network anomali detect base eigen equat compress mine abnorm pattern heterogen time dseri irrelev featur fault event detect address issu detect fault event multivari time seri suppos realist situat featur multivari time seri correspond heterogen rel larg number normal exampl small number exampl fault event advanc featur irrelev fault event includ situat requir real time high accuraci process propos algorithm resolv issu kei idea includ transform time seri featur sequenc anomali score order map heterogen featur homogen featur anomali score degre anomali rel ordin sequenc repres pattern fault event term anomali score vector select featur fault event mean iter optim normal fault anomali score vector monitor degre abnorm regard test anomali score vector match abnorm pattern demonstr effect propos algorithm applic actual automobil fault diagnosi data set latent variabl mine applic anomal behavior detect paper propos approach anomali detect latent variabl space make step latent anomali detect convent approach anomali detect concern track data larg deviat ordinari pattern paper concern issu track happen latent variabl space consist meta inform exist observ data case masquerad detect convent task detect anomal command line relat masquerad malici behavior attempt track behavior pattern write mail make softwar inform abstract level command line kei idea method construct model variat vector introduc rel latent variabl space latent anomali detect reduc issu chang point detect time seri model variat vector form demonstr experiment result artifici data set unix command data set method significantli enhanc accuraci exist anomali detect method latent variabl mine applic anomal behavior detect paper propos approach anomali detect latent variabl space make step latent anomali detect convent approach anomali detect concern track data larg deviat ordinari pattern paper concern issu track happen latent variabl space consist meta inform exist observ data case masquerad detect convent task detect anomal command line relat masquerad malici behavior attempt track behavior pattern write mail make softwar inform abstract level command line kei idea method construct model variat vector introduc rel latent variabl space latent anomali detect reduc issu chang point detect time seri model variat vector form demonstr experiment result artifici data set unix command data set method significantli enhanc accuraci exist anomali detect method dynam model select applic novelti detect unifi framework detect outlier chang point time seri dynam syslog mine network failur monitor syslog monitor technolog recent receiv vast attent area network manag network monitor address wide rang import issu includ network failur symptom detect event correl discoveri syslog intrins dynam sens form time seri behavior chang time paper propos methodolog dynam syslog mine order detect failur symptom higher confid discov sequenti alarm pattern comput devic kei idea dynam syslog mine repres syslog behavior mixtur hidden markov model adapt learn model line discount learn algorithm combin dynam select optim number mixtur compon give anomali score univers test statist dynam optim threshold real syslog data demonstr valid methodolog scenario failur symptom detect emerg pattern identif correl discoveri track dynam topic trend finit mixtur model dline unsupervis outlier detect finit mixtur discount learn algorithm outlier detect fundament issu data mine speci calli fraud detect network intrus detect network monitor smartsift abbrevi ss outlier detect engin adrress problem viewpoint statist learn theori paper theoret basi ss empir demonstr effect ss detect outlier onlin process line unsupervis learn probabilist model finit mixtur model inform sourc time datum input ss emploi line discount learn algorithm learn probabilist model score datum base learn model high score indic high possibl statist outlier featur ss adapt stationari sourc data score clear statist inform theoret mean computation inexpens handl categor continu variabl experiment applic network intrus detect show ss identifi data high score correspond attack low comput cost experiment applic identifi number meaning rare case actual health insur patholog data australia health insur commiss distribut cooper mine inform consortia topic analysi finit mixtur model address issu topic analysi determin text topic structur represent indic topic includ text topic chang text topic analysi consist main task topic identif text segment reflect topic topic analysi extrem varieti text process applic previou studi suifici address mine product reput web know reput competitor product import market custom relationship manag costli collect analyz survei data manual paper present framework mine product reput internet automat collect peopl opinion target product web page text mine techniqu obtain reput product advanc gener basi human test exampl syntact linguist rule determin statement opinion posit neg natur opinion collect statement target product gener search engin rule extract opinion attach opinion label unifi framework detect outlier chang point dstationari time seri data concern vith issu outlier detect chang point detect data stream area data mine increas interest issu relat fraud detect rare event discoveri vhile relat event trend chang detect activ monitor specif import situat data sourc stationari natur data sourc chang time real applic previou work outlier detect chang point