gaussian mixtur model local consist gaussian mixtur model gmm popular data cluster method view linear combin gaussian compon gmm cluster obei gaussian distribut task cluster group observ compon estim cluster paramet expect maxim algorithm involv estim problem previou studi shown natur occur data resid close underli submanifold paper case probabl distribut support submanifold ambient space account smooth condit probabl distribut geodes data manifold observ close intrins geometri distribut gaussian compon similar simpli speak introduc method base manifold structur data cluster call local consist gaussian mixtur model lcgmm specif construct nearest neighbor graph adopt kullback leibler diverg distanc measur regular object function gmm experi data set demonstr effect regular laplacian dhash term document unsupervis featur select multi dcluster data dtaught hash fast similar search document recommend social tag servic dtaught hash fast similar search semi dsupervis topic model imag annot propos techniqu semi supervis imag annot introduc harmon regular base graph laplacian data probabilist semant model learn latent topic imag probabilist semant model connect visual featur textual annot imag latent topic incorpor manifold assumpt model probabl latent topic imag drawn manifold imag share similar visual featur annot probabl distribut latent topic similar creat nearest neighbor graph model manifold propos regular em algorithm simultan learn gener model assign probabl densiti latent topic imag discrimin databas label imag annot previou work probabilist dyadic data analysi local global consist dyadic data aris real world applic social network analysi inform retriev order discov underli hidden structur dyadic data topic model techniqu propos typic algorithm includ probabilist latent semant analysi plsa latent dirichlet alloc lda probabl densiti function obtain algorithm support euclidean space previou studi shown natur occur data resid close underli submanifold introduc probabilist framework model topic geometr structur dyadic data explicitli take account local manifold structur specif local manifold structur model graph graph laplacian analog laplac beltrami oper manifold appli smooth probabl densiti function result obtain probabilist distribut concentr data manifold experiment result real data set demonstr effect propos approach local preserv nonneg matrix factor matrix factor techniqu frequent appli inform process task neg matrix factor nmf receiv consider attent due psycholog physiolog interpret natur occur data represent part base human brain hand geometr perspect data sampl low dimension manifold embed high dimension ambient space hope find compact represent uncov hidden topic simultan respect intrins geometr structur paper propos algorithm call local preserv neg matrix factor lpnmf purpos data point kl diverg evalu similar hidden topic optim map obtain featur valu hidden topic restrict neg vari smoothli geodes data manifold empir studi show encourag result propos algorithm comparison state ofth art algorithm larg high dimension databas spars project graph recent studi shown canon algorithm princip compon analysi pca linear discrimin analysi lda obtain graph base dimension reduct framework algorithm yield project map linear combin origin featur result difficult interpret psycholog physiolog paper present techniqu learn spars project graph data reduc subspac repres linear combin subset relev featur compar pca lda result obtain spars project easier interpret algorithm base graph embed model encod discrimin geometr structur term data affin embed result obtain appli regular regress learn set spars basi function specif norm regular lasso spars project effici comput experiment result document databas demonstr effect method model hidden topic document manifold train linear discrimin analysi linear time abstract linear discrimin analysi lda popular method extract featur preserv class separ wide field inform process machin learn data mine inform retriev pattern recognit comput lda involv dens matric eigen decomposit computation expens time memori specif lda mnt time complex requir mn mt nt memori number sampl number featur min larg infeas appli lda paper propos algorithm discrimin analysi call spectral regress discrimin analysi srda spectral graph analysi srda cast discrimin analysi regress framework facilit effici comput regular techniqu theoret analysi show srda comput ms time ms memori averag number featur sampl extens experiment result real world data set demonstr effect effici algorithm dneg matrix factor manifold recent neg matrix factor nmf receiv lot attent inform retriev comput vision pattern recognit nmf aim find neg matric product approxim origin matrix size matric smaller origin matrix result compress version origin data matrix solut nmf yield natur part base represent data nmf appli data represent major disadvantag fail geometr structur data paper develop graph base approach part base data represent order overcom limit construct affin graph encod geometr inform seek matrix factor respect graph structur demonstr success algorithm appli real world problem topic model network regular paper formal defin problem topic model network structur tmn propos solut problem regular statist topic model harmon regular base graph structur data propos method combin topic model social