Stability Region based Expectation Maximization for Model-based Clustering

Size: px
Start display at page:

Download "Stability Region based Expectation Maximization for Model-based Clustering"

Transcription

1 Stablty Regon based Expectaton Maxmzaton for Model-based Clusterng Chandan K. Reddy, Hsao-Dong Chang School of Electrcal and Computer Engneerng, Cornell Unversty, Ithaca, NY Bala Rajaratnam Department of Statstcal Scence Cornell Unversty, Ithaca, NY Abstract In spte of the ntalzaton problem, the Expectaton- Maxmzaton (EM) algorthm s wdely used for estmatng the parameters n several data mnng related tasks. Most popular model-based clusterng technques mght yeld poor clusters f the parameters are not ntalzed properly. To reduce the senstvty of ntal ponts, a novel algorthm for learnng mxture models from multvarate data s ntroduced n ths paper. The proposed algorthm takes advantage of TRUST-TECH (TRansformaton Under STabltyreTanng Equlbra CHaracterzaton) to compute neghborhood local maxma on lkelhood surface usng stablty regons. Bascally, our method coalesces the advantages of the tradtonal EM wth that of the dynamc and geometrc characterstcs of the stablty regons of the correspondng nonlnear dynamcal system of the log-lkelhood functon. Two phases namely, the EM phase and the stablty regon phase, are repeated alternatvely n the parameter space to acheve mprovements n the maxmum lkelhood. Though appled to Gaussan mxtures n ths paper, our technque can be easly generalzed to any other parametrc fnte mxture model. The algorthm has been tested on both synthetc and real datasets and the mprovements n the performance compared to other approaches are demonstrated. The robustness wth respect to ntalzaton s also llustrated expermentally. 1 Introducton Fnte mxtures allow a probablstc model-based approach to unsupervsed learnng [10] whch plays an mportant role n predctve data mnng applcatons. One of the most popular methods used for fttng mxture models to the observed data s the Expectaton-Maxmzaton (EM) algorthm whch converges to the maxmum lkelhood estmate of the mxture parameters locally [4, 6]. The usual steepest Correspondng author : emal - ckr6@cornell.edu descent, conjugate gradent, or Newton-Raphson methods are too complcated for use n solvng ths problem [19]. EM has become a popular method snce t takes advantages of problem specfc propertes. EM based approaches have been successfully used to solve problems that arse n varous other applcatons [12, 2]. In ths paper, we consder the problem of learnng parameters of Gaussan Mxture Models (GMM). Fg 1 shows data generated by three Gaussan components wth dfferent mean and varance. Note that every data pont has a probablstc (or soft) membershp that gves the probablty wth whch t belongs to each of the components. Ponts that belong to component 1 wll have hgh probablty of membershp for component 1. On the other hand, data ponts belongng to components 2 and 3 are not well separated. The problem of learnng mxture models nvolves not only estmatng the parameters of these components but also fndng the probabltes wth whch each data pont belongs to these components. Gven the number of components and an ntal set of parameters, EM algorthm can be appled to compute the optmal estmates of the parameters that maxmze the lkelhood of the data gven the estmates of these components. However, the man problem wth the EM algorthm s that t s a greedy method whch s very senstve to the gven ntal set of parameters. To overcome ths problem, a novel two phase algorthm based on stablty regon analyss s proposed. The man research concerns that motvated the new algorthm presented n ths paper are : EM algorthm for mxture modelng converges to a local maxmum of the lkelhood functon very quckly. There are many other promsng local optmal solutons n the close vcnty of the solutons obtaned from the methods that provde good ntal guesses of the soluton. Model selecton crtera usually assumes that the global optmal soluton of the log-lkelhood functon can be obtaned. However, achevng ths s computatonally ntractable.

2 2 Relevant Background Fgure 1. Data generated by three Gaussan components. The problem of learnng mxture models s to obtan the parameters of these Gaussan components and the membershp probabltes of each datapont. Some regons n the search space do not contan any promsng solutons. The promsng and nonpromsng regons coexst and t becomes challengng to avod wastng computatonal resources to search n non-promsng regons. Of all the concerns mentoned above, the fact that most of the local maxma are not dstrbuted unformly [16] makes t mportant for us to develop algorthms that not only help us to avod searchng n the low-lkelhood regons but also emphasze the mportance of explorng promsng subspaces more thoroughly. Ths subspace search wll also be useful for makng the soluton less senstve to the ntal set of parameters. In ths paper, we propose a novel two phase algorthm for estmatng the parameters of mxture models. Usng concepts of dynamcal systems and EM algorthm smultaneously to explot the problem specfc features of the mxture models, our algorthm obtans the optmal set of parameters by searchng for the global maxmum on the lkelhood surface n a systematc manner. The rest of ths paper s organzed as follows: Secton 2 gves some relevant background about varous methods proposed n the lterature for solvng the problem of learnng mxture models. Secton 3 dscusses some prelmnares about mxture models, EM algorthm and stablty regons. Secton 4 dscusses our new framework and the detals of our mplementaton are gven n Secton 5. Secton 6 shows the expermental results of our algorthm on synthetc and real datasets. Fnally, Secton 7 concludes our dscusson wth future research drectons. Although EM and ts varants have been extensvely used for learnng mxture models, several researchers have approached the problem by dentfyng new technques that gve good ntalzaton. More generc technques lke determnstc annealng [16], genetc algorthms [13] have been appled to obtan a good set of parameters. Though, these technques have asymptotc guarantees, they are very tme consumng and hence cannot be used for most of the practcal applcatons. Some problem specfc algorthms lke splt and merge EM [17], component-wse EM [6], greedy learnng [18], ncremental verson for sparse representatons[11], parameter space grd [8] are also proposed n the lterature. Some of these algorthms are ether computatonally very expensve or nfeasble when learnng mxtures n hgh dmensonal spaces [8]. Inspte of all the expense n these methods, very lttle effort has been taken to explore promsng subspaces wthn the larger parameter space. Most of these algorthms eventually apply the EM algorthm to move to a locally maxmal set of parameters on the lkelhood surface. Smpler practcal approaches lke runnng EM from several random ntalzatons, and then choosng the fnal estmate that leads to the local maxmum wth hgher value of the lkelhood are also successful to certan extent [15]. Though some of these methods apply other addtonal mechansms (lke perturbatons [5]) to escape out of the local optmal solutons, systematc methods are yet to be developed for searchng the subspace. The dynamcal system of the log-lkelhood functon reveals more nformaton on the neghborhood stablty regons and ther correspondng local maxma [3]. Hence, the dffcultes of fndng good solutons when the error surface s very rugged can be overcome by addng stablty regon based mechansms to escape out of the convergence zone of the local maxma. Though ths method mght ntroduce some addtonal cost, one has to realze that exstng approaches are much more expensve due to ther stochastc nature. Specfcally, for a problem n ths context, where there s a non-unform dstrbuton of local maxma, t s dffcult for most of the methods to search neghborng regons [20]. For ths reason, t s more desrable to apply TRUST-TECH based Expectaton Maxmzaton (TT-EM) algorthm after obtanng some pont n a promsng regon. The man advantages of the proposed algorthm are that t: Explores most of the neghborhood local optmal solutons unlke the tradtonal stochastc algorthms. Acts as a flexble nterface between the EM algorthm and other global methods. Allows the user to work wth exstng clusters obtaned from the tradtonal approaches and mproves the qual-

3 ty of the solutons based on the maxmum lkelhood crtera. Helps the expensve global methods to truncate early. Explots the fact that promsng solutons are obtaned by faster convergence of the EM algorthm. 3 Prelmnares We now ntroduce some necessary prelmnares on mxture models, EM algorthm and stablty regons. Frst, we descrbe the notaton used n the rest of the paper: Table 1. Descrpton of the Notatons used n the paper Notaton Descrpton d number of features n number of data ponts k number of components s total number of parameters Θ parameter set θ parameters of th component α mxng weghts for th component X observed data Z mssng data Y complete data t tmestep for the estmates 3.1 Mxture Models Lets assume that there are k Gaussans n the mxture model. The form of the probablty densty functon s as follows: p(x Θ) = k α p(x θ ) (1) =1 where x = [x 1,x 2,..., x d ] T s the feature vector of d dmensons. The α k s represent the mxng weghts. Θ represents the parameter set (α 1,α 2,...α k,θ 1,θ 2,...θ k )andp s a unvarate Gaussan densty parameterzed by θ (.e. µ and σ ): 1 p(x θ )= e (x µ ) (2π)σ 2 2σ 2 (2) Also, t should be notced that beng probabltes α must satsfy 0 α 1, =1,.., k, and k α =1 (3) =1 Gven a set of n..d samples X = {x (1),x (2),.., x (n) }, the log-lkelhood correspondng to a mxture s log p(x Θ) = log = n log p(x (j) Θ) k α p(x (j) θ ) =1 The goal of learnng mxture models s to obtan the parameters Θ from a set of n data ponts whch are the samples of a dstrbuton wth densty gven by (1). The Maxmum Lkelhood Estmate (MLE) s gven by : Θ MLE = arg max { log p(x Θ) } (5) Θ where Θ ndcates the entre parameter space. Snce, ths MLE cannot be found analytcally for mxture models, one has to rely on teratve procedures that can fnd the global maxmum of log p(x Θ). The EM algorthm descrbed n the next secton has been used successfully to fnd the local maxmum of such a functon [9]. 3.2 Expectaton Maxmzaton The EM algorthm assumes X to be observed data. The mssng part, termed as hdden data, s a set of n labels Z = {z (1), z (2),.., z (n) } assocated wth n samples, ndcatng whch component produced each sample [9]. Each label z (j) = [z (j),.., z(j) ] s a bnary vector where z (j) =1and z (j) m 1,z(j) 2 k (4) =0 m, means the sample x (j) was produced by the th component. Now, the complete log-lkelhood (.e. the one from whch we would estmate Θ f the complete data Y = {X, Z}s log p(x, Z Θ) = log p(y Θ) = k log k =1 =1 [ α p(x (j) θ )] z(j) z (j) log [ α p(x (j) θ )] (6) The EM algorthm produces a sequence of estmates { Θ(t),t =0, 1, 2,...} by alternately applyng the followng two steps untl convergence: E-Step : Compute the condtonal expectaton of the hdden data, gven X and the current estmate Θ(t). Snce log p(x, Z Θ) s lnear wth respect to the mssng data Z, we smply have to compute the condtonal expectaton W E[Z X, Θ(t)], and plug t nto log p(x, Z Θ). ThsgvestheQ-functon as follows:

