Cross-Domain Metric Learning Based on Information Theory

Size: px
Start display at page:

Download "Cross-Domain Metric Learning Based on Information Theory"

Transcription

1 Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence Cross-Doain Metric Learning Based on Inforation Theory Hao Wang,2, Wei Wang 2,3, Chen Zhang 2, Fanjiang Xu 2. State Key Laboratory of Coputer Science 2. Science and Technology on Integrated Inforation Syste Laboratory Institute of Software, Chinese Acadey of Sciences, Beijing 0090, China 3. Departent of Autoation, University of Science and Technology of China Abstract Supervised etric learning plays a substantial role in statistical classification. Conventional etric learning algoriths have liited utility when the training data and testing data are drawn fro related but different doains (i.e., source doain and target doain). Although this issue has got soe progress in feature-based transfer learning, ost of the work in this area suffers fro non-trivial optiization and pays little attention to preserving the discriinating inforation. In this paper, we propose a novel etric learning algorith to transfer knowledge fro the source doain to the target doain in an inforation-theoretic setting, where a shared Mahalanobis distance across two doains is learnt by cobining three goals together: ) reducing the distribution difference between different doains; 2) preserving the geoetry of target doain data; 3) aligning the geoetry of source doain data with its label inforation. Based on this cobination, the learnt Mahalanobis distance effectively transfers the discriinating power and propagates standard classifiers across these two doains. More iportantly, our proposed ethod has closed-for solution and can be efficiently optiized. Experients in two real-world applications deonstrate the effectiveness of our proposed ethod. Introduction Distance etric learning is of fundaental iportance in achine learning. Previous research has deonstrated that appropriate distance etrics learnt fro labeled training data can greatly iprove classification accuracy (Jin, Wang, and Zhou 2009). Depending on whether the geoetry inforation is used, state-of-the-art supervised etric learning ethods can be classified into two categories, i.e., globality and locality. Globality etric learning ethods ai at keeping all the data points in the sae class close together for copactness while ensuring those fro different classes far apart for separability (Davis et al. 2007; Globerson and Roweis 2006; Wang and Jin 2009; Xing et al. 2002). Locality etric learning ethods incorporate the Corresponding author who ade ain idea and contribution to this work. Copyright c 204, Association for the Advanceent of Artificial Intelligence ( All rights reserved. geoetry of data with the label inforation to accoodate ultiodal data distributions and to further iprove classification perforance (Weinberger and Saul 2009; Yang et al. 2006). Existing etric learning ethods always perfor well when there are sufficient labeled training saples. However, in soe real-world applications, obtaining the label inforation of data points drawn fro the task-specific doain (i.e., target doain) is extreely expensive or even ipossible. One ay turn to find labeled data drawn fro a related but different doain (i.e., source doain) and apply it as prior knowledge. Apparently, distance etrics learnt only in source doain cannot be directly reused in target doain, although these two doains are closely related. It is because that the significant distribution difference between the data drawn fro source and target doains is not explicitly taken into considerations, and this difference will ake classifiers trained in source doain invalid in target doain. Therefore, it is iportant and necessary to reduce the distribution difference between labeled source doain data and unlabeled target doain data in distance etric learning. Recently, soe feature extraction approaches in transfer learning (Caruana 997; Pan and Yang 200) have been proposed to address this proble by iplicitly exploring a etric (siilarity) as a bridge for inforation transfer fro the source doain to the target doain (Geng, Tao, and Xu 20; Long et al. 203; Pan, Kowok, and Yang 2008; Pan et al. 20; Si, Tao, and Geng 200). These feature extraction ethods learn a shared feature representation across doains by ) reducing the distribution difference, 2) preserving the iportant properties (e.g., variance or geoetry) of data, especially the target doain data. However, ost work in this area does not focus on incorporating the geoetry with the label inforation of source doain data to iprove the classification perforance in target doain. Moreover, these ethods forulate a seidefinite prograing (SDP) (Boyd and Vandenberghe 2004) or a non-convex optiization proble, resulting in expensive coputation. In this paper, we address the transfer learning proble fro the etric learning view and propose a novel algorith naed Cross-Doain Metric Learning (CDML). Specifically, CDML first iniizes the distance between different distributions such that the arginal distributions of target doain and source doain data are close under the learnt distance etric. Second, two Gaussian distributions are con- 2099