network analysi leverag power statist topic model discret regular output model summar topic text map topic network discov topic commun instanti topic model graph base regular model appli wide rang text mine problem authortop analysi commun discoveri spatial text mine empir experi data set genr show approach effect outperform text orient method network orient method propos model gener appli text collect mixtur topic network structur srda effici algorithm larg dscale discrimin analysi abstract linear discrimin analysi lda popular method extract featur preserv class separ project function lda commonli obtain maxim class covari simultan minim class covari wide field inform process machin learn data mine inform retriev pattern recognit comput lda involv dens matric eigendecomposit computation expens time memori specif lda mnt time complex requir mn mt nt memori number sampl number featur min larg infeas appli lda paper propos algorithm discrimin analysi call spectral regress discrimin analysi srda spectral graph analysi srda cast discrimin analysi regress framework facilit effici comput regular techniqu specif srda solv set regular squar problem eigenvector comput involv huge save time memori theoret analysi show srda comput ms time ms memori averag number nonzero featur sampl extens experiment result real world data set demonstr effect effici algorithm index term linear discrimin analysi spectral regress dimension reduct learn maximum margin subspac imag retriev abstract fundament problem content base imag retriev cbir gap low level visual featur high level semant concept narrow gap relev feedback introduc imag retriev user provid inform classifi learn distinguish posit neg exampl real world applic number user feedback small compar dimension imag space order cope high dimension propos semisupervis method dimension reduct call maximum margin project mmp mmp aim maxim margin posit neg exampl local neighborhood tradit dimension reduct algorithm princip compon analysi pca linear discrimin analysi lda effect global euclidean structur mmp design discov local manifold structur mmp suitabl imag retriev nearest neighbor search involv project imag lower dimension subspac relev imag closer queri imag retriev perform enhanc experiment result corel imag databas demonstr effect propos algorithm index term multimedia inform system imag retriev relev feedback dimension reduct isometr project recent problem dimension reduct receiv lot interest field inform process includ data mine inform retriev pattern recognit case data sampl low dimension manifold embed high dimension euclidean space popular manifold learn algorithm includ local linear embed isomap laplacian eigenmap algorithm nonlinear provid embed result train sampl paper propos linear dimension reduct algorithm call isometr project isometr project construct weight data graph weight discret approxim geodes distanc data manifold linear subspac obtain preserv pairwis distanc algorithm perform origin space reproduc kernel hilbert space lead kernel isometr project isometr project defin compar princip compon analysi pca wide data process algorithm capabl discov intrins geometr structur special pca optim data space linear algorithm assumpt handl complex data space present experiment result algorithm appli synthet data set real life data exampl illustr effect propos method spectral regress unifi subspac learn framework content dbase imag retriev relev feedback establish effect framework narrow gap low level visual featur high level semant concept content base imag retriev tradit implement relev feedback distanc metric classifi learn user provid neg posit exampl due limit user feedback high dimension featur space confront issu curs dimension recent research consid manifold wai address issu local preserv project augment relat embed semant subspac project paper techniqu spectral graph embed regress propos unifi framework call spectral regress learn imag subspac framework facilit analysi differ connect algorithm mention crucial faster comput make retriev system capabl respond user queri effici spectral regress unifi subspac learn framework content dbase imag retriev relev feedback establish effect framework narrow gap low level visual featur high level semant concept content base imag retriev tradit implement relev feedback distanc metric classifi learn user provid neg posit exampl due limit user feedback high dimension featur space confront issu curs dimension recent research consid manifold wai address issu local preserv project augment relat embed semant subspac project paper techniqu spectral graph embed regress propos unifi framework call spectral regress learn imag subspac framework facilit analysi differ connect algorithm mention crucial faster comput make retriev system capabl respond user queri effici learn spatial smooth subspac face recognit subspac learn base face recognit method attract consider interest recent year includ semi dsupervis discrimin analysi linear discrimin analysi lda popular method extract featur preserv class separ project vector commonli obtain maxim class covari simultan minim class covari practic suffici train sampl covari matrix class accur estim paper propos method call semisupervis discrimin analysi sda make label unlabel sampl label data point maxim separ class unlabel data point estim intrins geometr structur