4 Q(Θ Θ(t)) E Z [log p(x, Z) X, Θ(t)] (7) Snce Z s a bnary vector, ts condtonal expectaton sgvenby: E [ z (j) X, Θ(t)] = Pr [ z (j) =1 x (j), Θ(t)] = α (t)p(x (j) θ (t)) k =1 α (t)p(x (j) θ (t)) (8) where the last equalty s smply the Bayes law (α s the a pror probablty that z (j) =1), whle s the a posteror probablty that z (j) =1gven the observaton x (j). M-Step : The estmates of the new parameters are updated usng the followng equaton : 3.3 EM for GMMs Θ(t +1)=arg max Θ {Q(Θ, Θ(t))} (9) Several varants of the EM algorthm have been extensvely used to solve ths problem. The convergence propertes of the EM algorthm for Gaussan mxtures are thoroughly dscussed n [19]. The Q functon for GMM s gven by: Q(Θ Θ(t)) = where = k =1 k =1 1 [log σ 2π (x(j) µ ) 2 2σ 2 α (t) 1 σ (t) e 2σ (t) 2 (x(j) µ (t)) 2 + log α ] (10) (11) α (t) 1 σ (t) e 2σ (t) 2 (x(j) µ (t)) 2 The maxmzaton step s gven by the followng equaton : Q(Θ Θ(t)) = 0 (12) Θ k where Θ k s the parameters for the k th component. Because of the assumpton made that each data pont comes from a sngle component, solvng the above equaton becomes trval. The updates for the maxmzaton step n the case of GMMs are gven as follows: µ (t +1)= σ 2 (t +1)= α (t +1)= 1 n n w(j) x (j) n w(j) n w(j) (x (j) µ (t +1)) 2 n w(j) 3.4 Stablty Regons Ths secton manly deals wth the transformaton of the orgnal log-lkelhood functon nto ts correspondng nonlnear dynamcal system and ntroduces some termnology pertnent to comprehend our algorthm. Ths transformaton gves the correspondence between all the crtcal ponts of the s-dmensonal lkelhood surface and that of ts dynamcal system. For the case of sphercal Gaussan mxtures wth k components, we have the number of unknown parameters s =3k 1. For convenence, the maxmzaton problem s transformed nto a mnmzaton problem defned by the followng objectve functon : mn Θ f(θ) = mn { log p(y Θ) } Θ = max Θ { log p(y Θ) } (13) where f(θ) s assumed to be n C 2 (R s, R). Defnton 1 Θ s sad to be a crtcal pont of (13) f t satsfes the followng condton f( Θ) = 0 (14) A crtcal pont s sad to be nondegenerate f at the crtcal pont Θ R s, d T 2 f( Θ)d 0 ( d 0). We construct the followng gradent system n order to locate crtcal ponts of the objectve functon (13): Θ(t) = f(θ) (15) where the state vector Θ belongs to the Eucldean space R s, and the vector feld f : R s R s satsfes the suffcent condton for the exstence and unqueness of the solutons. The soluton curve of Eq. (15) startng from Θ at tme t =0 s called a trajectory and t s denoted by Φ(Θ, ) :R R s. A state vector Θ s called an equlbrum pont of Eq. (15) f f(θ) = 0. An equlbrum pont s sad to be hyperbolc f the Jacoban of f at pont Θ has no egenvalues wth zero

5 real part. The gradent system for the log-lkelhood functon n the case of sphercal Gaussans s constructed as follows : [ µ 1 (t).. µ k (t) σ 1 (t).. σ k (t) α 1 (t).. α k 1 (t)] T [ f f f f =.... µ 1 µ k σ 1 σ k where f µ = f σ = f α = 1 α (x (j) µ ) 2σ 2 [ 1σ + (x(j) µ ) 2 ] σ 3 f α 1.. f α k 1 =1,.., k ] T =1,.., k (16) =1,.., k 1 For smplcty, we show the constructon of the gradent system for the case of sphercal Gaussans. It can be easly extended to the full covarance Gaussan mxture case. It should be noted that only (k-1) α values are consdered n the gradent system because of the unty constrant. The dependent varable α k s wrtten as follows: k 1 α k =1 α j (17) Defnton 2 A hyperbolc equlbrum pont s called a (asymptotcally) stable equlbrum pont (SEP) f all the egenvalues of ts correspondng Jacoban have negatve real part. Conversely, t s an unstable equlbrum pont f some egenvalues have a postve real part. An equlbrum pont s called a type-k equlbrum pont f ts correspondng Jacoban has exact k egenvalues wth postve real part. The stable (W s ( x)) andunstable (W u ( x)) manfolds of an equlbrum pont, say x, sdefned as: W s ( x) ={x R s : lm Φ(x, t) = x} (18) t W u ( x) ={x R s : lm Φ(x, t) = x} (19) t The task of fndng multple local maxma on the loglkelhood surface s transformed nto the task of fndng multple stable equlbrum ponts on ts correspondng gradent system. The advantage of our approach s that ths transformaton nto the correspondng dynamcal system wll yeld more knowledge about the varous dynamc and geometrc characterstcs of the orgnal surface and leads to the development a powerful method for fndng mproved solutons. In ths paper, we are partcularly nterested n the propertes of the local maxma and ther one-to-one correspondence to the stable equlbrum ponts. To comprehend the transformaton, we need to defne energy functon. A smooth functon V ( ) : R s R s satsfyng V (Φ(Θ,t)) < 0, x / {set of equlbrum ponts (E)} and t R + s termed as energy functon. Theorem 3.1 [3]: f(θ) s a energy functon for the gradent system (15). Defnton 3 A type-1 equlbrum pont x d (k=1) on the practcal stablty boundary of a stable equlbrum pont x s s called a decomposton pont. Defnton 4 The practcal stablty regon of a stable equlbrum pont x s of a nonlnear dynamcal system (15), denoted by A p (x s ) and s the nteror of closure of the stablty regon A(x s ) whch s gven by : A(x s )={x R s : lm t Φ(x, t) =x s } (20) The boundary of practcal stablty regon s called the practcal stablty boundary of x s and wll be denoted by A p (x s ). Theorem 3.2 asserts that the practcal stablty boundary s contaned n the unon of the closure of the stable manfolds of all the decomposton ponts on the practcal stablty boundary. Hence, f the decomposton ponts can be dentfed, then an explct characterzaton of the practcal stablty boundary can be establshed usng (21). Ths theorem gves an explct descrpton of the geometrcal and dynamcal structure of the practcal stablty boundary. Theorem 3.2 (Characterzaton of practcal stablty boundary)[7]: Consder a negatve gradent system descrbed by (15). Let σ, =1,2,... be the decomposton ponts on the practcal stablty boundary A p (x s ) of a stable equlbrum pont, say x s.then A p (x s )= σ A p W s (σ ). (21)

