636 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 44, NO. 5, MAY 2014

Size: px
Start display at page:

Download "636 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 44, NO. 5, MAY 2014"

Transcription

1 636 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 44, NO. 5, MAY 2014 Constraint Neighborhood Projetions for Sei-Supervised Custering Hongjun Wang, Tao Li, Tianrui Li, and Yan Yang Abstrat Sei-supervised ustering ais to inorporate the known prior knowedge into the ustering agorith. Pairwise onstraints and onstraint projetions are two popuar tehniques in sei-supervised ustering. However, both of the ony onsider the given onstraints and do not onsider the neighbors around the data points onstrained by the onstraints. This paper presents a new tehnique by utiizing the onstrained pairwise data points and their neighbors, denoted as onstraint neighborhood projetions that requires fewer abeed data points (onstraints) and an naturay dea with onstraint onfits. It inudes two steps: 1) the onstraint neighbors are hosen aording to the pairwise onstraints and a given radius so that the pairwise onstraint reationships an be extended to their neighbors, and 2) the origina data points are projeted into a new ow-diensiona spae earned fro the pairwise onstraints and their neighbors. A CNP-Keans agorith is deveoped based on the onstraint neighborhood projetions. Extensive experients on University of Caifornia Irvine (UCI) datasets deonstrate the effetiveness of the proposed ethod. Our study aso shows that onstraint neighborhood projetions (CNP) has soe favorabe features opared with the previous tehniques. Index Ters Constraint neighborhood projetions (CNP), pairwise onstraints, sei-supervised ustering. I. Introdution IN MANY situations when we disover new patterns using ustering, there exists known prior knowedge about the probe. We wish to inorporate the knowedge into the ustering agorith. Reenty, sei-supervised ustering (i.e., ustering with knowedge-based onstraints) has eerged as an iportant variant of the traditiona ustering paradig [1], [2] Current sei-supervised ustering ethods fous on the use of bakground inforation in the for of instane eve Manusript reeived Juy 17, 2012; revised January 19, 2013; aepted May 3, Date of pubiation January 9, 2014; date of urrent version Apri 11, This work was supported in part by the NSFC under Grants , , , and , the State Key Laboratory of Hydrauis and Mountain River Engineering, Sihuan University under Grant 1010, the Fundaenta Researh Funds for the Centra Universities under Grant SWJTU12CX092, and the Constrution Pan for Sientifi Researh Innovation Groups of Leshan Nora University. This paper was reoended by Assoiate Editor P. P. Angeov. H. Wang, T. Li, and Y. Yang are with the Shoo of Inforation Siene and Tehnoogy, Southwest Jiaotong University, Chengdu, Ihuan , China (e-ai: wanghongjun@swjtu.edu.n; tri@swjtu.edu.n; yyang@swjtu.edu.n). T. Li is with the Shoo of Coputer Siene, Forida Internationa University, Miai, FL USA (e-ai: taoi@s.fiu.edu). Coor versions of one or ore of the figures in this paper are avaiabe onine at Digita Objet Identifier /TCYB Fig. 1. Constraint projetion. (a) An exape dataset with onstraints. (b) Resuts of onstraint projetion. ust-ink (ML) and annot-ink (CL) onstraints. An ML onstraint enfores that two instanes ust be paed in the sae uster whie a CL onstraint enfores that two instanes ust not be paed in the sae uster [3]. Typia sei-supervised ustering ethods based on onstraints inude onstrained opete-ink [4], Constrained EM [5], HMRFKeans [6], MPCKeans [7], Kerne ethods [8], [9], [10], [11], [12], [13], [14], and onstraint projetion [15], [16]. Agoriths by axiizing onstraint argin are iustrated in detai by Wang [17] and Zeng et a. [18], [19], respetivey. Spetra agoriths-based parewise onstraints to earn distane etris are stated by the authors in [20], [21]. Pairwise onstraint is aso used for feature seetion [22]. Most existing sei-supervised ustering ethods ony ensure that the given onstraints are onsidered, i.e., pairs of data points with ML (or CL) are not aoated to different usters (or sae usters) or pairs of data points with ML (or CL) are ose to (or distant fro) eah other in the transfored spae or with the earned siiarity easure. They do not onsider the neighbors around the data points onstrained by the given ust/annot-inks. As a resut, in order to ensure the perforane of sei-supervised ustering, generay, a arge nuber of onstraints are needed. There are soe researh efforts reported on weighting propagated onstraints with a Gaussian funtion [23], [24]. However, these ethods are opex and annot work we if the data set is disrete. In addition, they use ots of onstraints (e.g., onstraints were used in their experients whie the nuber of instanes in Iris dataset is ony 150). In this paper, the key otivation is ess training abes and better resuts that are aso the goa of sei-supervised earning. We propose a sei-supervised ustering ethod based on onstraint neighborhood projetions (CNP), where the onstrained pairwise data points and their neighbors are IEEE. Persona use is peritted, but repubiation/redistribution requires IEEE perission. See standards/pubiations/rights/index.ht for ore inforation.

2 WANG et a.: CONSTRAINT NEIGHBORHOOD PROJECTIONS FOR SEMI-SUPERVISED CLUSTERING 637 Fig. 2. Exape iustrates onstraint neighborhood projetions that has four steps. (a) Exape dataset with onstraints. (b) Seeted neighbors. () Neighbor onstraints. (d) Resuts of projetion. used to transfor the input data into a ow-diensiona spae. The proposed ethod requires fewer abeed data points for sei-supervised earning and an naturay dea with the onstraint onfits. Consequenty, our proposed ethod has better generaization apabiity and ore fexibiity than soe state-of-the-art ethods. The rest of the paper is organized as foows. In Setion II, we introdue CNP in detai. A sei-supervised ustering ethod based on CNP is aso proposed. Experienta resuts are presented in Setion III. The paper ends with onusion in Setion IV. II. Constraint Neighborhood Projetions In onstraint projetions (CP), ony the given onstraints are used to deterine the ow-diensiona spae [15], [16]. In our proposed CNP, both the given onstraints and the neighbors of the onstraint points are used to find the transforation. An exape (Figs. 1 and 2) is epoyed to iustrate the differene between CNP and CP. Fig. 1 shows a sape dataset with two ML and one CL onstraints. The dataset is generated fro 2-D Gaussian distributions and is divided into two different usters: the seven points on the eft side of the figure beong to one uster and the eight points on the right side beong to another uster. Fig. 1 presents the resuts of CP. It an be observed fro Fig. 1 that the given onstraints do not propagate to the neighbors in the projetions. Fig. 2 presents the proess of CNP. The origina data points and onstraints are shown in Fig. 2. Aording to the neighborhood size, the seeted onstraint neighbors are shown in Fig. 2. Various derived onstraints of the seeted neighbor points are shown in Fig. 2. Finay, the projetion using a the onstraints (inuding both the input onstraints and the derived onstraints) is shown in Fig. 2. Coparing the two ethods, CNP needs fewer onstraints to ahieve siiar earning outoes as CP. Fig. 3 presents the resuts of CNP on the Wine dataset fro UCI data repository. Fro the figure, it is ear that the uster strutures of data points beoe ore evident after CNP as there are fewer overapping points in the projeted spae. A. Probe Foruation ML onstraints speify that two data points have to be in the sae uster. ML onstraints are transitive [25]. Let ki and Fig. 3. Resuts on the Wine dataset. Different oors are used to show different usters and an ova is used to approxiatey represent the range of a uster. (a) Origina datasets. (b) CNP Resuts. kj be two onneted oponents, and et xi and xj be the instanes in ki and kj, respetivey. Let M be the set of ML onstraints. Then, (xi, xj ) M, xi ki, xj kj (a, b) M, a, b : a ki, b kj. CL onstraints speify that two instanes ust be paed in two different usters and CL onstraints an aso be entaied. Let ki and kj be onneted oponents (opetey onneted subgraphs by ML onstraints), and et xi and xj be the instanes in ki and kj, respetivey. Let C be the set of CL onstraints. Then (xi, xj ) C, xi ki, xj kj (a, b) C, a, b : a ki, b kj. Given a set of p diensiona data X = {x1, x2,..., xn }, the orresponding pairwise ML onstraint set is M = {(xi, xj ) xi ki ; xj ki }, the pairwise CL onstraint set is C = {(xi, xj ) xi ki ; xj kj ; ki = kj }, and the neighbors set of xi is μ = {x ρ xi x 2 ; x X; = (1,..., n)}. CP seeks a set of projetive vetors W = [w1, w2,..., wd ], suh that the pairwise onstraints in C and M are ost faithfuy

