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 discrimin codeword select imag represent music recommend unifi hypergraph combin social media inform music content dwai min dmax cut imag cluster junk imag filter googl imag unsupervis featur select multi dcluster data document recommend social tag servic dasa dissatisfact dorient advertis base sentiment analysi laplacian regular doptim design activ learn applic imag retriev abstract increasingli case interest comput vision pattern recognit confront situat data size larg label expens challeng determin unlabel sampl inform improv classifi label train sampl problem activ learn regress model context experiment design classic optim experiment design approach base squar error measur sampl fail account unmeasur sampl paper propos activ learn algorithm oper graph algorithm base graph laplacian regular regress model simultan minim squar error measur sampl preserv local geometr structur data space construct nearest neighbor graph geometr structur data space graph laplacian discuss result field optim experiment design guid select subset data point amount inform experi demonstr superior perform comparison convent algorithm index term activ learn experiment design imag retriev regular approxim harmon project theoret analysi algorithm convex experiment design manifold structur imag retriev content base imag retriev cbir activ research area comput scienc relev feedback cbir system bridg semant gap typic user ask make relev judgement queri result feedback inform rank imag databas effect relev feedback algorithm provid user inform imag respect rank function paper propos activ learn algorithm call convex laplacian regular optim design claprid relev feedback imag retriev algorithm base regress model minim squar error label imag simultan preserv intrins geometr structur imag space select inform imag minim averag predict varianc optim problem claprid cast semidefinit program sdp problem solv interior point method experiment result corel databas demonstr effect propos algorithm relev feedback imag retriev categori subject descriptor inform storag retriev inform search retriev relev feedback mathemat comput probabl statist experiment design 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 unifi activ semi dsupervis learn framework imag compress problem lossi imag compress machin learn perspect typic imag compress algorithm transform imag spatial domain represent frequenc domain represent transform techniqu discret cosin transform discret wavelet transform code transform valu recent perform frequenc transform machin learn base approach propos color inform repres pixel learn model predict color rest pixel select repres pixel essenti activ learn problem color semi supervis learn problem paper propos activ learn algorithm call graph regular experiment design gred share principl semisupervis learn algorithm color activ semi supervis learn unifi singl framework pixel select color experiment result suggest propos approach achiev higher compress ratio imag qualiti compress time significantli reduc 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 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 graph embed constraint recent graph base dimension reduct receiv lot interest field inform process central graph structur model geometr discrimin structur data manifold label inform incorpor graph structur modifi weight data point paper propos dimension reduct algorithm call constrain graph embed consid label inform addit constraint specif constrain space solut explor embed result consist label experiment result real life data set illustr effect propos method new articl extract templat dindepend wrapper problem templat independ new extract state art new extract method base templat level wrapper induct limit correctli extract page belong unseen templat wrapper templat gener costli maintain date wrapper hundr websit chang templat lead invalid wrapper paper formal new extract machin learn problem learn templat independ wrapper small number label new page singl site featur dedic new titl bodi develop correl new titl new bodi exploit templat independ wrapper extract new page site templat experi wrapper learn page singl new site achiev accuraci new page new site advertis keyword gener activ learn tag dorient document summar social annot web document highli gener descript topic contain page tag frequenc user attent degre make annot good resourc summar multipl topic web page paper present tag orient web document summar approach document content tag annot document improv summar perform tag rank algorithm name eigentag propos paper reduc nois tag associ mine techniqu emploi expand tag set tackl sparsiti problem experiment result show tag orient summar signific improv tag probabilist question recommend question answer commun user interact question answer qa commun yahoo answer grow popular qa site thousand question post daili difficult user find question interest delai answer question rise question recommend techniqu user locat interest question paper adopt probabilist latent semant analysi plsa model question recommend propos metric evalu perform approach experiment result show recommend approach effect categori subject descriptor imag retriev nonlinear manifold embed 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 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 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 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 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 local sensit semi dsupervis featur select regular queri classif search click inform hundr million user dai submit queri web search engin user queri typic short make queri understand challeng problem paper propos approach queri represent classif submit queri web search engin queri repres set term found web page return search engin queri consid point high dimension space standard classif algorithm regress appli tradit regress flexibl situat larg number highli correl predictor variabl suffer overfit problem search click inform semant relationship queri incorpor learn system regular specif function minim empir loss label queri select preserv semant relationship queri present experiment evid suggest regular regress algorithm search click inform effect queri classif 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 regular local preserv index spectral regress problem document index represent recent local preserv index lpi propos learn compact document subspac latent semant index lsi optim sens global euclidean structur lpi optim sens local manifold structur lpi effici time memori make difficult appli larg data set specif comput lpi involv eigen decomposit dens matric expens paper propos algorithm call regular local preserv index rlpi benefit recent progress spectral graph analysi cast origin lpi algorithm regress framework enabl avoid eigen decomposit dens matric regress base framework kind regular natur incorpor algorithm make flexibl extens experiment result show rlpi obtain similar result compar lpi significantli faster make effici effect data preprocess method larg scale text cluster classif retriev 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 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 queri rewrit activ learn sponsor search sponsor search major revenu sourc search compani web searcher issu queri advertis keyword limit queri rewrit techniqu effect match user queri relev