data specif aim learn discrimin function smooth data manifold experiment result singl train imag face recognit relev feedback imag retriev demonstr effect algorithm spectral regress effici regular subspac learn subspac learn base face recognit method attract consider interest recent year includ princip compon analysi pca linear discrimin analysi lda local preserv project lpp neighborhood preserv embed npe margin fisher analysi mfa disadvantag approach comput involv eigendecomposit dens matric expens time memori paper propos dimension reduct framework call spectral regress sr effici regular subspac learn sr cast problem learn project function regress framework avoid eigen decomposit dens matric regress base framework kind regular natur incorpor algorithm make flexibl comput analysi show sr linear time complex huge speed compar cubic time complex ordinari approach experiment result face recognit demonstr effect effici method effici kernel discrimin analysi spectral regress linear discrimin analysi lda popular method extract featur preserv class separ project vector commonli obtain maxim class covari simultan minim class covari lda perform origin input space reproduc kernel hilbert space rkh data point map lead kernel discrimin analysi kda data highli nonlinear distribut kda achiev perform lda comput project function kda involv eigen decomposit kernel matrix expens larg number train sampl exist paper present algorithm kernel discrimin analysi call spectral regress kernel discrimin analysi srkda spectral graph analysi srkda cast discrimin analysi regress framework facilit effici comput regular techniqu specif srkda solv set regular regress problem eigenvector comput involv huge save comput cost comput analysi show srkda time faster ordinari kda formul make easi develop increment version algorithm fulli util comput result exist train sampl experi face recognit demonstr effect effici propos algorithm spectral regress unifi approach spars subspac learn recent problem dimension reduct subspac learn receiv lot interest field inform process includ data mine inform retriev pattern recognit popular method includ princip compon analysi pca linear discrimin analysi lda local preserv project lpp disadvantag approach learn project function linear combin origin featur difficult interpret result paper propos dimension reduct framework call unifi spars subspac learn ussl learn spars project ussl cast problem learn project function regress framework facilit kind regular norm regular lasso spars project effici comput experiment result real world classif cluster problem demonstr effect method local sensit discrimin analysi linear discrimin analysi lda popular data analyt tool studi class relationship data point major disadvantag lda fail discov local geometr structur data manifold paper introduc linear algorithm discrimin analysi call local sensit discrimin analysi lsda suffici train sampl local structur gener import global structur discrimin analysi discov local manifold structur lsda find project maxim margin data point class local area specif data point map subspac nearbi point label close nearbi point label experi carri standard face databas show clear improv result lda base recognit regular regress imag manifold retriev recent consider interest geometr base method imag retriev method imag space smooth manifold appli manifold learn techniqu find euclidean embed euclidean distanc embed space approxim geodes distanc manifold main advantag method relev feedback retriev natur incorpor system prior inform paper retriev problem classif problem manifold learn distanc measur aim learn classif function imag manifold effici kei issu imag retriev web scale propos approach imag retriev manifold approach base regular linear regress framework local manifold structur user provid relev feedback incorpor imag retriev system local preserv regular extens experi carri larg imag databas demonstr effici effect propos approach laplacian optim design imag retriev relev feedback power techniqu enhanc content base imag retriev cbir perform solicit user relev judgment retriev imag return cbir system user label learn classifi distinguish relev irrelev imag top return imag inform challeng determin unlabel imag inform improv classifi label train sampl paper propos activ learn algorithm call laplacian optim design lod relev feedback imag retriev algorithm base regress model minim squar error measur label imag simultan preserv local geometr structur imag space specif assum imag suffici close measur label close construct nearest neighbor graph geometr structur imag space graph laplacian discuss result field optim experiment design guid select subset imag amount inform experiment result corel databas suggest propos approach achiev higher precis relev feedback imag retriev categori subject descriptor inform storag retriev inform search retriev relev feedback mathemat cluster search www imag link page layout analysi due rapid growth number digit imag web increas demand effect effici method organ retriev imag paper describ ifind system cluster ing search www imag vision base page segment algorithm web page partit block textual link inform imag accur extract block imag textual inform imag represent extract page block block imag block page relationship link structur page layout analysi construct imag graph method sensit