3 638 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 44, NO. 5, MAY 2014 preserved in the transfored ower-diensiona representations z i = W T x i. That is, data points invoved by M shoud be ose whie data points invoved by C shoud be far away fro eah other in the ower-diensiona spae. We define the objetive funtion as axiizing J(W) w.r.t. W T W = I, where ( 1 J(W) = W T (x i x j ) 2 + 2n (x i,x j ) C 1 W T (x i x j ) 2) 2n μ (x i,x j ) μ ( 1 W T (x i x j ) 2 + 2n (x i,x j ) M 1 W T (x i x j ) 2). (1) 2n μ (x i,x j ) μ The objetive funtion in (1) an be reforuated in a ore onvenient way as J(W) = 1 2n μ W T (x i x j ) 2 (x i,x j ) C μ 1 2n μ W T (x i x j ) 2, (2) where C μ and M μ are the sets of C and M with their neighbors, respetivey. B. Inferene Ceary, the probe in (2) an be foruated as a typia Eigen-probe. It an be effiienty soved by oputing the eigenvetors and their orresponding eigenvaues. For onveniene, we define S C and S M as foows: S C = 1 2n μ W T (x i x j ) 2, (3) and S M = 1 2n μ W T (x i x j ) 2. (4) Fro (2), to axiize J(W), S C shoud be axiized whie S M shoud be iniized. There exists an anaytia soution to the optiization probe of finding the optia projetion atrix W in (3). We have S C = 1 2n μ W T (x i x j ) 2 = 1 2n μ = 1 2n μ = 1 2n μ W T W T (x 1 x 2 )(x 1 x 2 ) T W (x 1 x 2 )(x 1 x 2 ) T W W T (CC T )W = 1 2n μ W T W,. (5) Aording to (5), S M is S M = 1 2n μ W T W. (6) Fro (5) and (6), the objetion funtion now beoes J(W) = 1 2n μ W T W 1 2n μ W T W = ϑ W T ( ) W W T ( )W. (7) Using the Lagrange utipier optiization tehnique, the Lagrangian an be represented as k L W1,...,W k = J(W 1,..., W k ) δ (W T W 1). (8) By taking the partia derivative of L W1,...,W k with respet to eah W and setting it to zero, we obtain L =2 W 2δ W =0, =1,..., k W =1 W = δ W, =1,..., k. (9) It is ear fro (9) that the soution W is an eigenvetor of and δ is the orresponding eigenvaue of. To axiize J, W ust be the first k eigenvetors of that akes J the su of the k argest eigenvaues of. Suppose W =[W 1,W 2,..., W d ] is the soution to (9), and the orresponding eigenvaues are γ 1 γ 2,..., γ d. Denote = diag(γ 1,γ 2,..., γ d ). Then, the optiization probe beoes hoosing i to Maxiize( γ i ). (10) i Fro (10), if the non-negative eigenvaues of γ i are hosen for the su, it is axiized. After obtaining the onstraint-guided feature projetion as desribed above, we an represent the origina instanes in a ow-diensiona spae that onfors to the uster inforation given in the for of pairwise onstraints. If we add a saing paraeter γ to adjust ratio of onstraints, (2) an be represented as foows: J(W) = 1 2n μ W T x i W T x j 2 γ 2n μ W T x i W T x j 2. (11) Sine, the distane between data points in the sae uster is typiay saer than those in different usters, a saing paraeter is used to baane the ontributions of the two ters and its vaue an be estiated by γ = nμ x i x j 2 x i x j. (12) 2 n μ The objet funtion ontains two ters. One is to ake the average distane in the ower-diensiona spae between instanes invoved by the annot-ink set C as arge as possibe, whie the other is to et the average distanes between

4 WANG et a.: CONSTRAINT NEIGHBORHOOD PROJECTIONS FOR SEMI-SUPERVISED CLUSTERING 639 instanes invoved by the ust-ink set M as sa as possibe. Fro (12), γ > 1 shows that the ust-ink onstraints ontribute the objetion funtion ore and vise versa. At ast γ an be set aording to requireent and an aso be estiated by the two ters of the objet funtion. C. Constraint Confits If a pair (x a,x b ) beongs to C μ and M μ, we a it as a onstraint onfit. CNP an naturay hande this probe. If (x a,x b ) is a onstraint onfit (i.e., it beongs to both C μ and M μ ), we have the foowing objetive funtion J(W) = 1 2n μ W T (x i x j )(x a x b ) }{{} 2 1 2n μ W T (x i x j )(x a x b ) }{{} 2. (13) The ter of (x a x b ) is in both sets of M and C and it is a }{{} onstraint onfit. In order to sove the probe, we suppose (x a x b ) ω, (13) an then be written as J(W) = 1 2n μ W T (x i x j )ω 2 1 2n μ W T (x i x j )ω 2 = ω2 2n μ ω2 2n μ W T (x i x j ) 2 W T (x i x j ) 2. (14) Fro (2) and (14), we find that J(W) J(W) and J(W) = J(W)/ω 2,so J(W) is the sae inferene as J(W). And there are no onstraint onfits in J(W), the onstraint onfits in J(W) are soved by the natura inferene. In another way, ω an be set as a paraeter (e.g., γ in (12)) to baane the ontributions of annot-ink and ust-ink. In this way, the onstraint onfit probe has been addressed. D. Agorith Desription Aording to the above inferene, we design an agorith based on CNP for sei-supervised ustering. The agorith proedure is desried step-by-step as Agorith 1. III. Epiria Study A. Experient Setup In this setion, we perfor experients on 16 rea-word datasets fro UCI ahine earning repository. The nuber of objets, features and asses in eah data set are isted in Tabe II. For evauation, we use iro-preision [26] to Agorith 1: (Constraint Neighborhood Projetions for Sei-supervised K-eans [CNP-Keans]) Input: Training data, {x i,y i } n i=1, where x i is the data points; The set of ust-inks, M; The set of annot-inks, C; The radius of their neighborhood, L. 1) Copute J(w) aording to (5) (7). 2) Copute a of eigenvaues and eigenvetors aording to (8) and (9). 3) Choose the non-negative eigenvaues and orresponding eigenvetors for axiizing J(w). 4) Use eigenvetors orresponding to non-negative eigenvaues to get W. 5) Copute Z = W T X. 6) Run the standard K-eans agorith with input Z. Output: Custer ebership of every point. easure the auray of the uster with respet to the true abes. The iro-preision is defined as MP = k h=1 a h/n, where k is the nuber of usters and n is the nuber of objets, a h denotes the nuber of objets in the uster h that are orrety assigned to the orresponding ass. We identify the orresponding ass for a uster h as the true ass with the argest overap with the uster, and assign a objets in uster h to that ass. Note that 0 MP 1 with 1 indiating the best possibe onsensus ustering, whih has to be in fu agreeent with the ass abes. In our experients, the onstraints are generated as foows: for eah onstraint, one pair of data points are piked out randoy fro exepars of the input data sets (the abes of whih are avaiabe for evauation purpose but unavaiabe for ustering). If the abes of this pair of points are the sae, then a ML onstraint is generated. If the abes are different, a CL onstraint is generated. The aounts of onstraints are deterined by the size of input data. On a the datasets, the experients are perfored 20 ties to eiinate the differene aused by onstraints. B. Perforane Coparison To deonstrate how our ethod works for sei-supervised ustering probe and iproves the ustering perforane, we opare the foowing ethods. 1) MPC-Keans [6]: perforing MPC-Keans on the origina datasets with the nuber of onstraints in Tabe I. 2) CP-Keans [15]: perforing K-eans after onstraints projetions with the nuber of onstraints in Tabe I. 3) CNP-Keans: perforing CNP-Keans after onstraint neighbor projetions with the nuber of onstraints in Tabe I. 4) PCA-Keans: Prinipa oponent anaysis (PCA) is first appied to redue data diensionaity foowed by perforing K-eans ustering. 5) LDA-Keans [27]: adaptive data ustering by integrating K-eans ustering and inear disriinant anaysis