2 structed, one based on the Mahalanobis distance to be learnt and the other based on the geoetry of target doain data. By iniizing the relative entropy between these two distributions, the geoetry of target doain data is preserved in the learnt distance etric. Third, another two Gaussian distributions are constructed, one based on the Mahalanobis distance to be learnt as well and the other based on the labels and the geoetry of source doain data. By iniizing the relative entropy between these two distributions, the learnt distance etric pulls the source doain data in the sae class close together, while pushing differently labeled data far apart. Finally, the three ters above are cobined into the unified loss function of CDML. This cobination effectively transfers the discriinating power gained fro the labeled source doain data to the unlabeled target doain data. To the best of our knowledge, our ethod has ade the first attept to cross-doain etric learning based on relative entropy. We ephasize that CDML has the closedfor solution, leading to efficient optiization. In suary, the contribution of this paper is two-fold. Fro the perspective of etric learning, we ai at addressing the challenge of distribution difference. Fro the perspective of transfer learning, a novel algorith is proposed to transfer knowledge by finding a shared Mahalanobis distance across doains. The optial etric can be found efficiently in closed-for. Under this optial etric, the data distributions are close and points fro different classes can be well separated. As a result, we can train standard classifiers in the source doain and reuse the to correctly classify the target doain data. Experiental results in realworld applications verify the effectiveness and efficiency of CDML copared with state-of-the-art etric learning ethods and transfer learning ethods. Metric Learning Related Work Significant efforts in etric learning have been spent on learning a Mahalanobis distance fro labeled training data for classification. Existing Mahalanobis distance learning ethods can be classified into two categories, i.e., globality and locality. A natural intention in globality learning is to forulates an SDP for keeping the sae labeled points siilar (i.e., the distances between the should be sall) and differently labeled points dissiilar (i.e., the distances should be larger) (Globerson and Roweis 2006; Xing et al. 2002). Other notable work in globality learning is based on inforation theory (Davis et al. 2007; Wang and Jin 2009). In particular, Inforation-Theoretic Metric Learning (ITML) (Davis et al. 2007) forulates the relative entropy as a Bregan optiization proble subject to linear constraints. Inforation Geoetry Metric Learning (IGML) (Wang and Jin 2009) iniizes the Kullback-Leibler (K-L) divergence between two Gaussian distributions and finds the closed-for solution. Locality etric learning ethods axially align the geoetry of data with its label inforation (Weinberger and Saul 2009; Yang et al. 2006) to further iprove their perforance. However, the supervised algoriths discussed above are liited by the underlying assuption that training data and testing data are drawn fro the sae distribution. Transfer Learning State-of-the-art transfer learning can be organized into instance reweighing (Dai et al. 2007a) and feature extraction. In the feature extraction category, recent work tries to find a subspace shared by both doains, such that the distribution difference is explicitly reduced and the iportant properties of original data are preserved (Geng, Tao, and Xu 20; Long et al. 203; Pan, Kowok, and Yang 2008; Si, Tao, and Geng 200). In this subspace, classifiers can be propagated between doains. Specifically, Maxiu Mean Discrepancy Ebedding (MMDE) (Pan, Kowok, and Yang 2008) eploys Maxiu Mean Discrepancy (MMD) (Gretton et al. 2006) to estiate the distance between different distributions and learns a kernel atrix by preserving the data variance at the sae tie. Joint Distribution Adaption (JDA) (Long et al. 203) extends MMD and constructs feature subspace by Principal Coponent Analysis (PCA) (Jolliffe 986). Transfer Subspace Learning (TSL) (Si, Tao, and Geng 200) integrates the Bregan divergence with soe diension reduction algoriths, e.g., PCA and Fisher s linear discriinant analysis (FLDA) (Fisher 936). However, these ethods forulate an SDP or a non-convex optiization, which has high coputational coplexity and requires iteratively updating paraeters. Even worse, the non-convex probles are prone to being trapped in local solutions. In coparison, our etric learning ethod has efficient closedfor solution and optially transfers the discriinating power. We would also like to ention that Transfer Coponent Analysis (TCA) (Pan et al. 20) is an efficient kernel learning ethod to extend MMDE. Our work differs fro TCA significantly in the proposed optiization. In this paper, an optial Mahalanobis distance is searched by utilizing the relationship between Gaussian distributions. Cross-Doain Metric Learning Based on Inforation Theory In this section, we present the proposed algorith naed Cross-Doain Metric Learning (CDML) in detail. Proble Definition We begin with the proble definition. Table lists the iportant notations used in this paper. Definition. (The Mahalanobis Distance) Denote x i, x j R d, and then the Mahalanobis distance between x i and x j is calculated as follows: d A (x i, x j ) = (x i x j ) T A(x i x j ), () where A R d d is positively sei-definite. In fact, there is a close link between Mahalanobis distance and linear transforation. If we define a linear projection W: W T W = A which aps x i to Wx i, the Euclidean distance between Wx and Wx 2, i.e., Wx Wx 2 2 = (x x 2 ) T W T W(x x 2 ) = (x x 2 ) T A(x x 2 ), is actually the Mahalanobis distance between x and x