6 (a) Parameter Space (b) Functon Space Fgure 2. Varous stages of our algorthm n (a) Parameter space - the sold lnes ndcate the practcal stablty boundary. Ponts hghlghted on the stablty boundary (σ 1,σ 2 ) are the decomposton ponts. The dotted lnes ndcate the convergence of the EM algorthm. d j are the promsng drectons generated at the local maxmum LM. The dashed lnes ndcate the stablty regon phase. x 1, x 2 and x 3 are the ext ponts on the practcal stablty boundary (b) Dfferent varables n the functon space and ther correspondng log-lkelhood values. Our approach takes advantage of TRUST-TECH (TRansformaton Under STablty-reTanng Equlbra CHaracterzaton) to compute neghborhood local maxma on lkelhood surface usng stablty regons. Orgnally, the basc dea of our algorthm was to fnd decomposton ponts on the practcal stablty boundary. Snce, each decomposton pont connects two local maxma unquely, t s mportant to obtan the saddle ponts from the gven local maxmum and then move to the next local maxmum through ths decomposton pont [14]. Though, ths procedure gves a guarantee that the local maxmum s not revsted, the computatonal expense for tracng the stablty boundary and dentfyng the decomposton pont s hgh compared to the cost of applyng the EM algorthm drectly usng the ext pont wthout consderng the decomposton pont. One can use the saddle pont tracng procedure descrbed n [14] for applcatons where the local methods lke EM are much expensve. 4 Our Algorthm Our framework conssts manly of two phases whch are repeated n the promsng subspaces of the parameter search space. It s more effectve to use our algorthm at only these promsng subspaces whch are usually obtaned by stochastc global methods. The frst phase s the local phase (or the EM phase) where the promsng solutons are refned to the correspondng locally optmal parameter set. The second phase whch s the man contrbuton of ths paper, s the stablty regon phase, where the ext ponts are computed and the neghborhood solutons are systematcally explored through these ext ponts. Fg. 2 shows the dfferent steps of our algorthm both n (a) the parameter space and (b) the functon space. Ths approach can be treated as a hybrd between global methods for ntalzaton and the EM algorthm whch gves the local maxma. One of the man advantages of our approach s that t searches the parameter space more determnstcally. Ths approach dffers from tradtonal local methods by computng multple local solutons n the neghborhood regon. Ths also enhances user flexblty by allowng the users to choose between dfferent sets of good clusterngs. Though global methods gve promsng subspaces, t s mportant to explore ths subspace more thoroughly especally n problems lke parameter estmaton. Algorthm 1 descrbes our approach. In order to escape out of ths local maxmum, our methods needs to compute certan promsng drectons based on the local behavour of the functon. One can realze that generatng these promsng drectons s one of the mportant aspects of our algorthm. Surprsngly, choosng random drectons to move out of the local maxmum works

7 Algorthm 1 Stablty Regon based EM Algorthm Input: Parameters Θ,DataX, tolerance τ, StepS p Output: Θ MLE Algorthm: Apply global method and store the q promsng solutons Θ nt = {Θ 1, Θ 2,.., Θ q } Intalze E= φ whle Θ nt φ do Choose Θ Θ nt,setθ nt =Θ nt \{Θ } LM = EM(Θ, X,τ) E = E {LM } Generate promsng drecton vectors d j from LM for each d j do Compute Ext Pont (X j ) along d j startng from LM by evaluatng the log-lkelhood functon gven by (4) New j = EM(X j + ɛ d j, X,τ) f new j / E then E = E New j end f end for end whle Θ MLE = max{val(e )} well for ths problem. One mght also use other drectons lke egenvectors of the Hessan or ncorporate some doman-specfc knowledge (lke nformaton about prors, approxmate locaton of cluster means, user preferences on the fnal clusters) dependng on the applcaton that they are workng on and the level of computatonal expense that they can afford. We used random drectons n our work because they are very cheap to compute. Once the promsng drectons are generated, ext ponts are computed along these drectons. Ext ponts are ponts of ntersecton between any gven drecton and the practcal stablty boundary of that local maxmum along that partcular drecton. If the stablty boundary s not encountered along a gven drecton, t s very lkely that one mght not fnd any new local maxmum n that drecton. Wth a new ntal guess n the vcnty of the ext ponts, EM algorthm s appled agan to obtan a new local maxmum. 5 Implementaton Detals Our program s mplemented n MATLAB and runs on Pentum IV 2.8 GHz machne. The man procedure mplemented s TT EM descrbed n Algorthm 2. The algorthm takes the mxture data and the ntal set of parameters as nput along wth step sze for movng out and tolerance for convergence n the EM algorthm. It returns the set of parameters that correspond to the Ter-1 neghborng local optmal solutons. The procedure eval returns the log-lkelhood score gven by (4). The Gen Dr procedure generates promsng drectons from the local maxma. Ext ponts are obtaned along these generated drectons. The procedure update moves the current parameter to the next parameter set along a gven k th drecton Dr[k]. Someof the drectons mght have one of the followng two problems: () Ext ponts mght not be obtaned n these drectons. () Even f the ext pont s obtaned t mght converge to a less promsng soluton. If the ext ponts are not found along these drectons, search wll be termnated after Eval MAX number of evaluatons. For all ext ponts that are successfully found, EM procedure s appled and all the correspondng neghborhood set of parameters are stored n the Params[] 1. Snce, dfferent parameters wll be of dfferent range, care must be taken whle multplyng wth the step szes. It s mportant to use the current estmates to get an approxmaton of the step sze wth whch one should move out along each parameter n the search space. Fnally, the soluton wth the hghest lkelhood score amongst the orgnal set of parameters and the Ter-1 solutons s returned. Algorthm 2 Params[ ] TT EM(Pset, Data, T ol, Step) Val= eval(pset) Dr[ ]=Gen Dr(Pset) Eval MAX = 500 for k =1to sze(dr) do Params[k] =Pset ExtPt = OFF Prev Val= Val Cnt=0 whle (! ExtPt)&&(Cnt < Eval MAX) do Params[k] =update(params[k],dr[k],step) Cnt = Cnt +1 Next Val = eval(params[k]) f (Next Val > Prev Val) then ExtPt = ON end f Prev Val= Next Val end whle f count < Eval MAX then Params[k] =update(params[k],dr[k],asc) Params[k] =EM(Params[k], Data, T ol) else Params[k] =NULL end f end for Return max(eval(params[])) 6 Results and Dscusson Our algorthm has been tested on both synthetc and real datasets. The ntal values for the centers and the covar- 1 To ensure that the new ntal ponts are n the dfferent stablty regons, one should move along the drectons ɛ away from the ext ponts.

8 (a) (b) (c) (d) Fgure 3. Parameter estmates at varous stages of our algorthm on the three component Gaussan mxture model (a) Poor random ntal guess (b) Local maxmum obtaned after applyng EM algorthm wth the poor ntal guess (c) Ext pont obtaned by our algorthm (d) The fnal soluton obtaned by applyng the EM algorthm to the ntal pont n the neghborng stablty regon. ances were chosen unformly random. Unform prors were chosen for ntalzng the components. For real datasets, the centers were chosen randomly from the sample ponts. Fgure 4. Graph showng lkelhood vs Evaluatons. A corresponds to the orgnal local maxmum (L= ). B corresponds to the ext pont (L= ). C corresponds to the new ntal pont n the neghborng stablty regon (L= ) after movng out by ɛ. D corresponds to the new local maxmum (L= ). 6.1 Synthetc Datasets A smple synthetc data wth 40 samples and 5 sphercal Gaussan components was generated and tested wth our algorthm. Prors were unform and the standard devaton was The centers for the fve components are gven as follows: µ 1 =[0.30.3] T, µ 2 =[0.50.5] T, µ 3 =[0.70.7] T, µ 4 =[0.30.7] T and µ 5 =[0.70.3] T. The second dataset was that of a dagonal covarance case contanng n = 900 data ponts. The data generated from a two-dmensonal, three-component Gaussan mxture dstrbuton wth mean vectors at [0 2] T, [0 0] T, [0 2] T and same dagonal covarance matrx wth values 2 and 0.2 along the dagonal [16]. All the three mxtures have unform prors. Fg. 3 shows varous stages of our algorthm and demonstrates how the clusters obtaned from exstng algorthms are mproved usng our algorthm. The ntal clusters obtaned are of low qualty because of the poor ntal set of parameters. Our algorthm takes these clusters and apples the stablty regon step and the EM step smultaneously to obtan the fnal result. Fg. 4 shows the value of the log-lkelhood durng the stablty regon phase and the EM teratons. In the thrd synthetc dataset, a more complcated overlappng Gaussan mxtures are consdered [6]. The parameters are as follows: µ 1 = µ 2 =[ 4 4] T, µ 3 =[22] T and µ 4 =[ 1 6] T. α 1 = α 2 = α 3 =0.3and α 4 =0.1. Σ 1 = Σ 3 = [ [ ] ] 6.2 Real Datasets Σ 2 = Σ 4 = [ ] [ Two real datasets obtaned from the UCI Machne Learnng repostory [1] were also used for testng the performance of our algorthm. Most wdely used Irs data wth 150 samples, 3 classes and 4 features was used. Wne data set wth ]