5 640 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 44, NO. 5, MAY 2014 TABLE I Average Auraies Resuts of the Experients Dataset MPC-Keans CP-Keans CNP-Keans PCA-Keans GP-Keans LDA-Keans pia ± ± ± ± ± ± iris ± ± ± ± ± ± wdb ± ± ± ± ± ± wine ± ± ± ± ± ± ionosphere ± ± ± ± ± ± gass ± ± ± ± ± ± bupa ± ± ± ± ± ± baane ± ± ± ± ± ± Kdd99sub ± ± ± ± ± ± spetheart ± ± ± ± ± ± Hepatitis ± ± ± ± ± ± donationsub ± ± ± ± ± ± MiniBooNE ± ± ± ± ± ± bank ± ± ± ± ± ± shutte ± ± ± ± ± ± agi ± ± ± ± ± ± Average ± ± ± ± ± ± TABLE II Nuber of the Instanes, Features, Casses and Constraints Used in Eah Dataset Dataset Charateristi Instanes Features Categories Constraints pia rea iris rea wdb rea wine rea ionosphere rea gass rea bupa disrete baane disrete kdd99sub disrete spetheart rea hepatitis rea donationsub1 ixed iniboone rea bank ixed shutte rea agi rea (LDA) where K-eans ustering is used to generate uster abes and LDA is used to perfor subspae seetion. 6) GP-Keans [23]: perforing GP-Keans on the origina datasets with the nuber of onstraints in Tabe I. The perforane oparison is shown in Tabe I. We an see that CNP-Keans ahieves the best average MP with the vaue of , and CNP-Keans outperfors other ustering agoriths for ost of the ties, whih is obviousy shown on the arge datasets. Aong a the agoriths on the 16 datasets, it is 11 ties that CNP-Keans ahieves the best resuts, whie the other five agoriths ony ahieve six best resuts. To ake a arefu oparison of these agoriths, we perfor a 1 n oparison by eans of the aigned Friedan test [28]. The proposed ethod CNP-Keans is the ontro ethod. Tabe III dispays the aigned observations and the aigned ranks in the parentheses onsidering the known 6 agoriths and 16 data sets. On average, CNP-Keans ranks the first with a rank of ; CP-Keans ranks the seond with a vaue of ; GP-Keans ranks the third with ; PCA-Keans ranks the fourth with ; LDA- Keans ranks the fifth; and MPC-Keans ranks the ast. The Friedan aigned rank test heks whether the easured su of aigned ranks is signifianty different fro the tota aigned ranks R j = 776 expeted under the nu hypothesis: k R 2.,j = j= = , k R 2 i,. = j=1 = , {(6 1)( ( /4)(6 16+1) 2 )} T = ((6 16( )( )) )/6 = With six agoriths and 16 data sets, T is distributed aording to the hi-square distribution with 6 1 = 5 degrees of freedo. The p-vaue oputed by using the χ 2 (5) distribution are (for one taied test) and (for two taied test). Then, the nu hypothesis is rejeted at a high eve of signifiane. We an see the vaues are far ess than 0.05 whih shows that the resuts of agoriths are signifianty different. C. Paraeter Tuning We aso report the experient resuts on the radius of onstraint neighborhood. In this experient, the sae set of onstraints is used but the radius of their neighborhoods is inreasing graduay. In both Figs. 4 and 5, the x oordinate is defined as radius =, where L is the distane MAX(L) aong any two data points. Aong the 16 datasets, ost

6 WANG et a.: CONSTRAINT NEIGHBORHOOD PROJECTIONS FOR SEMI-SUPERVISED CLUSTERING 641 TABLE III Aigned Observations of Six Agoriths Seeted in the Experienta Study Dataset pia iris wdb wine ionosphere gass bupa baane Kdd99sub spetheart Hepatitis donationsub1 iniboone bank shutte agi Tota Average rank MPC-Keans (87) (66.5) (2) (73.5) (52) (56) (85) (90.5) (15) (70.5) (59) (63) (48.5) (90.5) (54) CP-Keans (62) (50.5) (96) (3) (29) (19) (8) (94) (18) (42) (43) (58) (39) (57) (45) CNP-Keans (33) (37) (93) (73.5) (20.5) (39) (7) (1) (20.5) (30) (16) (10) (23) (6) (11) PCA-Keans (33) (50.5) (95) (82) (64) (39) (86) (12.5) (81) (70.5) (36) (44) (48.5) (9) (54) GP-Keans (33) (66.5) (4.5) (77.5) (46) (41) (83) (88) (22) (31) (61) (65) (47) (75) (54) LDA-Keans (14) (17) (92) (4.5) (77.5) (79) (84) (69) (12.5) (89) (35) (72) (60) (80) (68) (76) Tota The ranks in the parentheses are used in the oputation of the friedan aigned ranks test. The saest one is the best. Fig. 4. Experienta resuts with the inreasing radius of onstraint neighborhood. In this experient the sae onstraints are used but the radius of their neighborhood are inreasing graduay.

7 642 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 44, NO. 5, MAY 2014 of a the pairwise onstraints. It is ear that using ony about 1% of a the pairwise onstraints the axiu MPs are obtained in the CNP, whie CP requires about 10% of a the onstraints. In other words, CP requires as any as 10 ties the nuber of onstraints needed in CNP. Therefore, one iportant feature of our proposed CNP is that it requires ess abeed inforation. IV. Conusion Fig. 5. Average MP of 16 datasets with the inreasing radius of onstraint neighborhood. In this experient, the sae onstraints are used but the radius of their neighborhood are inreasing graduay. In this paper, a nove approah, CNP, was proposed for sei-supervised ustering. To the best of our knowedge, this is the first work in sei-supervised ustering on using pairwise onstraints and their neighbors for projetion to a ow-diensiona spae. CNP uses onstraint neighborhoods to guide the projetion. The proposed ethod opares favoraby over severa state-of-the-art ethods. In other words, it requires fewer onstraints and an aso dea with onstraint onfits. Based on CNP, we design a CNP-Keans agorith and the experienta resuts deonstrate its effetiveness. In our future work, we wi fous on the investigation of better onstraints seetion and radius seetion for sei-supervised ustering. Aknowedgent The authors woud ike to thank the anonyous reviewers for their hepfu oents and suggestions. Referenes Fig. 6. Max auray vs. nuber of onstraints. (a the datasets exept two of Bupa and Ionosphere) of their ustering auray resuts are graduay inreasing with the inreasing size of the radius. In genera, the auray wi reah axiu when the radius reahes 0.2. However, when the radius ontinues to inrease, the auray resuts wi draatiay deine. This is beause when the radius beoes arge to soe degree, the ust-ink onstraint neighborhood ay overap with the annot-ink onstraint neighborhood eah other, and the orresponding neighborhood ay ontain data points fro different usters and thus generate soe onfiting onstraints. As a resut, the ustering perforane ay be degraded. In genera, the auray of ustering inreases with the radius beoing arge between 0 and 0.2, whie the auray dereases with the radius being ore than o.2. D. Nuber of Constraints Under Maxiu MP In this subsetion, we report the resuts on the nuber of onstraints that are needed in the CP and CNP to ahieve the axiu MPs. The resuts suarizing over the 16 datasets are shown in Fig. 6 where the Y -axis is ranging fro the saest MP to the argest and the X-axis shows the perentage [1] S. Basu, I. Davidson, and K. L. Wagstaff, Constrained Custering. Boa Raton, FL, USA: CRC Press, [2] O. Chapee, A. Zien, and B. Shokopf, Sei-Supervised Learning. Cabridge, MA, USA: MIT Press, [3] K. Wagstaff, C. Cardie, S. Rogers, and S. Shroed, Constrained Keans ustering with bakground knowedge, in Pro. ICML, 2001, pp [4] D. Kein, S. D. Kavar, and C. D. Manning, Fro instane-eve onstraints to spae-eve onstraints: Making the ost of prior knowedge in data ustering, in Pro. ICML, 2002, pp [5] N. Shenta, A. Bar-Hie, T. Hertz, and D. Weinsha, Coputing gaussian ixture odes with e using equivaene onstraints, in Pro. NIPS, [6] S. Basu, M. Bienko, and R. J. Mooney, A probabiisti fraework for sei-supervised ustering, in Pro. KDD, 2004, pp [7] M. Bienko, S. Basu, and R. J. Mooney, Integrating onstraints and etri earning in sei-supervised ustering, in Pro. ICML, 2004, pp [8] B. Kuis, S. Basu, I. Dhion, and R. J. Mooney, Seisupervisedgraphustering: A kerne approah, in Pro. ICML, 2005, pp [9] B. Yan and C. Doenioni, An adaptive kerne ethod for seisupervised ustering, in Pro. ECML, 2006, pp [10] D. Y. Yeung and H. Chang, A kerne approah for sei-supervised etri earning, IEEE Trans. Neura Netw., vo. 18, no. 1, pp , Ju [11] S. C. H. Hoi, R. Jin, M. R. Lyu, and J. Wu, Learning nonparaetri kerne atries fro pairwise onstraints, in Pro. ICML, 2007, pp [12] O. Masayuki and Y. Seiji, Learning siiarity atrix fro onstraints of reationa neighbors, J. Adv. Coput. Inte. Inte. Infor., vo. 14, no. 4, pp , [13] X. Yin, S. Chen, and E. H. D. Zhang, Sei-supervised ustering with etri earning: An adaptive kerne ethod, Pattern Reognit., vo. 43, no. 4, pp , 2010.