advertis keyword increas amount web advertis match relev critic click studi aim improv queri rewrit relev purpos activ learn algorithm call transduct experiment design select inform sampl train queri rewrit relev model experi show approach improv model accuraci rewrit relev 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 semi dsupervis support vector learn face recognit 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 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 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 face recognit laplacianfac abstract propos appear base face recognit method call laplacianfac approach local preserv project lpp face imag map face subspac analysi princip compon analysi pca linear discrimin analysi lda effect euclidean structur face space lpp find embed preserv local inform obtain face subspac detect essenti face manifold structur laplacianfac optim linear approxim eigenfunct laplac beltrami oper face manifold unwant variat result light facial express pose elimin reduc theoret analysi show pca lda lpp obtain graph model compar propos laplacianfac approach eigenfac fisherfac method face data set experiment result suggest propos laplacianfac approach represent achiev lower error rate face recognit index term face recognit princip compon analysi linear discrimin analysi local preserv project face manifold subspac learn imag retriev base increment subspac learn learn imag manifold retriev problem learn map function low level featur space high level semant space assumpt data lie submanifold embed high dimension euclidean space propos relev feedback scheme natur conduct imag manifold question total ambient space imag typic repres featur vector natur distanc distanc induc ambient space geodes distanc manifold measur similar imag number data point small hard discov intrins manifold structur base user interact relev feedback driven queri system intrins similar imag accur estim develop algorithm framework approxim optim map function radial basi function rbf neural network semant imag infer rbf neural network experiment result show approach effect improv perform content base imag retriev system increment semi dsupervis subspac learn imag retriev subspac learn techniqu widespread pattern recognit research includ princip compon analysi pca local preserv project lpp techniqu gener unsupervis model data absenc label categori relev feedback driven imag retriev system user provid inform describ intrins semant relationship imag paper propos semi supervis subspac learn algorithm increment learn adapt subspac preserv semant structur imag space base user interact relev feedback driven queri byexampl system algorithm capabl accumul knowledg user result featur represent imag databas system futur retriev perform enhanc experi larg collect imag shown effect effici propos algorithm 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 imag retriev dimension reduct 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 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 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 bayesian shape local face recognit global local textur face recognit problem face year lot research research requir order end robust face recognit system potenti close human perform face recognit system fr report high perform level achiev bayesian shape local face recognit global local textur face recognit problem face year lot research research requir order end robust face recognit system potenti close human perform face recognit system fr report high perform level achiev local pursuit embed exist nonlinear local method dimension reduct yield impress result data embed manifold visual open problem defin unifi project data embed subspac construct train sampl think global fit local present linear embed approach call local embed analysi lea dimension reduct featur select lea view linear approxim local linear embed lle solv linear eigenspac problem close form lea automat learn local neighborhood characterist discov compact linear subspac optim preserv intrins manifold structur highdimension data point lea find low dimension point subspac linear project embed process concentr adjac data point dens cluster effect discrimin analysi supervis classif unsupervis cluster test propos lea algorithm benchmark databas experiment result show lea data represent effici dimension reduct classic linear method know tree forest art imag retriev societi profil paper aim address problem art imag retriev air aim user find favorit paint imag air great interest applic potenti interest research challeng retriev base paint content style heavili base user prefer profil paper describ collabor ensembl learn statist learn approach task appli probabilist support vector machin svm model individu user profil base exampl dislik paint due high complex profil model svm weak predict prefer paint overcom problem combin societi user profil repres respect svm model predict user prefer paint imag demonstr combin scheme embed bayesian framework retain intuit interpret mind user share similar prefer report extens empir studi base experiment set includ control simul perform paint imag set report evalu base user prefer collect onlin web base survei experi demonstr propos approach achiev excel perform term captur user divers prefer local preserv project problem inform process involv form dimension reduct paper introduc local preserv project lpp linear project map aris solv variat problem optim preserv neighborhood structur data set lpp altern princip compon analysi pca classic linear techniqu project data direct maxim varianc high dimension data li low dimension manifold embed ambient space local preserv project obtain find optim linear approxim eigenfunct laplac beltrami oper manifold result lpp share data represent properti nonlinear techniqu laplacian eigenmap local linear embed lpp linear crucial defin ambient space train data point born illustr exampl high dimension data set learn local preserv subspac visual recognit previou work demonstr face recognit perform improv significantli low dimension linear subspac convention princip compon analysi pca linear discrimin analysi lda consid effect deriv face subspac effect euclidean structur face space paper propos approach map face imag subspac obtain local preserv project lpp face analysi call laplacianfac approach pca lda lpp find embed preserv local inform obtain face space detect essenti manifold structur unwant variat result light facial express pose elimin reduc compar propos laplacianfac approach eigenfac fisherfac method test dataset experiment result show propos laplacianfac approach represent achiev lower error rate face recognit local preserv project problem inform process involv form dimension reduct paper introduc local preserv project lpp linear project map aris solv variat problem optim preserv neighborhood structur data set lpp altern princip compon analysi pca classic linear techniqu project data direct maxim varianc high dimension data li low dimension manifold embed ambient space local preserv project obtain find optim linear approxim eigenfunct laplac beltrami oper manifold result lpp share data represent properti nonlinear techniqu laplacian eigenmap local linear embed lpp linear crucial defin ambient space train data point born illustr exampl high dimension data set learn semant space user relev feedback imag retriev learn infer semant space user relev feedback imag retriev 