noisi link previou method picashow imag graph reflect semant relationship imag notion markov chain infer semant de gree imag imagerank character import imag imagerank bine relev score produc final rank imag search graph model techniqu spectral graph theori imag cluster embed visual experi mental result million imag download web provid paper tensor space model document analysi vector space model vsm core inform retriev past decad vsm consid document vector high dimension space vector space techniqu latent semant index lsi support vector machin svm naiv bay appli index classif case dimension document space extrem larg make techniqu infeas due curs dimension paper propos tensor space model document analysi repres document order tensor matric correspondingli index algorithm call tensor latent semant index tensorlsi develop tensor space theoret analysi show tensorlsi computation effici convent latent semant index make applic extrem larg scale data set experiment result standard document data set demonstr effici effect algorithm categori subject descriptor orthogon laplacianfac face recognit patrascu buzuloiu imag dynam rang enhanc imag cluster tensor represent neighborhood preserv embed recent lot interest geometr motiv approach data analysi high dimension space case data drawn sampl probabl distribut support submanifold euclidean space paper propos subspac learn algorithm call neighborhood preserv embed npe princip compon analysi pca aim preserv global euclidean structur npe aim preserv local neighborhood structur data manifold npe sensit outlier pca compar recent propos manifold learn algorithm isomap local linear embed npe defin train data point npe conduct origin space reproduc kernel hilbert space data point map rise kernel npe experi face databas demonstr effect algorithm statist comput analysi local preserv project laplacian score featur select supervis learn scenario featur select studi wide literatur select featur unsupervis learn scenario harder problem due absenc class label guid search relev inform previou unsupervis featur select method wrapper techniqu requir learn algorithm evalu candid featur subset paper propos filter method featur select independ learn algorithm method perform supervis unsupervis fashion propos method base observ real world classif problem data class close import featur evalu power local preserv laplacian score compar method data varianc unsupervis fisher score supervis data set experiment result demonstr effect effici algorithm tensor subspac analysi previou work demonstr imag variat object human face variabl light effect model low dimension linear space typic linear subspac learn algorithm includ princip compon analysi pca linear discrimin analysi lda local preserv project lpp method imag high dimension vector imag repres plane intrins matrix paper propos algorithm call tensor subspac analysi tsa tsa consid imag order tensor vector space relationship column vector imag matrix row vector natur character tsa tsa detect intrins local geometr structur tensor space learn lower dimension tensor subspac compar propos approach pca lda lpp method standard databas experiment result demonstr tsa achiev recognit rate effici commun mine multi drelat network abstract social network analysi attract attent recent year commun mine major direct social network analysi exist method commun mine assum kind relat network mine result independ user prefer realiti exist multipl heterogen social network repres kind relationship kind relationship plai distinct role task paper systemat analyz problem mine hidden commun heterogen social network base observ relat import respect queri propos method learn optim linear combin relat meet user expect obtain relat perform achiev commun mine orthogon local preserv index problem document index represent recent local preserv index lpi propos learn compact document subspac latent semant index optim sens global euclidean structur lpi optim sens local manifold structur lpi extrem sensit number dimens make difficult estim intrins dimension inaccur estim dimension drastic degrad perform reason lead problem lpi orthogon orthogon distort metric structur document space paper propos algorithm call orthogon lpi orthogon lpi iter comput mutual orthogon basi function respect local geometr structur empir studi show olpi local preserv power lpi compar algorithm lsi lpi extens experiment result show orthogon lpi obtain perform lsi lpi crucial insensit number dimens make effici data preprocess method text cluster classif retriev document cluster local preserv index abstract propos document cluster method aim cluster document semant class document space gener high dimension cluster high dimension space infeas due curs dimension local preserv index lpi document project lower dimension semant space document relat semant close previou document cluster method base latent semant index lsi nonneg matrix factor nmf method discov geometr discrimin structur document space theoret analysi method show lpi unsupervis approxim supervis linear discrimin analysi lda method intuit motiv method extens experiment evalu perform reuter tdt data set index term document cluster local preserv index dimension reduct semant local preserv cluster imag databas import challeng make grow imag repositori easi search brows imag cluster techniqu help wai includ imag data