8 WANG et a.: CONSTRAINT NEIGHBORHOOD PROJECTIONS FOR SEMI-SUPERVISED CLUSTERING 643 [14] C. Doenioni, J. Peng, and B. Yan, Coposite kernes for seisupervised ustering, Know. Infor. Syst., vo. 28, pp , Aug [15] W. Tang, H. Xiong, S. Zhong, and J. Wu, Enhaning sei-supervised ustering: A feature projetion perspetive, in Pro. KDD, 2007, pp [16] D. Zhang, S. Chen, Z. Zhou, and Q. Yang, Constraint projetions for ensebe earning, in Pro. AAAI, 2008, pp [17] F. Wang, Seisupervised etri earning by axiizing onstraint argin, IEEE Trans. Neura Netw., vo. 41, no. 4, pp , Aug [18] H. Zeng and Y.-M. Cheung, Sei-supervised axiu argin ustering with pairwise onstraints, IEEE Trans. Know. Data Eng., vo. 24, no. 5, pp , May [19] A. Mignon and F. Jurie, Pa: A new approah for distane earning fro sparse pairwise onstraints, in Pro. CVPR, 2012, pp [20] W. Chen and G. Feng, Spetra ustering: A sei-supervised approah, Neurooputing, vo. 77, no. 1, pp , Jan [21] F. Shang, Y. Liu, and F. Wang, Learning spetra ebedding for seisupervised ustering, in Pro. ICDM, 2011, pp [22] D. Zhang, S. Chen, and Z. Zhou, Constraint sore: A new fiter ethod for feature seetion with pairwise onstraints, Pattern Reognit., vo. 41, no. 5, pp , May [23] E. R. Eaton, Custering with Propagated Constraints, Thesis, Univ. Maryand, Coege Park, MD, USA, [24] J. Huang and H. Sun, Lighty-supervised ustering using pairwise onstraint propagation, in Pro. 3rd Int. Conf. Inte. Syst. Know. Eng., 2008, pp [25] Z. Li, J. Liu, and X. Tang, Pairwise onstraint propagation by seidefinite prograing for sei-supervised assifiation, in Pro. ICML, 2008, pp [26] Z. H. Zhou and W. Tang, Custerer ensebe, Knowedge-Based Syst., vo. 19, no. 1, pp , [27] C. Ding and T. Li, Adaptive diension redution using disriinant anaysis and K-eans ustering, in Pro. ICML, 2007, pp [28] S. Gara, A. Fernndez, J. Luengo, and F. Herrera, Advaned nonparaetri tests for utipe oparisons in the design of experients in oputationa inteigene and data ining: Experienta anaysis of power, Infor. Si., vo. 180, no. 10, pp , May Tao Li reeived the Ph.D. degree in oputer siene fro the Departent of Coputer Siene, University of Rohester, Rohester, NY, USA, in He is urrenty an Assoiate Professor at the Shoo of Coputing and Inforation Sienes, Forida Internationa University, Miai, FL, USA. His urrent researh interests inude data ining, oputing syste anageent, inforation retrieva, and ahine earning. Dr. Li was a reipient of NSF CAREER Award and utipe IBM Fauty Researh Awards. Tianrui Li reeived the B.S., M.S., and Ph.D. degrees fro Southwest Jiaotong University, Chengdu, China, in 1992, 1995, and 2002, respetivey. Fro 2005 to 2006, he was a Post-Dotora Researher at SCK*CEN, Begiu, and a Visiting Professor at Hasset University, Begiu, in 2008, and the University of Tehnoogy, Sydney, Austraia, in He is urrenty a Professor and the Diretor at the Key Laboratory of Coud Coputing and Inteigent Tehniques, Southwest Jiaotong University. Sine 2000, he has o-edited three textbooks, ten proeedings, five speia issues of internationa journas and pubished over 100 researh papers in refereed journas and onferenes. His urrent researh interests inude data ining and knowedge disovery, granuar oputing and rough sets, oud oputing, traffi inforation engineering and ontro. Dr. Li has been the Vie Chair of the IEEE CIS Chengdu Chapter sine 2011 and the Area Editor of the Internationa Journa of Coputationa Inteigene Systes (IJCIS, SCI indexed). He has served as ISKE2007, ISKE2008, ISKE2009, ISKE2010, ISKE2011, JRS2012, ISKE2012 progra hairs, IEEE GrC 2009 Progra Vie Chair and RSKT2008, FLINS2010 Organizing Chair. He has been a reviewer for severa eading aadei journas. aadei journas. Hongjun Wang reeived the Ph.D. degree in oputer siene fro Sihuan University, Chengdu, China, in He is urrenty an Assoiate Professor of the Key Laboratory of Coud Coputing and Inteigent Tehniques, Southwest Jiaotong University, Chengdu, China. His urrent researh interests inude ahine earning, data ining, and ensebe earning. He has pubished over 30 researh papers in journas and onferenes and he is a eber of ACM and CCF. He has been a reviewer for severa Yan Yang reeived the Ph.D. degree in traffi inforation engineering and ontro fro Southwest Jiaotong University of Chengdu, China, in She is urrenty a Professor and Assoiate Dean of the Shoo of Inforation Siene and Tehnoogy, Southwest Jiaotong University. Her urrent researh interests inude data ining, oputationa inteigene, and ensebe earning. She has been the Vie Chair of the ACM Chengdu Chapter sine 2012, eber of IEEE-CS sine 2012, and senior eber of CCF sine 2010.

Noise Source Identification Applied in Electric Power Industry Using Microphone Arrays

Noise Source Identification Applied in Electric Power Industry Using Microphone Arrays Engineering, 213, 5, 152-156 doi:1.4236/eng.213.51b28 Pubished Onine January 213 (http://www.sirp.org/journa/eng Noise Soure Identifiation Appied in Eetri Power Industry Using Mirophone Arrays Pengxiao

More information

A REGULARIZED GMRES METHOD FOR INVERSE BLACKBODY RADIATION PROBLEM

A REGULARIZED GMRES METHOD FOR INVERSE BLACKBODY RADIATION PROBLEM Wu, J., Ma. Z.: A Reguarized GMRES Method for Inverse Backbody Radiation Probe THERMAL SCIENCE: Year 3, Vo. 7, No. 3, pp. 847-85 847 A REGULARIZED GMRES METHOD FOR INVERSE BLACKBODY RADIATION PROBLEM by

More information

Optimizing Single Sweep Range and Doppler Processing for FMCW Radar using Inverse Filtering

Optimizing Single Sweep Range and Doppler Processing for FMCW Radar using Inverse Filtering Optiizing Single Sweep and Doppler Proessing for FMCW Radar using Inverse Filtering AJ de Jong and Ph van Dorp Oude Waalsdorperweg 63 2597 AK, Den Haag The Netherlands ajdejong@feltnonl ABSTRACT We disuss

More information

Laboratory exercise No. 2 Basic material parameters of porous building materials

Laboratory exercise No. 2 Basic material parameters of porous building materials Laboratory exerise No. 2 Basi ateria paraeters of porous buiding aterias Materias (buiding aterias) an be assified aording to the different riteria, e.g. based on their properties, funtion, heia oposition

More information

Approximate dynamic programming using model-free Bellman Residual Elimination

Approximate dynamic programming using model-free Bellman Residual Elimination Approxiate dynaic prograing using ode-free Bean Residua Eiination The MIT Facuty has ade this artice openy avaiabe. Pease share how this access benefits you. Your story atters. Citation As Pubished Pubisher

More information

Research Article Some Applications of Second-Order Differential Subordination on a Class of Analytic Functions Defined by Komatu Integral Operator

Research Article Some Applications of Second-Order Differential Subordination on a Class of Analytic Functions Defined by Komatu Integral Operator ISRN Mathematia Anaysis, Artie ID 66235, 5 pages http://dx.doi.org/1.1155/214/66235 Researh Artie Some Appiations of Seond-Order Differentia Subordination on a Cass of Anayti Funtions Defined by Komatu

More information

A New Method of Transductive SVM-Based Network Intrusion Detection

A New Method of Transductive SVM-Based Network Intrusion Detection A New Method of Transductive SVM-Based Network Intrusion Detection Manfu Yan and Zhifang Liu 2 Departent of Matheatics, Tangshan Teacher s Coege, Tangshan Hebei, China 3005@tstc.edu.cn 2 Network Technoogy

More information

Estimating Mutual Information Using Gaussian Mixture Model for Feature Ranking and Selection

Estimating Mutual Information Using Gaussian Mixture Model for Feature Ranking and Selection Estiating utual Inforation Using Gaussian ixture odel for Feature Ranking and Seletion Tian Lan, Deniz Erdogus, Uut Ozerte, Yonghong Huang Abstrat Feature seletion is a ritial step for pattern reognition

More information

Design of Output Feedback Compensator

Design of Output Feedback Compensator Design of Output Feedbak Copensator Vanita Jain, B.K.Lande Professor, Bharati Vidyapeeth s College of Engineering, Pashi Vihar, New Delhi-0063 Prinipal, Shah and Anhor Kuthhi Engineering College, Chebur,

More information

SPEECH RECOGNITION USING LPC AND HMM APPLIED FOR CONTROLLING MOVEMENT OF MOBILE ROBOT

SPEECH RECOGNITION USING LPC AND HMM APPLIED FOR CONTROLLING MOVEMENT OF MOBILE ROBOT Seinar asiona Teknoogi Inforasi 200 SPEECH RECOGITIO USIG LPC AD HMM APPLIED FOR COTROLLIG MOVEMET OF MOBILE ROBOT Thiang ) Wanto ) ) Eectrica Engineering Departent Petra Christian university Siwaankerto