9 Table 2. Performance of our algorthm on an average of 100 runs on varous synthetc and real datasets Dataset Samples Clusters Features EM (mean ± std) TRUST-TECH-EM (mean ± std) Sphercal ± ±0.6 Ellptcal ± ±0.03 Full covarance ± ± Full covarance ± ±37.02 Irs ± ±11.72 Wne ± ± samples was also used for testng. Wne data had 3 classes and 13 features. For these real data sets, the class labels were deleted thus treatng t as unsupervsed learnng problem. Table 2 summarzes our results over 100 runs. The mean and the standard devatons of the log-lkelhood values are reported. The tradtonal EM algorthm wth random starts s compared aganst our algorthm on both synthetc and real data sets. Our algorthm not only obtans hgher lkelhood value but also produces t wth hgh confdence. The low standard devaton of our results ndcates the robustness of obtanng the global maxmum. In the case of the wne data, the mprovements wth our algorthm are not much sgnfcant compared to the other datasets. Ths mght be due to the fact that the dataset mght not have Gaussan components. Our method assumes that the underlyng dstrbuton of the data s mxture of Gaussans. Table 3 gves the results of TRUST-TECH compared wth other methods lke splt and merge EM and k-means+em proposed n the lterature. Table 3. Comparson of TRUST-TECH-EM wth other methods Method Ellptcal Irs RS+EM ± ± 27 K-Means+EM ± ± 10 SMEM ± ± 6 TRUST-TECH-EM ± ± 11 clusters obtaned. Most of the focus n the lterature was on new methods for ntalzaton or new clusterng technques whch often do not take advantage of the exstng results and completely start the clusterng procedure from scratch. Though shown only for the case of multvarate Gaussan mxtures, our technque can be effectvely appled to any parametrc fnte mxture model. Table 4 summarzes the average number of teratons taken by the EM algorthm for the convergence to the local optmal soluton. We can see that the most promsng soluton produced by our TRUST-TECH methodology converges much faster. In other words, our method can effectvely take advantage of the fact that the convergence of the EM algorthm s much faster for hgh qualty solutons. Ths s an nherent property of the EM algorthm when appled to the mxture modelng problem. We explot ths property of the EM for mprovng the effcency of our algorthm. Hence, for obtanng the Ter-1 solutons usng our algorthm, the threshold for the number of teratons can be sgnfcantly lowered. Table 4. Number of teratons taken for the convergence of the best soluton. Dataset Avg. no. of No. of teratons teratons for the best soluton Sphercal Ellptcal Full covarance Dscusson It wll be effectve to use our algorthm for those solutons that appear to be promsng. Due to the nature of the problem, t s very lkely that the nearby solutons surroundng the exstng soluton wll be more promsng. One of the prmary advantages of our method s that t can be used along wth other popular methods avalable and mprove the qualty of the exstng solutons. In clusterng problems, t s an added advantage to perform refnement of the fnal 7 Concluson and Future Work A novel stablty regon based EM algorthm has been ntroduced for estmatng the parameters of mxture models. The EM phase and the stablty regon phase are appled alternatvely n the context of the well-studed mxture model parameter estmaton problem. The concept of stablty regon helps us to understand the topology of the orgnal loglkelhood surface. Our method computes the neghborhood

10 local maxma on lkelhood surface usng stablty regons of the correspondng nonlnear dynamcal system. The algorthm has been tested successfully on varous synthetc and real datasets and the mprovements n the performance are clearly manfested. Some propertes of the EM algorthm about the rate of convergence have been exploted effcently. Our algorthm can be easly extended to popularly used k-means clusterng technque. In the future, we plan to work on applyng these stablty regon based methods for other wdely used EM related parameter estmaton problems lke tranng Hdden Markov Models, Mxture of Factor Analyzers, Probablstc Prncpal Component Analyss, Bayesan Networks etc. We would also plan to extend our technque to Markov Chan Monte Carlo strateges lke Gbbs samplng for the estmaton of mxture models. References [1] C.L. Blake and C.J. Merz. UCI repostory of machne learnng databases. Unversty of Calforna, Irvne, Dept. of Informaton and Computer Scences, [2] C. Carson, S. Belonge, H. Greenspan, and J. Malk. Blobworld: Image segmentaton usng expectatonmaxmzaton and ts applcaton to mage queryng. IEEE Transactons on Pattern Analyss and Machne Intellgence, 24(8): , [3] H.D. Chang and C.C. Chu. A systematc search method for obtanng multple local optmal solutons of nonlnear programmng problems. IEEE Transactons on Crcuts and Systems: I Fundamental Theory and Applcatons, 43(2):99 109, [4] A. P. Demspter, N. A. Lard, and D. B. Rubn. Maxmum lkelhood from ncomplete data va the EM algorthm. Journal of the Royal Statstcal Socety Seres B, 39(1):1 38, [5] G. Eldan, M. Nno, N. Fredman, and D. Schuurmans. Data perturbaton for escapng local maxma n learnng. In Proceedngs of the Eghteenth Natonal Conference on Artfcal Intellgence, pages , [6] M. Fgueredo and A.K. Jan. Unsupervsed learnng of fnte mxture models. IEEE Transactons on Pattern Analyss and Machne Intellgence, 24(3): , [7] J. Lee and H.D. Chang. A dynamcal trajectory-based methodology for systematcally computng multple optmal solutons of general nonlnear programmng problems. IEEE Transactons on Automatc Control, 49(6): , [8] J.Q.L. Estmaton of Mxture Models. PhD thess, Department of Statstcs,Yale Unversty, [9] G. McLachlan and T. Krshnan. The EM Algorthm and Extensons. John Wley and Sons, New York, [10] G. J. McLachlan and K. E. Basford. Mxture models: Inference and applcatons to clusterng. Marcel Dekker, New York, [11] R. M. Neal and G. E. Hnton. A new vew of the EM algorthm that justfes ncremental, sparse and other varants. In M. I. Jordan, edtor, Learnng n Graphcal Models, pages Kluwer Academc Publshers, [12] K. Ngam, A. McCallum, S. Thrun, and T. Mtchell. Text classfcaton from labeled and unlabeled documents usng EM. Machne Learnng, 39(2-3): , [13] F. Pernkopf and D. Bouchaffra. Genetc-based EM algorthm for learnng gaussan mxture models. IEEE Transactons on Pattern Analyss and Machne Intellgence, 27(8): , [14] C. K. Reddy and H.D. Chang. A stablty boundary based method for fndng saddle ponts on potental energy surfaces. Journal of Computatonal Bology, 13(3): , [15] S. J. Roberts, D. Husmeer, I. Rezek, and W. Penny. Bayesan approaches to gaussan mxture modelng. IEEE Transactons on Pattern Analyss and Machne Intellgence, 20(11): , [16] N. Ueda and R. Nakano. Determnstc annealng EM algorthm. Neural Networks, 11(2): , [17] N. Ueda, R. Nakano, Z. Ghahraman, and G.E. Hnton. SMEM algorthm for mxture models. Neural Computaton, 12(9): , [18] J. J. Verbeek, N. Vlasss, and B. Krose. Effcent greedy learnng of gaussan mxture models. Neural Computaton, 15(2): , [19] L. Xu and M. I. Jordan. On convergence propertes of the EM algorthm for gaussan mxtures. Neural Computaton, 8(1): , [20] B. Zhang, C. Zhang, and X. Y. Compettve EM algorthm for fnte mxture models. Pattern Recognton, 37(1): , 2004.

Learning the Kernel Parameters in Kernel Minimum Distance Classifier

Learning the Kernel Parameters in Kernel Minimum Distance Classifier Learnng the Kernel Parameters n Kernel Mnmum Dstance Classfer Daoqang Zhang 1,, Songcan Chen and Zh-Hua Zhou 1* 1 Natonal Laboratory for Novel Software Technology Nanjng Unversty, Nanjng 193, Chna Department

More information

CS 534: Computer Vision Model Fitting

CS 534: Computer Vision Model Fitting CS 534: Computer Vson Model Fttng Sprng 004 Ahmed Elgammal Dept of Computer Scence CS 534 Model Fttng - 1 Outlnes Model fttng s mportant Least-squares fttng Maxmum lkelhood estmaton MAP estmaton Robust

More information

Biostatistics 615/815

Biostatistics 615/815 The E-M Algorthm Bostatstcs 615/815 Lecture 17 Last Lecture: The Smplex Method General method for optmzaton Makes few assumptons about functon Crawls towards mnmum Some recommendatons Multple startng ponts

More information

Outline. Type of Machine Learning. Examples of Application. Unsupervised Learning

Outline. Type of Machine Learning. Examples of Application. Unsupervised Learning Outlne Artfcal Intellgence and ts applcatons Lecture 8 Unsupervsed Learnng Professor Danel Yeung danyeung@eee.org Dr. Patrck Chan patrckchan@eee.org South Chna Unversty of Technology, Chna Introducton

More information

Parallelism for Nested Loops with Non-uniform and Flow Dependences

Parallelism for Nested Loops with Non-uniform and Flow Dependences Parallelsm for Nested Loops wth Non-unform and Flow Dependences Sam-Jn Jeong Dept. of Informaton & Communcaton Engneerng, Cheonan Unversty, 5, Anseo-dong, Cheonan, Chungnam, 330-80, Korea. seong@cheonan.ac.kr

More information

Unsupervised Learning

Unsupervised Learning Pattern Recognton Lecture 8 Outlne Introducton Unsupervsed Learnng Parametrc VS Non-Parametrc Approach Mxture of Denstes Maxmum-Lkelhood Estmates Clusterng Prof. Danel Yeung School of Computer Scence and

More information

SLAM Summer School 2006 Practical 2: SLAM using Monocular Vision

SLAM Summer School 2006 Practical 2: SLAM using Monocular Vision SLAM Summer School 2006 Practcal 2: SLAM usng Monocular Vson Javer Cvera, Unversty of Zaragoza Andrew J. Davson, Imperal College London J.M.M Montel, Unversty of Zaragoza. josemar@unzar.es, jcvera@unzar.es,

More information

The Research of Support Vector Machine in Agricultural Data Classification

The Research of Support Vector Machine in Agricultural Data Classification The Research of Support Vector Machne n Agrcultural Data Classfcaton Le Sh, Qguo Duan, Xnmng Ma, Me Weng College of Informaton and Management Scence, HeNan Agrcultural Unversty, Zhengzhou 45000 Chna Zhengzhou

