster relat graph number real world domain social network commerc involv heterogen data describ relat multipl class entiti understand natur structur type heterogen relat data essenti exploratori analysi perform predict model task paper propos principl multi cluster framework relat data type entiti simultan cluster base intrins attribut valu multipl relat entiti achiev introduc relat graph model describ relat entiti class relat set entiti class repres form multi modal tensor domain multi cluster formul driven object captur maxim inform origin relat graph accur approxim set tensor relat formul applic bregman diverg broad famili loss function includ squar euclidean distanc kl diverg permit analysi mix data type convex combin bregman loss function present larg famili structur multi cluster scheme preserv linear summari statist origin data accomplish gener extend recent propos kei theoret result minimum bregman inform principl relat graph set describ effici multi cluster algorithm base altern minim gener number recent propos cluster method empir result dataset obtain real world domain movi recommend newsgroup articl demonstr gener efficaci framework topic model text stream studi batch onlin unsupervis learn bayesian bound show import bayesian bound studi machin learn batch onlin set aris applic simpl compress lemma deriv pac bayesian bound batch set ii bayesian log loss bound iii bayesian bound loss bound onlin set compress lemma set semant prior posterior loss show core bound argument paper simplifi understand import appar dispar result bring light power tool develop similar argument method scalabl cluster algorithm balanc constraint abstract cluster method data mine problem extrem scalabl addit data mine applic demand cluster obtain balanc approxim size import paper propos gener framework scalabl balanc cluster data cluster process broken step sampl small repres subset point cluster sampl data popul initi cluster remain data refin show simpl uniform sampl origin data suffici repres subset high probabl propos framework larg class algorithm cluster sampl set focu popular parametr algorithm eas exposit present algorithm popul refin cluster algorithm popul cluster base gener stabl marriag problem refin algorithm constrain iter reloc scheme complex method kn log obtain balanc cluster data point compar favor exist techniqu balanc cluster addit provid balanc guarante cluster perform obtain propos framework compar unconstrain solut experiment result dataset includ cluster base approach perceptu imag hash model dbase overlap cluster vast major cluster algorithm partit real world dataset inher overlap cluster approach find overlap cluster work analysi biolog dataset paper interpret overlap cluster model propos segal al gener gaussian mixtur model extend overlap cluster model base mixtur regular exponenti famili distribut bregman diverg provid algorithm modif extens present result synthet data subset newsgroup eachmovi dataset optim condit expect bregman predictor cluster unit hyperspher von mise dfisher distribut larg scale data mine applic text categor gene express analysi involv high dimension data inher direct natur data normal li surfac unit hyperspher popular model mixtur multi variat gaussian inadequ character data paper propos gener mixtur model approach cluster direct data base von mise fisher vmf distribut aris natur data distribut unit hyperspher deriv analyz variant expect maxim em framework estim concentr paramet mixtur numer estim concentr paramet trivial high dimens involv function invers ratio bessel function formul cluster algorithm variant em deriv approach theoret basi cosin similar wide emploi inform retriev commun obtain spheric kmean algorithm kmean cosin similar special case variant empir result cluster high dimension text gene express data base mixtur vmf distribut cluster bregman diverg inform theoret analysi maximum likelihood mixtur estim exponenti famili import task unsupervis learn maximum likelihood mixtur estim mlme exponenti famili paper prove mathemat equival mlme problem rate distort problem bregman diverg present theoret result rate distort theori bregman diverg analysi problem trade compress preserv inform present yield inform bottleneck method interest special case gener maximum entropi approach bregman dcluster matrix approxim cluster power data mine techniqu vari applic text cluster microarrai analysi recommend system recent informationtheoret cluster approach applic empir joint probabl distribut propos situat cluster gener matric desir paper present substanti gener cluster framework bregman diverg object function condit expect base constraint consid base statist preserv analysi coclust problem lead minimum bregman inform principl gener maximum entropi principl yield eleg meta algorithm guarante achiev local optim methodolog yield algorithm encompass previous cluster cluster algorithm base altern minim object evalu criterion cluster propos test object criterion evalu cluster perform cluster algorithm run unlabel data aid classif algorithm accuraci quantifi pac mdl bound semisupervis set cluster algorithm natur separ data hidden label small number cluster perform simpl extens argument lead object model select method experiment result text analysi dataset demonstr approach empir result competit bound test set perform natur dataset activ semi dsupervis pairwis constrain cluster semi supervis cluster small amount supervis data aid unsupervis learn typic approach specifi limit number link cannotlink constraint pair exampl paper present pairwis constrain cluster framework method activ select inform pairwis constraint improv cluster perform cluster activ learn method easili scalabl larg dataset handl high dimension data experiment theoret result confirm activ queri pairwis constraint significantli improv accuraci cluster small amount supervis cluster bregman diverg gener model dbase cluster direct data semi dsupervis cluster seed semi supervis cluster small amount label data aid bia cluster unlabel data paper explor label data gener initi seed cluster constraint gener label data guid cluster process introduc semi supervis variant kmean cluster view instanc em algorithm label data prior inform condit distribut hidden categori label experiment result demonstr advantag method standard random seed cop kmean previous develop semi supervis cluster algorithm scale balanc cluster algorithm amp quot rand amp quot page 