More information

AN INVESTIGATION ON SEISMIC ANALYSIS OF SHALLOW TUNEELS IN SOIL MEDIUM

AN INVESTIGATION ON SEISMIC ANALYSIS OF SHALLOW TUNEELS IN SOIL MEDIUM The 4 th October -7, 8, Beijing, China AN INVESTIGATION ON SEISMIC ANALYSIS OF SHALLOW TUNEELS IN SOIL MEDIUM J. Boouri Bazaz and V. Besharat Assistant Professor, Dept. of Civi Engineering, Ferdowsi University,

More information

Sparse Structured Associative Memories as Efficient Set-Membership Data Structures

Sparse Structured Associative Memories as Efficient Set-Membership Data Structures Sparse Strutured Assoiative Memories as Effiient Set-Membership Data Strutures Vinent Gripon Eetronis Department Tééom Bretagne, Brest, Frane vinent.gripon@ens-ahan.org Vitay Skahek Institute of Computer

More information

Involutions and representations of the finite orthogonal groups

Involutions and representations of the finite orthogonal groups Invoutions and representations of the finite orthogona groups Student: Juio Brau Advisors: Dr. Ryan Vinroot Dr. Kaus Lux Spring 2007 Introduction A inear representation of a group is a way of giving the

More information

Derivation of Non-Einsteinian Relativistic Equations from Momentum Conservation Law

Derivation of Non-Einsteinian Relativistic Equations from Momentum Conservation Law Asian Journal of Applied Siene and Engineering, Volue, No 1/13 ISSN 35-915X(p); 37-9584(e) Derivation of Non-Einsteinian Relativisti Equations fro Moentu Conservation Law M.O.G. Talukder Varendra University,

More information

FEATURE SELECTION BASED ON SURVIVAL CAUCHY-SCHWARTZ MUTUAL INFORMATION

FEATURE SELECTION BASED ON SURVIVAL CAUCHY-SCHWARTZ MUTUAL INFORMATION 04 IEEE International Conferene on Aousti, peeh and ignal Proessing (ICAP FEATURE ELECTIO BAED O URVIVAL CAUCHY-CHWARTZ MUTUAL IFORMATIO Badong Chen, iaohan Yang, Hua Qu, Jihong Zhao, anning Zheng, Jose

More information

FRST Multivariate Statistics. Multivariate Discriminant Analysis (MDA)

FRST Multivariate Statistics. Multivariate Discriminant Analysis (MDA) 1 FRST 531 -- Mutivariate Statistics Mutivariate Discriminant Anaysis (MDA) Purpose: 1. To predict which group (Y) an observation beongs to based on the characteristics of p predictor (X) variabes, using

More information

DETERMINATION OF TWO LAYER EARTH STRUCTURE PARAMETERS

DETERMINATION OF TWO LAYER EARTH STRUCTURE PARAMETERS DETERMINATION OF TWO LAYER EARTH STRUCTURE PARAMETERS Ioannis F. GONOS, Vassiliki T. KONTARGYRI, Ioannis A. STATHOPULOS, National Tehnial University of Athens, Shool of Eletrial and Coputer Engineering,

More information

Alphanumeric Character Recognition

Alphanumeric Character Recognition Amerian Journa of Signa Proessing. 0; (): 34-39 DOI: 0. 593/j.ajsp.000.06 Training Tangent Simiarities with N-SVM for Aphanumeri Charater Reognition Hassiba Nemmour *, Youef Chibani Signa Proessing Laboratory,

More information

Transforms, Convolutions, and Windows on the Discrete Domain

Transforms, Convolutions, and Windows on the Discrete Domain Chapter 3 Transfors, Convoutions, and Windows on the Discrete Doain 3. Introduction The previous two chapters introduced Fourier transfors of functions of the periodic and nonperiodic types on the continuous

More information

Robust Thermal Boundary Condition Using Maxwell- Boltzmann Statistics and its Application

Robust Thermal Boundary Condition Using Maxwell- Boltzmann Statistics and its Application Robust Thera Boundary Condition Using Maxwe- Botzann Statistis and its Aiation Jae Wan Shi a and Renée Gatigno b a Interdisiinary Fusion Tehnoogy Division, Korea Institute of Siene and Tehnoogy, 6-79,

More information

Relationship between the number of labeled samples and classification accuracy based on sparse representation

Relationship between the number of labeled samples and classification accuracy based on sparse representation Relationship between the nuber of labeled saples and lassifiation auray based on sparse representation 1 Shool of Coputer Siene and Engineering, Beifang University for Nationalities,Yinhuan, 75001,China

More information

Maximum Entropy and Exponential Families

Maximum Entropy and Exponential Families Maximum Entropy and Exponential Families April 9, 209 Abstrat The goal of this note is to derive the exponential form of probability distribution from more basi onsiderations, in partiular Entropy. It

More information

Intelligent Systems: Reasoning and Recognition. Perceptrons and Support Vector Machines

Intelligent Systems: Reasoning and Recognition. Perceptrons and Support Vector Machines Intelligent Systes: Reasoning and Recognition Jaes L. Crowley osig 1 Winter Seester 2018 Lesson 6 27 February 2018 Outline Perceptrons and Support Vector achines Notation...2 Linear odels...3 Lines, Planes

More information

Fractional Order Controller for PMSM Speed Servo System Based on Bode s Ideal Transfer Function

Fractional Order Controller for PMSM Speed Servo System Based on Bode s Ideal Transfer Function Sensors & Transduers, Vol. 73, Issue 6, June 24, pp. -7 Sensors & Transduers 24 by IFSA Publishing, S. L. http://www.sensorsportal.o Frational Order Controller for PMSM Speed Servo Syste Based on Bode

More information

Danielle Maddix AA238 Final Project December 9, 2016

Danielle Maddix AA238 Final Project December 9, 2016 Struture and Parameter Learning in Bayesian Networks with Appliations to Prediting Breast Caner Tumor Malignany in a Lower Dimension Feature Spae Danielle Maddix AA238 Final Projet Deember 9, 2016 Abstrat

More information

Optimization of the CBSMAP Queueing Model

Optimization of the CBSMAP Queueing Model July 3-5 23 London UK Optiization of the CBSMAP Queueing Model Kondrashova EV Kashtanov VA Abstrat The present paper is devoted to the researh of ontrolled queueing odels at ontrol of CBSMAP-flow Controlled

More information

A Queueing Model for Call Blending in Call Centers

A Queueing Model for Call Blending in Call Centers A Queueing Model for Call Blending in Call Centers Sandjai Bhulai and Ger Koole Vrije Universiteit Amsterdam Faulty of Sienes De Boelelaan 1081a 1081 HV Amsterdam The Netherlands E-mail: {sbhulai, koole}@s.vu.nl

More information

An Integrated Statistical Model for Multimedia Evidence Combination

An Integrated Statistical Model for Multimedia Evidence Combination An Integrated Statistial Model for Multiedia Evidene Cobination Sheng Gao, Joo-Hwee Li and Qibin Sun Institute for Infoo Researh, 21 Heng Mui Keng Terrae, Singapore, 119613 {gaosheng, joohwee, qibin}@i2r.a-star.edu.sg

More information

A complete set of ladder operators for the hydrogen atom

A complete set of ladder operators for the hydrogen atom A copete set of adder operators for the hydrogen ato C. E. Burkhardt St. Louis Counity Coege at Forissant Vaey 3400 Persha Road St. Louis, MO 6335-499 J. J. Leventha Departent of Physics University of

More information

LONG-TERM PREDICTIVE VALUE INTERVAL WITH THE FUZZY TIME SERIES

LONG-TERM PREDICTIVE VALUE INTERVAL WITH THE FUZZY TIME SERIES Journal of Marine Science and Technology, Vol 19, No 5, pp 509-513 (2011) 509 LONG-TERM PREDICTIVE VALUE INTERVAL WITH THE FUZZY TIME SERIES Ming-Tao Chou* Key words: fuzzy tie series, fuzzy forecasting,

More information

Millennium Relativity Acceleration Composition. The Relativistic Relationship between Acceleration and Uniform Motion

Millennium Relativity Acceleration Composition. The Relativistic Relationship between Acceleration and Uniform Motion Millennium Relativity Aeleration Composition he Relativisti Relationship between Aeleration and niform Motion Copyright 003 Joseph A. Rybzyk Abstrat he relativisti priniples developed throughout the six

More information

FOR many years the authors of this paper have worked on

FOR many years the authors of this paper have worked on The Fast Parametri Integra Equations System for Poygona D Potentia Probems Andrzej Kużeewski and Eugeniusz Zieniuk Abstrat Appiation of tehniques for modeing of boundary vaue probems impies three onfiting

More information

CS229 Lecture notes. Andrew Ng

CS229 Lecture notes. Andrew Ng CS229 Lecture notes Andrew Ng Part IX The EM agorithm In the previous set of notes, we taked about the EM agorithm as appied to fitting a mixture of Gaussians. In this set of notes, we give a broader view