More information

User Authentication Based On Behavioral Mouse Dynamics Biometrics

User Authentication Based On Behavioral Mouse Dynamics Biometrics User Authentcaton Based On Behavoral Mouse Dynamcs Bometrcs Chee-Hyung Yoon Danel Donghyun Km Department of Computer Scence Department of Computer Scence Stanford Unversty Stanford Unversty Stanford, CA

More information

EECS 730 Introduction to Bioinformatics Sequence Alignment. Luke Huan Electrical Engineering and Computer Science

EECS 730 Introduction to Bioinformatics Sequence Alignment. Luke Huan Electrical Engineering and Computer Science EECS 730 Introducton to Bonformatcs Sequence Algnment Luke Huan Electrcal Engneerng and Computer Scence http://people.eecs.ku.edu/~huan/ HMM Π s a set of states Transton Probabltes a kl Pr( l 1 k Probablty

More information

Hermite Splines in Lie Groups as Products of Geodesics

Hermite Splines in Lie Groups as Products of Geodesics Hermte Splnes n Le Groups as Products of Geodescs Ethan Eade Updated May 28, 2017 1 Introducton 1.1 Goal Ths document defnes a curve n the Le group G parametrzed by tme and by structural parameters n the

More information

Determining the Optimal Bandwidth Based on Multi-criterion Fusion

Determining the Optimal Bandwidth Based on Multi-criterion Fusion Proceedngs of 01 4th Internatonal Conference on Machne Learnng and Computng IPCSIT vol. 5 (01) (01) IACSIT Press, Sngapore Determnng the Optmal Bandwdth Based on Mult-crteron Fuson Ha-L Lang 1+, Xan-Mn

More information

Three supervised learning methods on pen digits character recognition dataset

Three supervised learning methods on pen digits character recognition dataset Three supervsed learnng methods on pen dgts character recognton dataset Chrs Flezach Department of Computer Scence and Engneerng Unversty of Calforna, San Dego San Dego, CA 92093 cflezac@cs.ucsd.edu Satoru

More information

A Robust Method for Estimating the Fundamental Matrix

A Robust Method for Estimating the Fundamental Matrix Proc. VIIth Dgtal Image Computng: Technques and Applcatons, Sun C., Talbot H., Ourseln S. and Adraansen T. (Eds.), 0- Dec. 003, Sydney A Robust Method for Estmatng the Fundamental Matrx C.L. Feng and Y.S.

More information

Problem Definitions and Evaluation Criteria for Computational Expensive Optimization

Problem Definitions and Evaluation Criteria for Computational Expensive Optimization Problem efntons and Evaluaton Crtera for Computatonal Expensve Optmzaton B. Lu 1, Q. Chen and Q. Zhang 3, J. J. Lang 4, P. N. Suganthan, B. Y. Qu 6 1 epartment of Computng, Glyndwr Unversty, UK Faclty

More information

Classifier Selection Based on Data Complexity Measures *

Classifier Selection Based on Data Complexity Measures * Classfer Selecton Based on Data Complexty Measures * Edth Hernández-Reyes, J.A. Carrasco-Ochoa, and J.Fco. Martínez-Trndad Natonal Insttute for Astrophyscs, Optcs and Electroncs, Lus Enrque Erro No.1 Sta.

More information

Unsupervised Learning and Clustering

Unsupervised Learning and Clustering Unsupervsed Learnng and Clusterng Why consder unlabeled samples?. Collectng and labelng large set of samples s costly Gettng recorded speech s free, labelng s tme consumng 2. Classfer could be desgned

More information

Feature Reduction and Selection

Feature Reduction and Selection Feature Reducton and Selecton Dr. Shuang LIANG School of Software Engneerng TongJ Unversty Fall, 2012 Today s Topcs Introducton Problems of Dmensonalty Feature Reducton Statstc methods Prncpal Components

More information

Lecture 4: Principal components

Lecture 4: Principal components /3/6 Lecture 4: Prncpal components 3..6 Multvarate lnear regresson MLR s optmal for the estmaton data...but poor for handlng collnear data Covarance matrx s not nvertble (large condton number) Robustness

More information

Simulation: Solving Dynamic Models ABE 5646 Week 11 Chapter 2, Spring 2010

Simulation: Solving Dynamic Models ABE 5646 Week 11 Chapter 2, Spring 2010 Smulaton: Solvng Dynamc Models ABE 5646 Week Chapter 2, Sprng 200 Week Descrpton Readng Materal Mar 5- Mar 9 Evaluatng [Crop] Models Comparng a model wth data - Graphcal, errors - Measures of agreement

More information

Mathematics 256 a course in differential equations for engineering students

Mathematics 256 a course in differential equations for engineering students Mathematcs 56 a course n dfferental equatons for engneerng students Chapter 5. More effcent methods of numercal soluton Euler s method s qute neffcent. Because the error s essentally proportonal to the

More information

Content Based Image Retrieval Using 2-D Discrete Wavelet with Texture Feature with Different Classifiers

Content Based Image Retrieval Using 2-D Discrete Wavelet with Texture Feature with Different Classifiers IOSR Journal of Electroncs and Communcaton Engneerng (IOSR-JECE) e-issn: 78-834,p- ISSN: 78-8735.Volume 9, Issue, Ver. IV (Mar - Apr. 04), PP 0-07 Content Based Image Retreval Usng -D Dscrete Wavelet wth

More information

Support Vector Machines

Support Vector Machines /9/207 MIST.6060 Busness Intellgence and Data Mnng What are Support Vector Machnes? Support Vector Machnes Support Vector Machnes (SVMs) are supervsed learnng technques that analyze data and recognze patterns.

More information

Backpropagation: In Search of Performance Parameters

Backpropagation: In Search of Performance Parameters Bacpropagaton: In Search of Performance Parameters ANIL KUMAR ENUMULAPALLY, LINGGUO BU, and KHOSROW KAIKHAH, Ph.D. Computer Scence Department Texas State Unversty-San Marcos San Marcos, TX-78666 USA ae049@txstate.edu,

More information

X- Chart Using ANOM Approach

X- Chart Using ANOM Approach ISSN 1684-8403 Journal of Statstcs Volume 17, 010, pp. 3-3 Abstract X- Chart Usng ANOM Approach Gullapall Chakravarth 1 and Chaluvad Venkateswara Rao Control lmts for ndvdual measurements (X) chart are

More information

Smoothing Spline ANOVA for variable screening

Smoothing Spline ANOVA for variable screening Smoothng Splne ANOVA for varable screenng a useful tool for metamodels tranng and mult-objectve optmzaton L. Rcco, E. Rgon, A. Turco Outlne RSM Introducton Possble couplng Test case MOO MOO wth Game Theory

More information

Subspace clustering. Clustering. Fundamental to all clustering techniques is the choice of distance measure between data points;

Subspace clustering. Clustering. Fundamental to all clustering techniques is the choice of distance measure between data points; Subspace clusterng Clusterng Fundamental to all clusterng technques s the choce of dstance measure between data ponts; D q ( ) ( ) 2 x x = x x, j k = 1 k jk Squared Eucldean dstance Assumpton: All features

More information

Meta-heuristics for Multidimensional Knapsack Problems

Meta-heuristics for Multidimensional Knapsack Problems 2012 4th Internatonal Conference on Computer Research and Development IPCSIT vol.39 (2012) (2012) IACSIT Press, Sngapore Meta-heurstcs for Multdmensonal Knapsack Problems Zhbao Man + Computer Scence Department,

More information

A New Approach For the Ranking of Fuzzy Sets With Different Heights

A New Approach For the Ranking of Fuzzy Sets With Different Heights New pproach For the ankng of Fuzzy Sets Wth Dfferent Heghts Pushpnder Sngh School of Mathematcs Computer pplcatons Thapar Unversty, Patala-7 00 Inda pushpndersnl@gmalcom STCT ankng of fuzzy sets plays

More information

S1 Note. Basis functions.

S1 Note. Basis functions. S1 Note. Bass functons. Contents Types of bass functons...1 The Fourer bass...2 B-splne bass...3 Power and type I error rates wth dfferent numbers of bass functons...4 Table S1. Smulaton results of type

More information

Hybridization of Expectation-Maximization and K-Means Algorithms for Better Clustering Performance

Hybridization of Expectation-Maximization and K-Means Algorithms for Better Clustering Performance BULGARIAN ACADEMY OF SCIENCES CYBERNETICS AND INFORMATION TECHNOLOGIES Volume 16, No 2 Sofa 2016 Prnt ISSN: 1311-9702; Onlne ISSN: 1314-4081 DOI: 10.1515/cat-2016-0017 Hybrdzaton of Expectaton-Maxmzaton

More information

A Binarization Algorithm specialized on Document Images and Photos

A Binarization Algorithm specialized on Document Images and Photos A Bnarzaton Algorthm specalzed on Document mages and Photos Ergna Kavalleratou Dept. of nformaton and Communcaton Systems Engneerng Unversty of the Aegean kavalleratou@aegean.gr Abstract n ths paper, a

More information

An Iterative Solution Approach to Process Plant Layout using Mixed Integer Optimisation

An Iterative Solution Approach to Process Plant Layout using Mixed Integer Optimisation 17 th European Symposum on Computer Aded Process Engneerng ESCAPE17 V. Plesu and P.S. Agach (Edtors) 2007 Elsever B.V. All rghts reserved. 1 An Iteratve Soluton Approach to Process Plant Layout usng Mxed