detect relat explicitli paper present unifi frame vork deal vith basi theori line learn stationari time seri framevork probabilist model data sourc inerement learn line discount learn algorithm track chang data sourc adapl forget effect past data gradual score data calcul measur deviat learn model vith higher score indic high possibl outlier chang point data stream detect appli score method time seri move averag loss predict learn model specif ve develop effici algorithm line discount learn auto regress model time seri data demonstr valid framework simul experiment applic stock market data analysi mine open answer questionnair data text classif esc dbase stochast decis list propos method text classif stochast decis list stochast decis list order sequenc rule method view rule base method text classif advantag readabl refin acquir knowledg method uniqu decis list automat construct basi principl minim extend stochast complex esc construct decis list fewer error classif accuraci classif achiev method appear compar exist rule base method empir demonstr rule base method result high classif accuraci categori text assign specif text tend short empir verifi advantag rule base method rule base discov outlier filter rule unlabel data combin supervis learner unsupervis learner mine open answer questionnair data dline unsupervis outlier detect finit mixtur discount learn algorithm outlier detect fundament issu data mine speci calli fraud detect network intrus detect network monitor smartsift abbrevi ss outlier detect engin adrress problem viewpoint statist learn theori paper theoret basi ss empir demonstr effect ss detect outlier onlin process line unsupervis learn probabilist model finit mixtur model inform sourc time datum input ss emploi line discount learn algorithm learn probabilist model score datum base learn model high score indic high possibl statist outlier featur ss adapt stationari sourc data score clear statist inform theoret mean computation inexpens handl categor continu variabl experiment applic network intrus detect show ss identifi data high score correspond attack low comput cost experiment applic identifi number meaning rare case actual health insur patholog data australia health insur commiss extend stochast complex minimax rel loss analysi concern problem sequenti predict hypothesi class continu predict strategi effect perform measur minimax rel cumul loss rcl minimum worst case differ cumul loss predict algorithm assign hypothesi class purpos paper evalu minimax rcl gener continu hypothesi class gener loss deriv asymptot upper lower bound minimax rcl show match ln error ln dimens paramet hypothesi class sampl size constant depend loss function show cumul loss attain minimax rcl asymptot coincid extend stochast complex esc extens rissanen stochast complex sc decis text classif esc dbase stochast decis list propos method text classif stochast decis list stochast decis list order sequenc rule method view rule base method text classif advantag readabl refin acquir knowledg method uniqu decis list automat construct basi principl minim extend stochast complex esc construct decis list fewer error classif accuraci classif achiev method appear compar exist rule base method empir demonstr rule base method result high classif accuraci categori text assign specif text tend short empir verifi advantag rule base method rule base distribut cooper bayesian learn strategi minimax rel loss analysi sequenti predict algorithm parametr hypothes decis dtheoret extens stochast complex applic learn document classif finit mixtur model propos method classifi document categori defin categori finit mixtur model base soft cluster word treat problem classifi document conduct statist hypothesi test finit mixtur model emploi em algorithm effici estim paramet finit mixtur model exper iment result method outperform exist method distribut cooper bayesian learn strategi document classif finit mixtur model propos method classifi document categori defin categori finit mixtur model base soft cluster word treat problem classifi document conduct statist hypothesi test finit mixtur model emploi em algorithm effici estim paramet finit mixtur model exper iment result method outperform exist method dline maximum likelihood predict respect gener loss function random approxim mdl stochast model hidden variabl random approxim aggreg strategi applic predict discrimin dline maximum likelihood predict respect gener loss function alpha dhelix region predict stochast rule learn loss bound model dline stochast predict algorithm discrimin learn minimum dcomplex algorithm applic learn dparametr rule polynomi dtime discrimin learnabl protein secondari structur predict base stochast drule learn discrimin learn learn criterion stochast rule wk weiss kulikowski comput system learn morgan kaufmann san mateo ca yam kenji yamanishi learn criterion stochast rule proceed workshop comput learn theori page publish morgan kaufmann pol david pollard converg stochast process springer verlag pol david pollard rate uniform converg empir process index unbound class function manuscript pol david pollard empir process theori applic volum nsf cbm region confer seri probabl statist institut math stat stat assoc qui adolfo quiroz metric entropi learnabl unpublish manuscript universidad sim bol ivar caraca venezuela ren renyi probabl theori loss bound model dline stochast predict strategi learn stochast motif genet sequenc 