preprocess user interfac design search result represent spectral cluster method promis cluster method year cluster data complex structur global optimum guarante exist spectral cluster algorithm normal cut difficult handl data point train set paper propos cluster algorithm name local preserv cluster lpc share data represent properti nonlinear spectral method lpc explicit map function defin train data point test point experiment result show lpc accur direct kmean pca kmean show lpc produc gener compar result normal cut effici normal cut hierarch cluster www imag search result visual problem cluster web imag search result gener imag search result return imag search engin multipl topic organ result semant cluster facilit user brows paper propos hierarch cluster method visual textual link analysi vision base page segment algorithm web page partit block textual link inform imag accur extract block imag block level link analysi techniqu imag graph construct appli spectral techniqu find euclidean embed imag respect graph structur imag kind represent visual featur base represent textual featur base represent graph base represent spectral cluster techniqu cluster search result semant cluster imag search illustr potenti techniqu organ www imag base analysi page layout web link structur due rapid growth number digit imag web increas demand effect effici method organ retriev imag paper describ method cluster embed www imag vision base page segment algorithm web page partit block textual link inform imag accur extract block imag extract page block block toimag block page relationship link structur page layout analysi construct imag graph imag graph model techniqu spectral graph theori imag cluster embed experiment result paper block dlevel link analysi link analysi shown great potenti improv perform web search pagerank hit popular algorithm exist link analysi algorithm treat web page singl node web graph case web page multipl semant web page consid atom node paper web page partit block visionbas page segment algorithm extract page toblock block page relationship link structur page layout analysi construct semant graph www node repres singl semant topic graph describ semant structur web base block level link analysi propos algorithm block level pagerank block level hit perform studi extens web data local preserv index document represent document represent index kei problem document analysi process cluster classif retriev convention latent semant index lsi consid effect deriv index lsi essenti detect repres featur document represent discrimin featur lsi optim discrimin document semant paper algorithm call local preserv index lpi propos document index document repres vector low dimension contrast lsi discov global structur document space lpi discov local structur obtain compact document represent subspac detect essenti semant structur compar propos lpi approach lsi standard databas experiment result show lpi represent sens semant structur block dbase web search multipl topic vari length web page neg factor significantli affect perform web search paper explor page segment algorithm partit web page block investig advantag block level evid improv retriev perform web context special characterist web page page segment method impact web search perform compar type method includ fix length page segment dom base page segment vision base page segment combin method integr semant fix length properti experi block level queri expans retriev perform approach combin method achiev perform web search experiment result show semant partit web page effect deal problem multipl drift topic mix length great potenti boost perform current web search engin local preserv index document represent document represent index kei problem document analysi process cluster classif retriev convention latent semant index lsi consid effect deriv index lsi essenti detect repres featur document represent discrimin featur lsi optim discrimin document semant paper algorithm call local preserv index lpi propos document index document repres vector low dimension contrast lsi discov global structur document space lpi discov local structur obtain compact document represent subspac detect essenti semant structur compar propos lpi approach lsi standard databas experiment result show lpi represent sens semant structur extract content structur web page base visual represent abstract web content structur base visual represent propos paper web applic inform retriev inform extract automat page adapt benefit structur paper present automat top tag tree independ approach detect web content structur simul user understand web layout structur base visual percept compar exist techniqu approach independ underli document represent html work html structur layout structur experi show satisfactori result microsoft research asia web track trec improv pseudo drelev feedback web inform retriev web page segment contrast tradit document retriev web page good inform unit search multipl topic lot irrelev inform navig decor interact part page paper propos vision base page segment vip algorithm detect semant content structur web page compar simpl dom base segment method page segment scheme util visual cue obtain partit page semant level vip algorithm assist select queri expans term pseudo relev feedback web inform retriev achiev perform improv web track dataset 