More information

An Algorithm for Pruning Redundant Modules in Min-Max Modular Network

An Algorithm for Pruning Redundant Modules in Min-Max Modular Network An Agorithm for Pruning Redundant Modues in Min-Max Moduar Network Hui-Cheng Lian and Bao-Liang Lu Department of Computer Science and Engineering, Shanghai Jiao Tong University 1954 Hua Shan Rd., Shanghai

More information

Kernel Methods and Support Vector Machines

Kernel Methods and Support Vector Machines Intelligent Systes: Reasoning and Recognition Jaes L. Crowley ENSIAG 2 / osig 1 Second Seester 2012/2013 Lesson 20 2 ay 2013 Kernel ethods and Support Vector achines Contents Kernel Functions...2 Quadratic

More information

Complexity of Regularization RBF Networks

Complexity of Regularization RBF Networks Complexity of Regularization RBF Networks Mark A Kon Department of Mathematis and Statistis Boston University Boston, MA 02215 mkon@buedu Leszek Plaskota Institute of Applied Mathematis University of Warsaw

More information

THE NONLINEAR NATURE OF PREFERENCES, ITS IMPACT ON THE SENSITIVITY AND EFFECTIVENESS OF MULTIPLE CRITERIA ALTERNATIVES

THE NONLINEAR NATURE OF PREFERENCES, ITS IMPACT ON THE SENSITIVITY AND EFFECTIVENESS OF MULTIPLE CRITERIA ALTERNATIVES IJAHP Artile: Mu, Saaty/A Style Guide for Paper Proposals o Be Subitted to the International Syposiu of the Analyti Hierarhy Proess 2014, Washington D.C., U.S.A. HE NONLINEAR NAURE OF PREFERENCES, IS IMPAC

More information

BP neural network-based sports performance prediction model applied research

BP neural network-based sports performance prediction model applied research Avaiabe onine www.jocpr.com Journa of Chemica and Pharmaceutica Research, 204, 6(7:93-936 Research Artice ISSN : 0975-7384 CODEN(USA : JCPRC5 BP neura networ-based sports performance prediction mode appied

More information

Identites and properties for associated Legendre functions

Identites and properties for associated Legendre functions Identites and properties for associated Legendre functions DBW This note is a persona note with a persona history; it arose out off y incapacity to find references on the internet that prove reations that

More information

Using EM To Estimate A Probablity Density With A Mixture Of Gaussians

Using EM To Estimate A Probablity Density With A Mixture Of Gaussians Using EM To Estiate A Probablity Density With A Mixture Of Gaussians Aaron A. D Souza adsouza@usc.edu Introduction The proble we are trying to address in this note is siple. Given a set of data points

More information

Phase Diagrams. Chapter 8. Conditions for the Coexistence of Multiple Phases. d S dt V

Phase Diagrams. Chapter 8. Conditions for the Coexistence of Multiple Phases. d S dt V hase Diaras Chapter 8 hase - a for of atter that is unifor with respect to cheica coposition and the physica state of areation (soid, iquid, or aseous phases) icroscopicay and acroscopicay. Conditions

More information

Numerical Studies of Counterflow Turbulence

Numerical Studies of Counterflow Turbulence Nonae anusript No. will be inserted by the editor Nuerial Studies of Counterflow Turbulene Veloity Distribution of Vorties Hiroyuki Adahi Makoto Tsubota Reeived: date Aepted: date Abstrat We perfored the

More information

Using the Green s Function to find the Solution to the Wave. Equation:

Using the Green s Function to find the Solution to the Wave. Equation: Using the Green s Funtion to find the Soution to the Wave Exampe 1: 2 1 2 2 t 2 Equation: r,t q 0 e it r aẑ r aẑ r,t r 1 r ; r r,t r 1 r 2 The Green s funtion soution is given by r,t G R r r,t t Fr,t d

More information

Weighted K-Nearest Neighbor Revisited

Weighted K-Nearest Neighbor Revisited Weighted -Nearest Neighbor Revisited M. Biego University of Verona Verona, Italy Email: manuele.biego@univr.it M. Loog Delft University of Tehnology Delft, The Netherlands Email: m.loog@tudelft.nl Abstrat

More information

Lightpath routing for maximum reliability in optical mesh networks

Lightpath routing for maximum reliability in optical mesh networks Vol. 7, No. 5 / May 2008 / JOURNAL OF OPTICAL NETWORKING 449 Lightpath routing for maximum reliability in optial mesh networks Shengli Yuan, 1, * Saket Varma, 2 and Jason P. Jue 2 1 Department of Computer

More information

Support Vector Machine Classification of Uncertain and Imbalanced data using Robust Optimization

Support Vector Machine Classification of Uncertain and Imbalanced data using Robust Optimization Recent Researches in Coputer Science Support Vector Machine Classification of Uncertain and Ibalanced data using Robust Optiization RAGHAV PAT, THEODORE B. TRAFALIS, KASH BARKER School of Industrial Engineering

More information

Dual-beard sampling method for fibre length measurements

Dual-beard sampling method for fibre length measurements Indian Journa of Fibre & Textie Research Vo. 39, March 14, pp. 7-78 Dua-beard saping ethod for fibre ength easureents H Y Wu & F M Wang a Coege of Texties, Donghua University, Shanghai, China Received

More information

Some Properties of Interval Quadratic Programming Problem

Some Properties of Interval Quadratic Programming Problem International Journal of Sstes Siene Applied Matheatis 07; (5): 05-09 http://www.sienepublishinggroup.o/j/ijssa doi: 0.648/j.ijssa.07005.5 ISSN: 575-5838 (Print); ISSN: 575-5803 (Online) Soe Properties

More information

Hankel Optimal Model Order Reduction 1

Hankel Optimal Model Order Reduction 1 Massahusetts Institute of Tehnology Department of Eletrial Engineering and Computer Siene 6.245: MULTIVARIABLE CONTROL SYSTEMS by A. Megretski Hankel Optimal Model Order Redution 1 This leture overs both

More information

Structural Design for Vibration Reduction in Brushless DC Stator

Structural Design for Vibration Reduction in Brushless DC Stator J Eletr Eng Tehnol.017; 1(5): 184-1850 http://doi.org/10.5370/jeet.017.1.5.184 ISSN(Print) 1975-010 ISSN(Online) 093-743 Strutural Design for Vibration Redution in Brushless DC Stator Mehrdad Jafarboland

More information

After the completion of this section the student should recall

After the completion of this section the student should recall Chapter I MTH FUNDMENTLS I. Sets, Numbers, Coordinates, Funtions ugust 30, 08 3 I. SETS, NUMERS, COORDINTES, FUNCTIONS Objetives: fter the ompletion of this setion the student should reall - the definition

More information

arxiv:gr-qc/ v2 6 Feb 2004

arxiv:gr-qc/ v2 6 Feb 2004 Hubble Red Shift and the Anomalous Aeleration of Pioneer 0 and arxiv:gr-q/0402024v2 6 Feb 2004 Kostadin Trenčevski Faulty of Natural Sienes and Mathematis, P.O.Box 62, 000 Skopje, Maedonia Abstrat It this

More information

Part B: Many-Particle Angular Momentum Operators.

Part B: Many-Particle Angular Momentum Operators. Part B: Man-Partice Anguar Moentu Operators. The coutation reations deterine the properties of the anguar oentu and spin operators. The are copete anaogous: L, L = i L, etc. L = L ± il ± L = L L L L =

More information

An Approach to Worst-Case Circuit Analysis

An Approach to Worst-Case Circuit Analysis Eena Nicuescu, Dorina-Mioara Purcaru and Mariuscristian Nicuescu An Approach to Worst-Case Circuit Anaysis ELENA NICULESCU*, DORINA-MIOARA PURCARU* and MARIUS- CRISTIAN NICULESCU** Eectronics and Instruentation

More information

Generation of Anti-Fractals in SP-Orbit

Generation of Anti-Fractals in SP-Orbit International Journal of Coputer Trends and Tehnology (IJCTT) Volue 43 Nuber 2 January 2017 Generation of Anti-Fratals in SP-Orbit Mandeep Kuari 1, Sudesh Kuari 2, Renu Chugh 3 1,2,3 Departent of Matheatis,

More information

MINIMIZATION OF FREQUENCY-WEIGHTED l 2 -SENSITIVITY FOR MULTI-INPUT/MULTI-OUTPUT LINEAR SYSTEMS

MINIMIZATION OF FREQUENCY-WEIGHTED l 2 -SENSITIVITY FOR MULTI-INPUT/MULTI-OUTPUT LINEAR SYSTEMS Automatique et ordinateurs MINIMIZAION OF FREQUENCY-WEIGHED -SENSIIVIY FOR MULI-INPU/MULI-OUPU LINEAR SYSEMS AKAO HINAMOO, OSAMU ANAKA, AKIMISU DOI Key words: MIMO inear disrete-time systems, Frequeny-weighted