More information

Support Vector Machines

Support Vector Machines Support Vector Machnes Decson surface s a hyperplane (lne n 2D) n feature space (smlar to the Perceptron) Arguably, the most mportant recent dscovery n machne learnng In a nutshell: map the data to a predetermned

More information

6.854 Advanced Algorithms Petar Maymounkov Problem Set 11 (November 23, 2005) With: Benjamin Rossman, Oren Weimann, and Pouya Kheradpour

6.854 Advanced Algorithms Petar Maymounkov Problem Set 11 (November 23, 2005) With: Benjamin Rossman, Oren Weimann, and Pouya Kheradpour 6.854 Advanced Algorthms Petar Maymounkov Problem Set 11 (November 23, 2005) Wth: Benjamn Rossman, Oren Wemann, and Pouya Kheradpour Problem 1. We reduce vertex cover to MAX-SAT wth weghts, such that the

More information

Machine Learning 9. week

Machine Learning 9. week Machne Learnng 9. week Mappng Concept Radal Bass Functons (RBF) RBF Networks 1 Mappng It s probably the best scenaro for the classfcaton of two dataset s to separate them lnearly. As you see n the below

More information

y and the total sum of

y and the total sum of Lnear regresson Testng for non-lnearty In analytcal chemstry, lnear regresson s commonly used n the constructon of calbraton functons requred for analytcal technques such as gas chromatography, atomc absorpton

More information

Classifying Acoustic Transient Signals Using Artificial Intelligence

Classifying Acoustic Transient Signals Using Artificial Intelligence Classfyng Acoustc Transent Sgnals Usng Artfcal Intellgence Steve Sutton, Unversty of North Carolna At Wlmngton (suttons@charter.net) Greg Huff, Unversty of North Carolna At Wlmngton (jgh7476@uncwl.edu)

More information

EXTENDED BIC CRITERION FOR MODEL SELECTION

EXTENDED BIC CRITERION FOR MODEL SELECTION IDIAP RESEARCH REPORT EXTEDED BIC CRITERIO FOR ODEL SELECTIO Itshak Lapdot Andrew orrs IDIAP-RR-0-4 Dalle olle Insttute for Perceptual Artfcal Intellgence P.O.Box 59 artgny Valas Swtzerland phone +4 7

More information

NAG Fortran Library Chapter Introduction. G10 Smoothing in Statistics

NAG Fortran Library Chapter Introduction. G10 Smoothing in Statistics Introducton G10 NAG Fortran Lbrary Chapter Introducton G10 Smoothng n Statstcs Contents 1 Scope of the Chapter... 2 2 Background to the Problems... 2 2.1 Smoothng Methods... 2 2.2 Smoothng Splnes and Regresson

More information

Type-2 Fuzzy Non-uniform Rational B-spline Model with Type-2 Fuzzy Data

Type-2 Fuzzy Non-uniform Rational B-spline Model with Type-2 Fuzzy Data Malaysan Journal of Mathematcal Scences 11(S) Aprl : 35 46 (2017) Specal Issue: The 2nd Internatonal Conference and Workshop on Mathematcal Analyss (ICWOMA 2016) MALAYSIAN JOURNAL OF MATHEMATICAL SCIENCES

More information

LECTURE : MANIFOLD LEARNING

LECTURE : MANIFOLD LEARNING LECTURE : MANIFOLD LEARNING Rta Osadchy Some sldes are due to L.Saul, V. C. Raykar, N. Verma Topcs PCA MDS IsoMap LLE EgenMaps Done! Dmensonalty Reducton Data representaton Inputs are real-valued vectors

More information

Course Introduction. Algorithm 8/31/2017. COSC 320 Advanced Data Structures and Algorithms. COSC 320 Advanced Data Structures and Algorithms

Course Introduction. Algorithm 8/31/2017. COSC 320 Advanced Data Structures and Algorithms. COSC 320 Advanced Data Structures and Algorithms Course Introducton Course Topcs Exams, abs, Proects A quc loo at a few algorthms 1 Advanced Data Structures and Algorthms Descrpton: We are gong to dscuss algorthm complexty analyss, algorthm desgn technques

More information

GSLM Operations Research II Fall 13/14

GSLM Operations Research II Fall 13/14 GSLM 58 Operatons Research II Fall /4 6. Separable Programmng Consder a general NLP mn f(x) s.t. g j (x) b j j =. m. Defnton 6.. The NLP s a separable program f ts objectve functon and all constrants are

More information

Implementation Naïve Bayes Algorithm for Student Classification Based on Graduation Status

Implementation Naïve Bayes Algorithm for Student Classification Based on Graduation Status Internatonal Journal of Appled Busness and Informaton Systems ISSN: 2597-8993 Vol 1, No 2, September 2017, pp. 6-12 6 Implementaton Naïve Bayes Algorthm for Student Classfcaton Based on Graduaton Status

More information

Chapter 6 Programmng the fnte element method Inow turn to the man subject of ths book: The mplementaton of the fnte element algorthm n computer programs. In order to make my dscusson as straghtforward

More information

Exercises (Part 4) Introduction to R UCLA/CCPR. John Fox, February 2005

Exercises (Part 4) Introduction to R UCLA/CCPR. John Fox, February 2005 Exercses (Part 4) Introducton to R UCLA/CCPR John Fox, February 2005 1. A challengng problem: Iterated weghted least squares (IWLS) s a standard method of fttng generalzed lnear models to data. As descrbed

More information

Sum of Linear and Fractional Multiobjective Programming Problem under Fuzzy Rules Constraints

Sum of Linear and Fractional Multiobjective Programming Problem under Fuzzy Rules Constraints Australan Journal of Basc and Appled Scences, 2(4): 1204-1208, 2008 ISSN 1991-8178 Sum of Lnear and Fractonal Multobjectve Programmng Problem under Fuzzy Rules Constrants 1 2 Sanjay Jan and Kalash Lachhwan

More information

Active Contours/Snakes

Active Contours/Snakes Actve Contours/Snakes Erkut Erdem Acknowledgement: The sldes are adapted from the sldes prepared by K. Grauman of Unversty of Texas at Austn Fttng: Edges vs. boundares Edges useful sgnal to ndcate occludng

More information

Machine Learning: Algorithms and Applications

Machine Learning: Algorithms and Applications 14/05/1 Machne Learnng: Algorthms and Applcatons Florano Zn Free Unversty of Bozen-Bolzano Faculty of Computer Scence Academc Year 011-01 Lecture 10: 14 May 01 Unsupervsed Learnng cont Sldes courtesy of

More information

MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION

MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION Paulo Quntlano 1 & Antono Santa-Rosa 1 Federal Polce Department, Brasla, Brazl. E-mals: quntlano.pqs@dpf.gov.br and

More information

Improvement of Spatial Resolution Using BlockMatching Based Motion Estimation and Frame. Integration

Improvement of Spatial Resolution Using BlockMatching Based Motion Estimation and Frame. Integration Improvement of Spatal Resoluton Usng BlockMatchng Based Moton Estmaton and Frame Integraton Danya Suga and Takayuk Hamamoto Graduate School of Engneerng, Tokyo Unversty of Scence, 6-3-1, Nuku, Katsuska-ku,

More information

Cluster Analysis of Electrical Behavior

Cluster Analysis of Electrical Behavior Journal of Computer and Communcatons, 205, 3, 88-93 Publshed Onlne May 205 n ScRes. http://www.scrp.org/ournal/cc http://dx.do.org/0.4236/cc.205.350 Cluster Analyss of Electrcal Behavor Ln Lu Ln Lu, School

More information

An Application of the Dulmage-Mendelsohn Decomposition to Sparse Null Space Bases of Full Row Rank Matrices

An Application of the Dulmage-Mendelsohn Decomposition to Sparse Null Space Bases of Full Row Rank Matrices Internatonal Mathematcal Forum, Vol 7, 2012, no 52, 2549-2554 An Applcaton of the Dulmage-Mendelsohn Decomposton to Sparse Null Space Bases of Full Row Rank Matrces Mostafa Khorramzadeh Department of Mathematcal

More information

Outline. Self-Organizing Maps (SOM) US Hebbian Learning, Cntd. The learning rule is Hebbian like:

Outline. Self-Organizing Maps (SOM) US Hebbian Learning, Cntd. The learning rule is Hebbian like: Self-Organzng Maps (SOM) Turgay İBRİKÇİ, PhD. Outlne Introducton Structures of SOM SOM Archtecture Neghborhoods SOM Algorthm Examples Summary 1 2 Unsupervsed Hebban Learnng US Hebban Learnng, Cntd 3 A

More information

SVM-based Learning for Multiple Model Estimation

SVM-based Learning for Multiple Model Estimation SVM-based Learnng for Multple Model Estmaton Vladmr Cherkassky and Yunqan Ma Department of Electrcal and Computer Engneerng Unversty of Mnnesota Mnneapols, MN 55455 {cherkass,myq}@ece.umn.edu Abstract:

More information

Lecture 5: Multilayer Perceptrons

Lecture 5: Multilayer Perceptrons Lecture 5: Multlayer Perceptrons Roger Grosse 1 Introducton So far, we ve only talked about lnear models: lnear regresson and lnear bnary classfers. We noted that there are functons that can t be represented

More information

An Optimal Algorithm for Prufer Codes *

An Optimal Algorithm for Prufer Codes * J. Software Engneerng & Applcatons, 2009, 2: 111-115 do:10.4236/jsea.2009.22016 Publshed Onlne July 2009 (www.scrp.org/journal/jsea) An Optmal Algorthm for Prufer Codes * Xaodong Wang 1, 2, Le Wang 3,

More information

Journal of Chemical and Pharmaceutical Research, 2014, 6(6): Research Article

Journal of Chemical and Pharmaceutical Research, 2014, 6(6): Research Article Avalable onlne www.jocpr.com Journal of Chemcal and Pharmaceutcal Research, 2014, 6(6):2512-2520 Research Artcle ISSN : 0975-7384 CODEN(USA) : JCPRC5 Communty detecton model based on ncremental EM clusterng

More information

A MOVING MESH APPROACH FOR SIMULATION BUDGET ALLOCATION ON CONTINUOUS DOMAINS

A MOVING MESH APPROACH FOR SIMULATION BUDGET ALLOCATION ON CONTINUOUS DOMAINS Proceedngs of the Wnter Smulaton Conference M E Kuhl, N M Steger, F B Armstrong, and J A Jones, eds A MOVING MESH APPROACH FOR SIMULATION BUDGET ALLOCATION ON CONTINUOUS DOMAINS Mark W Brantley Chun-Hung

More information

A Unified Framework for Semantics and Feature Based Relevance Feedback in Image Retrieval Systems

A Unified Framework for Semantics and Feature Based Relevance Feedback in Image Retrieval Systems A Unfed Framework for Semantcs and Feature Based Relevance Feedback n Image Retreval Systems Ye Lu *, Chunhu Hu 2, Xngquan Zhu 3*, HongJang Zhang 2, Qang Yang * School of Computng Scence Smon Fraser Unversty

More information

Outline. Discriminative classifiers for image recognition. Where in the World? A nearest neighbor recognition example 4/14/2011. CS 376 Lecture 22 1

Outline. Discriminative classifiers for image recognition. Where in the World? A nearest neighbor recognition example 4/14/2011. CS 376 Lecture 22 1 4/14/011 Outlne Dscrmnatve classfers for mage recognton Wednesday, Aprl 13 Krsten Grauman UT-Austn Last tme: wndow-based generc obect detecton basc ppelne face detecton wth boostng as case study Today:

More information

Helsinki University Of Technology, Systems Analysis Laboratory Mat Independent research projects in applied mathematics (3 cr)

Helsinki University Of Technology, Systems Analysis Laboratory Mat Independent research projects in applied mathematics (3 cr) Helsnk Unversty Of Technology, Systems Analyss Laboratory Mat-2.08 Independent research projects n appled mathematcs (3 cr) "! #$&% Antt Laukkanen 506 R ajlaukka@cc.hut.f 2 Introducton...3 2 Multattrbute

More information

12/2/2009. Announcements. Parametric / Non-parametric. Case-Based Reasoning. Nearest-Neighbor on Images. Nearest-Neighbor Classification

12/2/2009. Announcements. Parametric / Non-parametric. Case-Based Reasoning. Nearest-Neighbor on Images. Nearest-Neighbor Classification Introducton to Artfcal Intellgence V22.0472-001 Fall 2009 Lecture 24: Nearest-Neghbors & Support Vector Machnes Rob Fergus Dept of Computer Scence, Courant Insttute, NYU Sldes from Danel Yeung, John DeNero

More information

Intra-Parametric Analysis of a Fuzzy MOLP

Intra-Parametric Analysis of a Fuzzy MOLP Intra-Parametrc Analyss of a Fuzzy MOLP a MIAO-LING WANG a Department of Industral Engneerng and Management a Mnghsn Insttute of Technology and Hsnchu Tawan, ROC b HSIAO-FAN WANG b Insttute of Industral

More information

FEATURE EXTRACTION. Dr. K.Vijayarekha. Associate Dean School of Electrical and Electronics Engineering SASTRA University, Thanjavur

FEATURE EXTRACTION. Dr. K.Vijayarekha. Associate Dean School of Electrical and Electronics Engineering SASTRA University, Thanjavur FEATURE EXTRACTION Dr. K.Vjayarekha Assocate Dean School of Electrcal and Electroncs Engneerng SASTRA Unversty, Thanjavur613 41 Jont Intatve of IITs and IISc Funded by MHRD Page 1 of 8 Table of Contents

More information

The Greedy Method. Outline and Reading. Change Money Problem. Greedy Algorithms. Applications of the Greedy Strategy. The Greedy Method Technique

The Greedy Method. Outline and Reading. Change Money Problem. Greedy Algorithms. Applications of the Greedy Strategy. The Greedy Method Technique //00 :0 AM Outlne and Readng The Greedy Method The Greedy Method Technque (secton.) Fractonal Knapsack Problem (secton..) Task Schedulng (secton..) Mnmum Spannng Trees (secton.) Change Money Problem Greedy

More information

NUMERICAL SOLVING OPTIMAL CONTROL PROBLEMS BY THE METHOD OF VARIATIONS

NUMERICAL SOLVING OPTIMAL CONTROL PROBLEMS BY THE METHOD OF VARIATIONS ARPN Journal of Engneerng and Appled Scences 006-017 Asan Research Publshng Network (ARPN). All rghts reserved. NUMERICAL SOLVING OPTIMAL CONTROL PROBLEMS BY THE METHOD OF VARIATIONS Igor Grgoryev, Svetlana

More information

High resolution 3D Tau-p transform by matching pursuit Weiping Cao* and Warren S. Ross, Shearwater GeoServices

High resolution 3D Tau-p transform by matching pursuit Weiping Cao* and Warren S. Ross, Shearwater GeoServices Hgh resoluton 3D Tau-p transform by matchng pursut Wepng Cao* and Warren S. Ross, Shearwater GeoServces Summary The 3D Tau-p transform s of vtal sgnfcance for processng sesmc data acqured wth modern wde

More information

CHAPTER 3 SEQUENTIAL MINIMAL OPTIMIZATION TRAINED SUPPORT VECTOR CLASSIFIER FOR CANCER PREDICTION

CHAPTER 3 SEQUENTIAL MINIMAL OPTIMIZATION TRAINED SUPPORT VECTOR CLASSIFIER FOR CANCER PREDICTION 48 CHAPTER 3 SEQUENTIAL MINIMAL OPTIMIZATION TRAINED SUPPORT VECTOR CLASSIFIER FOR CANCER PREDICTION 3.1 INTRODUCTION The raw mcroarray data s bascally an mage wth dfferent colors ndcatng hybrdzaton (Xue

More information

Module Management Tool in Software Development Organizations

Module Management Tool in Software Development Organizations Journal of Computer Scence (5): 8-, 7 ISSN 59-66 7 Scence Publcatons Management Tool n Software Development Organzatons Ahmad A. Al-Rababah and Mohammad A. Al-Rababah Faculty of IT, Al-Ahlyyah Amman Unversty,

More information

Collaboratively Regularized Nearest Points for Set Based Recognition

Collaboratively Regularized Nearest Points for Set Based Recognition Academc Center for Computng and Meda Studes, Kyoto Unversty Collaboratvely Regularzed Nearest Ponts for Set Based Recognton Yang Wu, Mchhko Mnoh, Masayuk Mukunok Kyoto Unversty 9/1/013 BMVC 013 @ Brstol,

More information

Parameter estimation for incomplete bivariate longitudinal data in clinical trials

Parameter estimation for incomplete bivariate longitudinal data in clinical trials Parameter estmaton for ncomplete bvarate longtudnal data n clncal trals Naum M. Khutoryansky Novo Nordsk Pharmaceutcals, Inc., Prnceton, NJ ABSTRACT Bvarate models are useful when analyzng longtudnal data

More information

A Background Subtraction for a Vision-based User Interface *

A Background Subtraction for a Vision-based User Interface * A Background Subtracton for a Vson-based User Interface * Dongpyo Hong and Woontack Woo KJIST U-VR Lab. {dhon wwoo}@kjst.ac.kr Abstract In ths paper, we propose a robust and effcent background subtracton

More information

Wishing you all a Total Quality New Year!