More information

Sensitivity Analysis in Markov Networks

Sensitivity Analysis in Markov Networks Sensitivity Analysis in Markov Networks Hei Chan and Adnan Darwihe Computer Siene Department University of California, Los Angeles Los Angeles, CA 90095 {hei,darwihe}@s.ula.edu Abstrat This paper explores

More information

A Unified View on Multi-class Support Vector Classification Supplement

A Unified View on Multi-class Support Vector Classification Supplement Journal of Mahine Learning Researh??) Submitted 7/15; Published?/?? A Unified View on Multi-lass Support Vetor Classifiation Supplement Ürün Doğan Mirosoft Researh Tobias Glasmahers Institut für Neuroinformatik

More information

Genetic Algorithm Based Recurrent Fuzzy Neural Network Modeling of Chemical Processes

Genetic Algorithm Based Recurrent Fuzzy Neural Network Modeling of Chemical Processes Journa of Universa Computer Siene, vo. 3, no. 9 (27), 332-343 submitted: 2/6/6, aepted: 24//6, appeared: 28/9/7 J.UCS Geneti Agorithm Based Reurrent Fuzzy Neura Network odeing of Chemia Proesses Jii Tao,

More information

AN IMPROVED DOA ESTIMATION ALGORITHM FOR ASYNCHRONOUS MULTIPATH CDMA SYSTEM 1

AN IMPROVED DOA ESTIMATION ALGORITHM FOR ASYNCHRONOUS MULTIPATH CDMA SYSTEM 1 Vo.23 No. JOURNA OF EERONIS (INA January 26 AN IMPROVED DOA ESIMAION AGORIM FOR ASYNRONOUS MUIPA DMA SYSEM Yang Wei hen Junshi an Zhenhui (Schoo of Eectronics and Info. Eng. Beijing Jiaotong University

More information

Continuous-time low-pass filter Butterworth filter

Continuous-time low-pass filter Butterworth filter Continuous-tie ow-pass fiter Butterworth fiter th -order Butterworth fiter with utoff frequeny : ( H Very fat @ : (- derivative @.5.5 x 5 x x.5.5.5 5 5 5 5 x Magnitude is onotoniay dereasing in Passband

More information

SINCE Zadeh s compositional rule of fuzzy inference

SINCE Zadeh s compositional rule of fuzzy inference IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 14, NO. 6, DECEMBER 2006 709 Error Estimation of Perturbations Under CRI Guosheng Cheng Yuxi Fu Abstrat The analysis of stability robustness of fuzzy reasoning

More information

Combining Feature-based and Model-based Approaches For Robust Ellipse Detection

Combining Feature-based and Model-based Approaches For Robust Ellipse Detection Cobining Feature-based and Model-based Approahes For Robust Ellipse Halil Ibrahi Cakir Departent of Coputer Engineering Dulupinar University, Kutahya, Turkey Eail: akirhal@dulupinar.edu.tr Burak Benligiray,

More information

SUPPLEMENTARY INFORMATION

SUPPLEMENTARY INFORMATION DOI: 10.1038/NPHOTON.013.10 Suppementary Information: Quantum teeportation using a ight emitting diode J. Nisson 1, R. M. Stevenson 1*, K. H. A. Chan 1,, J. Skiba-Szymanska 1, M. Luamarini 1, M. B. Ward

More information

Determining the optimum length of a bridge opening with a specified reliability level of water runoff

Determining the optimum length of a bridge opening with a specified reliability level of water runoff MATE Web o onerenes 7, 0004 (07) DOI: 0.05/ ateon/0770004 XXVI R-S-P Seinar 07, Theoretial Foundation o ivil Engineering Deterining the optiu length o a bridge opening with a speiied reliability level

More information

Constructing Gyro-free Inertial Measurement Unit from Dual Accelerometers for Gesture Detection

Constructing Gyro-free Inertial Measurement Unit from Dual Accelerometers for Gesture Detection Sensors & ransduers Vo. 7 Issue 5 May 04 pp. 34-40 Sensors & ransduers 04 by IFSA Pubishing S. L. http://www.sensorsporta.om Construting Gyro-free Inertia Measurement Unit from Dua Aeerometers for Gesture

More information

MODELING OF THE NON-AZEOTROPIC MIXTURE CONDENSATION ON A VERTICAL ISOTHERMAL PLATE

MODELING OF THE NON-AZEOTROPIC MIXTURE CONDENSATION ON A VERTICAL ISOTHERMAL PLATE Proeedings of the Asian Conferene on Therma Sienes 07, st ACTS Marh 6-30, 07, Jeju Isand, Korea ACTS-P00605 MODELING OF THE NON-AZEOTROPIC MIXTURE CONDENSATION ON A VERTICAL ISOTHERMAL PLATE Li-i Zhang,

More information

Department of Electronic and Optical Engineering, Ordnance Engineering College, Shijiazhuang, , China

Department of Electronic and Optical Engineering, Ordnance Engineering College, Shijiazhuang, , China 6th International Conference on Machinery, Materials, Environent, Biotechnology and Coputer (MMEBC 06) Solving Multi-Sensor Multi-Target Assignent Proble Based on Copositive Cobat Efficiency and QPSO Algorith

More information

1. Which two values of temperature are equivalent to the nearest degree when measured on the Kelvin and on the

1. Which two values of temperature are equivalent to the nearest degree when measured on the Kelvin and on the . Whih two values of teperature are equivalent to the nearest degree when easured on the Kelvin and on the Celsius sales of teperature? Kelvin sale Celsius sale A. 40 33 B. 273 00 C. 33 40 D. 373 0 2.

More information

Nonreversibility of Multiple Unicast Networks

Nonreversibility of Multiple Unicast Networks Nonreversibility of Multiple Uniast Networks Randall Dougherty and Kenneth Zeger September 27, 2005 Abstrat We prove that for any finite direted ayli network, there exists a orresponding multiple uniast

More information

CSC2515 Winter 2015 Introduc3on to Machine Learning. Lecture 5: Clustering, mixture models, and EM

CSC2515 Winter 2015 Introduc3on to Machine Learning. Lecture 5: Clustering, mixture models, and EM CSC2515 Winter 2015 Introdu3on to Mahine Learning Leture 5: Clustering, mixture models, and EM All leture slides will be available as.pdf on the ourse website: http://www.s.toronto.edu/~urtasun/ourses/csc2515/

More information

The Influences of Smooth Approximation Functions for SPTSVM

The Influences of Smooth Approximation Functions for SPTSVM The Influenes of Smooth Approximation Funtions for SPTSVM Xinxin Zhang Liaoheng University Shool of Mathematis Sienes Liaoheng, 5059 P.R. China ldzhangxin008@6.om Liya Fan Liaoheng University Shool of

More information

Kinematics of Elastic Neutron Scattering

Kinematics of Elastic Neutron Scattering .05 Reator Physis - Part Fourteen Kineatis of Elasti Neutron Sattering. Multi-Group Theory: The next ethod that we will study for reator analysis and design is ulti-group theory. This approah entails dividing

More information

Time-varying Stiffness Characteristics of Shaft with Slant Crack

Time-varying Stiffness Characteristics of Shaft with Slant Crack Internationa Conferene on Modeing, Simuation and Appied Mathematis (MSAM 05) Time-varying Stiffness Charateristis of Shaft with Sant Cra Hengheng Xia Shoo of Aeronautia Manufaturing Engineering Nanhang

More information

A Decision Theoretic Framework for Analyzing Binary Hash-based Content Identification Systems

A Decision Theoretic Framework for Analyzing Binary Hash-based Content Identification Systems A Deision Theoreti Fraework for Analyzing Binary Hash-based Content Identifiation Systes Avinash L Varna Departent of Eletrial and Coputer Engineering University of Maryland College Park, MD, USA varna@udedu

More information

Error Bounds for Context Reduction and Feature Omission

Error Bounds for Context Reduction and Feature Omission Error Bounds for Context Redution and Feature Omission Eugen Bek, Ralf Shlüter, Hermann Ney,2 Human Language Tehnology and Pattern Reognition, Computer Siene Department RWTH Aahen University, Ahornstr.

More information

The Effectiveness of the Linear Hull Effect

The Effectiveness of the Linear Hull Effect The Effetiveness of the Linear Hull Effet S. Murphy Tehnial Report RHUL MA 009 9 6 Otober 009 Department of Mathematis Royal Holloway, University of London Egham, Surrey TW0 0EX, England http://www.rhul.a.uk/mathematis/tehreports

More information

TRACKING CONTROL FOR WHEELED MOBILE ROBOTS USING NEURAL NETWORK MODEL ALGORITHM CONTROL

TRACKING CONTROL FOR WHEELED MOBILE ROBOTS USING NEURAL NETWORK MODEL ALGORITHM CONTROL Journa of Theoretica and Appied Inforation Technoogy 3 st Deceber. Vo. 46 No. 5 - JATIT & LLS. A rights reserved. ISSN: 99-8645 www.jatit.org E-ISSN: 87-395 TRACKING CONTROL FOR WHEELED OBILE ROBOTS USING

More information

The Ultimate Strategy to Search on m Rays??