Wishing you all a Total Quality New Year! Total Qualty Management and Sx Sgma Post Graduate Program 214-15 Sesson 4 Vnay Kumar Kalakband Assstant Professor Operatons & Systems Area 1 Wshng you all a Total Qualty New Year! Hope you acheve Sx sgma

More information

Compiler Design. Spring Register Allocation. Sample Exercises and Solutions. Prof. Pedro C. Diniz

Compiler Design. Spring Register Allocation. Sample Exercises and Solutions. Prof. Pedro C. Diniz Compler Desgn Sprng 2014 Regster Allocaton Sample Exercses and Solutons Prof. Pedro C. Dnz USC / Informaton Scences Insttute 4676 Admralty Way, Sute 1001 Marna del Rey, Calforna 90292 pedro@s.edu Regster

More information

Skew Angle Estimation and Correction of Hand Written, Textual and Large areas of Non-Textual Document Images: A Novel Approach

Skew Angle Estimation and Correction of Hand Written, Textual and Large areas of Non-Textual Document Images: A Novel Approach Angle Estmaton and Correcton of Hand Wrtten, Textual and Large areas of Non-Textual Document Images: A Novel Approach D.R.Ramesh Babu Pyush M Kumat Mahesh D Dhannawat PES Insttute of Technology Research

More information

Machine Learning. Support Vector Machines. (contains material adapted from talks by Constantin F. Aliferis & Ioannis Tsamardinos, and Martin Law)

Machine Learning. Support Vector Machines. (contains material adapted from talks by Constantin F. Aliferis & Ioannis Tsamardinos, and Martin Law) Machne Learnng Support Vector Machnes (contans materal adapted from talks by Constantn F. Alfers & Ioanns Tsamardnos, and Martn Law) Bryan Pardo, Machne Learnng: EECS 349 Fall 2014 Support Vector Machnes

More information

Solving two-person zero-sum game by Matlab

Solving two-person zero-sum game by Matlab Appled Mechancs and Materals Onlne: 2011-02-02 ISSN: 1662-7482, Vols. 50-51, pp 262-265 do:10.4028/www.scentfc.net/amm.50-51.262 2011 Trans Tech Publcatons, Swtzerland Solvng two-person zero-sum game by

More information

Review of approximation techniques

Review of approximation techniques CHAPTER 2 Revew of appromaton technques 2. Introducton Optmzaton problems n engneerng desgn are characterzed by the followng assocated features: the objectve functon and constrants are mplct functons evaluated

More information

Tsinghua University at TAC 2009: Summarizing Multi-documents by Information Distance

Tsinghua University at TAC 2009: Summarizing Multi-documents by Information Distance Tsnghua Unversty at TAC 2009: Summarzng Mult-documents by Informaton Dstance Chong Long, Mnle Huang, Xaoyan Zhu State Key Laboratory of Intellgent Technology and Systems, Tsnghua Natonal Laboratory for

More information

Incremental Learning with Support Vector Machines and Fuzzy Set Theory

Incremental Learning with Support Vector Machines and Fuzzy Set Theory The 25th Workshop on Combnatoral Mathematcs and Computaton Theory Incremental Learnng wth Support Vector Machnes and Fuzzy Set Theory Yu-Mng Chuang 1 and Cha-Hwa Ln 2* 1 Department of Computer Scence and

More information

Learning-Based Top-N Selection Query Evaluation over Relational Databases

Learning-Based Top-N Selection Query Evaluation over Relational Databases Learnng-Based Top-N Selecton Query Evaluaton over Relatonal Databases Lang Zhu *, Wey Meng ** * School of Mathematcs and Computer Scence, Hebe Unversty, Baodng, Hebe 071002, Chna, zhu@mal.hbu.edu.cn **

More information

CS246: Mining Massive Datasets Jure Leskovec, Stanford University

CS246: Mining Massive Datasets Jure Leskovec, Stanford University CS46: Mnng Massve Datasets Jure Leskovec, Stanford Unversty http://cs46.stanford.edu /19/013 Jure Leskovec, Stanford CS46: Mnng Massve Datasets, http://cs46.stanford.edu Perceptron: y = sgn( x Ho to fnd

More information

Edge Detection in Noisy Images Using the Support Vector Machines

Edge Detection in Noisy Images Using the Support Vector Machines Edge Detecton n Nosy Images Usng the Support Vector Machnes Hlaro Gómez-Moreno, Saturnno Maldonado-Bascón, Francsco López-Ferreras Sgnal Theory and Communcatons Department. Unversty of Alcalá Crta. Madrd-Barcelona

More information

Problem Set 3 Solutions

Problem Set 3 Solutions Introducton to Algorthms October 4, 2002 Massachusetts Insttute of Technology 6046J/18410J Professors Erk Demane and Shaf Goldwasser Handout 14 Problem Set 3 Solutons (Exercses were not to be turned n,

More information

Lecture #15 Lecture Notes

Lecture #15 Lecture Notes Lecture #15 Lecture Notes The ocean water column s very much a 3-D spatal entt and we need to represent that structure n an economcal way to deal wth t n calculatons. We wll dscuss one way to do so, emprcal

More information

A Bilinear Model for Sparse Coding

A Bilinear Model for Sparse Coding A Blnear Model for Sparse Codng Davd B. Grmes and Rajesh P. N. Rao Department of Computer Scence and Engneerng Unversty of Washngton Seattle, WA 98195-2350, U.S.A. grmes,rao @cs.washngton.edu Abstract

More information

B.N.Jagadesh* et al. /International Journal of Pharmacy & Technology

B.N.Jagadesh* et al. /International Journal of Pharmacy & Technology ISS: 0975-766X CODE: IJPTFI Avalable Onlne through Research Artcle www.jptonlne.com A STATISTICAL APPROACH FOR SKI COLOUR SEGMETATIO USIG HIERARCHICAL CLUSTERIG B..Jagadesh*, A. V. S.. Murty Department

More information

Intelligent Information Acquisition for Improved Clustering

Intelligent Information Acquisition for Improved Clustering Intellgent Informaton Acquston for Improved Clusterng Duy Vu Unversty of Texas at Austn duyvu@cs.utexas.edu Mkhal Blenko Mcrosoft Research mblenko@mcrosoft.com Prem Melvlle IBM T.J. Watson Research Center

More information

A Deflected Grid-based Algorithm for Clustering Analysis

A Deflected Grid-based Algorithm for Clustering Analysis A Deflected Grd-based Algorthm for Clusterng Analyss NANCY P. LIN, CHUNG-I CHANG, HAO-EN CHUEH, HUNG-JEN CHEN, WEI-HUA HAO Department of Computer Scence and Informaton Engneerng Tamkang Unversty 5 Yng-chuan

More information

Human Face Recognition Using Generalized. Kernel Fisher Discriminant

Human Face Recognition Using Generalized. Kernel Fisher Discriminant Human Face Recognton Usng Generalzed Kernel Fsher Dscrmnant ng-yu Sun,2 De-Shuang Huang Ln Guo. Insttute of Intellgent Machnes, Chnese Academy of Scences, P.O.ox 30, Hefe, Anhu, Chna. 2. Department of

More information

5 The Primal-Dual Method

5 The Primal-Dual Method 5 The Prmal-Dual Method Orgnally desgned as a method for solvng lnear programs, where t reduces weghted optmzaton problems to smpler combnatoral ones, the prmal-dual method (PDM) has receved much attenton

More information

Parallel matrix-vector multiplication

Parallel matrix-vector multiplication Appendx A Parallel matrx-vector multplcaton The reduced transton matrx of the three-dmensonal cage model for gel electrophoress, descrbed n secton 3.2, becomes excessvely large for polymer lengths more

More information

Fitting & Matching. Lecture 4 Prof. Bregler. Slides from: S. Lazebnik, S. Seitz, M. Pollefeys, A. Effros.

Fitting & Matching. Lecture 4 Prof. Bregler. Slides from: S. Lazebnik, S. Seitz, M. Pollefeys, A. Effros. Fttng & Matchng Lecture 4 Prof. Bregler Sldes from: S. Lazebnk, S. Setz, M. Pollefeys, A. Effros. How do we buld panorama? We need to match (algn) mages Matchng wth Features Detect feature ponts n both

More information

Investigating the Performance of Naïve- Bayes Classifiers and K- Nearest Neighbor Classifiers

Investigating the Performance of Naïve- Bayes Classifiers and K- Nearest Neighbor Classifiers Journal of Convergence Informaton Technology Volume 5, Number 2, Aprl 2010 Investgatng the Performance of Naïve- Bayes Classfers and K- Nearest Neghbor Classfers Mohammed J. Islam *, Q. M. Jonathan Wu,

More information

Topology Design using LS-TaSC Version 2 and LS-DYNA

Topology Design using LS-TaSC Version 2 and LS-DYNA Topology Desgn usng LS-TaSC Verson 2 and LS-DYNA Wllem Roux Lvermore Software Technology Corporaton, Lvermore, CA, USA Abstract Ths paper gves an overvew of LS-TaSC verson 2, a topology optmzaton tool

More information

Range images. Range image registration. Examples of sampling patterns. Range images and range surfaces

Range images. Range image registration. Examples of sampling patterns. Range images and range surfaces Range mages For many structured lght scanners, the range data forms a hghly regular pattern known as a range mage. he samplng pattern s determned by the specfc scanner. Range mage regstraton 1 Examples

More information

A Fast Content-Based Multimedia Retrieval Technique Using Compressed Data

A Fast Content-Based Multimedia Retrieval Technique Using Compressed Data A Fast Content-Based Multmeda Retreval Technque Usng Compressed Data Borko Furht and Pornvt Saksobhavvat NSF Multmeda Laboratory Florda Atlantc Unversty, Boca Raton, Florda 3343 ABSTRACT In ths paper,

More information

SHAPE RECOGNITION METHOD BASED ON THE k-nearest NEIGHBOR RULE

SHAPE RECOGNITION METHOD BASED ON THE k-nearest NEIGHBOR RULE SHAPE RECOGNITION METHOD BASED ON THE k-nearest NEIGHBOR RULE Dorna Purcaru Faculty of Automaton, Computers and Electroncs Unersty of Craoa 13 Al. I. Cuza Street, Craoa RO-1100 ROMANIA E-mal: dpurcaru@electroncs.uc.ro

More information