The Ultimate Strategy to Search on m Rays?? The Ultiate Strategy to Searh on Rays?? Alejandro López-Ortiz 1 and Sven Shuierer 2 1 Faulty of Coputer Siene, University of New Brunswik, Canada, eail alopez-o@unb.a 2 Institut für Inforatik, Universität

More information

International Journal of Thermodynamics, Vol. 18, No. 1, P (2015). Sergey G.

International Journal of Thermodynamics, Vol. 18, No. 1, P (2015).   Sergey G. International Journal of Therodynais Vol. 8 No. P. 3-4 (5). http://dx.doi.org/.554/ijot.5343 Four-diensional equation of otion for visous opressible and harged fluid with regard to the aeleration field

More information

ABSOLUTELY CONTINUOUS FUNCTIONS OF TWO VARIABLES IN THE SENSE OF CARATHÉODORY

ABSOLUTELY CONTINUOUS FUNCTIONS OF TWO VARIABLES IN THE SENSE OF CARATHÉODORY Eetroni Journa of Differentia Equations, Vo. 2010(2010), No. 154, pp. 1 11. ISSN: 1072-6691. URL: http://ejde.math.txstate.edu or http://ejde.math.unt.edu ftp ejde.math.txstate.edu ABSOLUTELY CONTINUOUS

More information

Supplementary Materials

Supplementary Materials Supplementary Materials Neural population partitioning and a onurrent brain-mahine interfae for sequential motor funtion Maryam M. Shanehi, Rollin C. Hu, Marissa Powers, Gregory W. Wornell, Emery N. Brown

More information

THREE-DIMENSIONAL NON-LINEAR EARTHQUAKE RESPONSE ANALYSIS OF REINFORCED CONCRETE STRUCTURES

THREE-DIMENSIONAL NON-LINEAR EARTHQUAKE RESPONSE ANALYSIS OF REINFORCED CONCRETE STRUCTURES HREE-DIMESIOAL O-LIEAR EARHQUAKE RESPOSE AALYSIS OF REIFORCED COCREE SRUCURES K. ishiura 1), K. akiguhi 2), and H. H. guen 3) 1) Assistant Professor, Dept. of Arhiteture and Building Engineering, oko Institute

More information

COS 424: Interacting with Data. Written Exercises

COS 424: Interacting with Data. Written Exercises COS 424: Interacting with Data Hoework #4 Spring 2007 Regression Due: Wednesday, April 18 Written Exercises See the course website for iportant inforation about collaboration and late policies, as well

More information

A NETWORK SIMPLEX ALGORITHM FOR THE MINIMUM COST-BENEFIT NETWORK FLOW PROBLEM

A NETWORK SIMPLEX ALGORITHM FOR THE MINIMUM COST-BENEFIT NETWORK FLOW PROBLEM NETWORK SIMPLEX LGORITHM FOR THE MINIMUM COST-BENEFIT NETWORK FLOW PROBLEM Cen Çalışan, Utah Valley University, 800 W. University Parway, Orem, UT 84058, 801-863-6487, en.alisan@uvu.edu BSTRCT The minimum

More information

Chapter 8 Hypothesis Testing

Chapter 8 Hypothesis Testing Leture 5 for BST 63: Statistial Theory II Kui Zhang, Spring Chapter 8 Hypothesis Testing Setion 8 Introdution Definition 8 A hypothesis is a statement about a population parameter Definition 8 The two

More information

Simple Harmonic Motion

Simple Harmonic Motion Chapter 3 Sipe Haronic Motion Practice Probe Soutions Student extboo pae 608. Conceptuaize the Probe - he period of a ass that is osciatin on the end of a sprin is reated to its ass and the force constant

More information

Session : Electrodynamic Tethers

Session : Electrodynamic Tethers Session : Eectrodynaic Tethers Eectrodynaic tethers are ong, thin conductive wires depoyed in space that can be used to generate power by reoving kinetic energy fro their orbita otion, or to produce thrust

More information

Available online at ScienceDirect

Available online at   ScienceDirect Avaiabe onine at www.sienediret.om SieneDiret Proedia Engineering 7 04 ) 9 4 Geoogia Engineering Driing Tehnoog Conferene IGEDTC), New Internationa Convention Eposition Center Chengdu Centur Cit on rd-5th

More information

Abstract code: Meta Heuristics to Minimize Line Stoppage Time in Mixed-Model Sequencing Problem

Abstract code: Meta Heuristics to Minimize Line Stoppage Time in Mixed-Model Sequencing Problem Abstrat ode: 015-0215 Meta Heuristis to Miniize Line Stoppage Tie in Mixed-Model Sequening Proble Takayoshi Taura *1, Tej S. Dhakar *2, Katsuhisa Ohno *3, Taiji Okuura *1 *1 Nagoya Institute of Tehnology,

More information

Planning with Uncertainty in Position: an Optimal Planner

Planning with Uncertainty in Position: an Optimal Planner Planning with Unertainty in Position: an Optimal Planner Juan Pablo Gonzalez Anthony (Tony) Stentz CMU-RI -TR-04-63 The Robotis Institute Carnegie Mellon University Pittsburgh, Pennsylvania 15213 Otober

More information

Model-based mixture discriminant analysis an experimental study

Model-based mixture discriminant analysis an experimental study Model-based mixture disriminant analysis an experimental study Zohar Halbe and Mayer Aladjem Department of Eletrial and Computer Engineering, Ben-Gurion University of the Negev P.O.Box 653, Beer-Sheva,

More information

Keywords: Controllers with PI and PID dynamics, Takagi-Sugeno fuzzy models, Takagi- Sugeno fuzzy controllers, stability analysis, winding system.

Keywords: Controllers with PI and PID dynamics, Takagi-Sugeno fuzzy models, Takagi- Sugeno fuzzy controllers, stability analysis, winding system. Ata Poytehnia Hungaria Vo. 2, No., 2005 Deveopment of Conventiona and Fuzzy Controers and Takagi-Sugeno Fuzzy Modes Dediated for Contro of Low Order Benhmarks with Time Variabe Parameters Stefan Preit

More information

Multi-events Earthquake Early Warning algorithm

Multi-events Earthquake Early Warning algorithm subitted to Geophys. J. Int. Muti-events Earthquake Eary Warning agorith using a Bayesian approach S. Wu 1, M. Yaada 2, K. Taaribuchi 3 and J.L. Beck 1 1 Caifornia Institute of Technoogy, Pasadena, Caifornia,

More information

Parametric and sensitivity analysis of a vibratory automobile model

Parametric and sensitivity analysis of a vibratory automobile model Louisiana State University LSU Digita Commons LSU Master's Theses Graduate Shoo Parametri and sensitivity anaysis of a vibratory automobie mode Kania Nioe Vesse Louisiana State University and Agriutura

More information

Support Vector Machines. Maximizing the Margin

Support Vector Machines. Maximizing the Margin Support Vector Machines Support vector achines (SVMs) learn a hypothesis: h(x) = b + Σ i= y i α i k(x, x i ) (x, y ),..., (x, y ) are the training exs., y i {, } b is the bias weight. α,..., α are the

More information

Probabilistic Graphical Models

Probabilistic Graphical Models Probabilisti Graphial Models David Sontag New York University Leture 12, April 19, 2012 Aknowledgement: Partially based on slides by Eri Xing at CMU and Andrew MCallum at UMass Amherst David Sontag (NYU)

More information

APPLICATION OF VIM, HPM AND CM TO THE SYSTEM OF STRONGLY NONLINEAR FIN PROBLEM. Islamic Azad University, Sari, Iran

APPLICATION OF VIM, HPM AND CM TO THE SYSTEM OF STRONGLY NONLINEAR FIN PROBLEM. Islamic Azad University, Sari, Iran Journal of Engineering and Tehnology APPLICATION OF VIM, HPM AND CM TO THE SYSTEM OF STRONGLY NONLINEAR FIN PROBLEM M. R. Shirkhani,H.A. Hoshyar *, D.D. Ganji Departent of Mehanial Engineering, Sari Branh,

More information

Physical Laws, Absolutes, Relative Absolutes and Relativistic Time Phenomena

Physical Laws, Absolutes, Relative Absolutes and Relativistic Time Phenomena Page 1 of 10 Physial Laws, Absolutes, Relative Absolutes and Relativisti Time Phenomena Antonio Ruggeri modexp@iafria.om Sine in the field of knowledge we deal with absolutes, there are absolute laws that

More information

Sensitivity analysis for linear optimization problem with fuzzy data in the objective function

Sensitivity analysis for linear optimization problem with fuzzy data in the objective function Sensitivity analysis for linear optimization problem with fuzzy data in the objetive funtion Stephan Dempe, Tatiana Starostina May 5, 2004 Abstrat Linear programming problems with fuzzy oeffiients in the

More information