3 Table : List of iportant notations used in this paper. Notation X src = {(x s, y), s..., (x s n, yn)} s X tar = {x t,..., x t } X = {x s,..., x s n, x t,..., x t } W A = W T W L K tar = [ Wx t i, Wxt j ] K T K src = [ Wx s i, Wxs j ] n n K S Description Source doain data set Target doain data set Input data set Linear transforation atrix Mahalanobis distance atrix The MMD atrix The linear kernel atrix for WX tar The ideal kernel atrix for X tar The linear kernel atrix for WX src The ideal kernel atrix for X src Proble. (Cross-Doain Metric Learning Based on Inforation Theory) Let X tar be a set of unlabeled testing saples drawn fro a target doain: X tar = {x t,..., x t }, where x t i R d. Let X src be a set of n labeled training saples drawn fro a related source doain: X src = {(x s, y), s..., (x s n, yn)}, s where x s i R d and yi s Y s is the class label. We denote P t (X tar ) and P s (X src ) as the arginal probability distributions of X tar and X src respectively, P t (X tar ) P s (X src ). Our task is to learn a shared etric distance A across doains under which ) the distribution difference between P s (X src ) and P t (X tar ) is explicitly reduced; 2) the geoetry of X tar is preserved; 3) the points fro X src with the sae label are kept siilar according to the geoetry and others are kept dissiilar. Miniizing Distribution Difference Conventional Mahalanobis distance learning ethods perfors well in the classification setting based on the assuption that training and testing points are drawn fro the sae distribution (i.e., P s (X src ) = P t (X tar )). When such a distance etric W c is learnt fro X src, it can iprove classification accuracy on X tar using standard classifiers such as KNN and SVM. However, P s (X src ) is usually different fro P t (X tar ) since X src and X tar are drawn fro different but related doains. In this case, P s (W c X src ) and P t (W c X tar ) are still significantly different and standard classification odels trained on W c X src cannot be directly applied on W c X tar. Therefore, it is necessary to find a etric W which can reduce the distance between different distributions. This issue is of particular iportance and gains its popularity in transfer learning. Inspired by the work (Long et al. 203; Pan, Kowok, and Yang 2008), we adopt the criterion Maxiu Mean Discrepancy (MMD) to easure the distance between P s (WX src ) and P t (WX tar ). The epirical estiate of MMD is as follows: n Wx s i Wx t i 2 = tr(xlx T A), (2) n i= i= where X = {x s,..., x s n, x t,..., x t } R d (n+), L R (n+) (n+) with: if x n i, x 2 j X src L(i, j) = if x i, x 2 j X tar (3) otherwise. n By iniizing Equation (2), P s (WX src ) and P t (WX tar ) are close to each other. Transferring Discriinating Power Based on Inforation Theory The etric distance W learnt by only iniizing the distribution difference ay erge all data points together, which is unsuitable for the classification task. To iprove classification accuracy, as stated in Proble, W should cobine iniizing the distribution difference with ) preserving the geoetry of X tar, 2) axially aligning the geoetry of X src with its label inforation. Based on this cobination, it is supposed that P s (Y s WX src ) P t (Y t WX tar ). W optially transfers discriinating power gained fro the source doain to the target doain, that is, the sae labeled points are kept close together and the differently labeled points are pushed far apart. In this way, if a classifier is trained on WX src and Y s, it can be reused to correctly classify WX tar. Note that the cobination can perfor well because X tar and X src share soe latent variables. Geoetry Preservation of X tar Preserving the geoetry of unlabeled X tar is particular useful for transfer learning (Long et al. 202; Wang and Mahadevan 20; Pan et al. 20). We construct a linear kernel K tar for WX tar : K tar = (WX tar) T (WX tar) = X T tarax tar. (4) To introduce the inforation theory into the space of positive definite atrices, K tar is related as the covariance atrix of a ultivariate Gaussian distribution with zero ean (Wang and Jin 2009): P r(z K tar) = (2π) /2 K tar /2 exp( zt K tarz/2), (5) where z R. In the ideal case, an ideal kernel atrix K T is expected to give a useful siilarity such that the geoetry of X tar is preserved. K T is related as the covariance atrix of another ultivariate Gaussian distribution: P r(z K T ) = (2π) /2 K T /2 exp( zt K T z/2), (6) where z R. The distance between K tar and K T, denoted as d(k tar K T ), can be derived by the K-L divergence between the two distributions in Equation (5) and (6): d(k tar K T ) = KL(P r(z K tar) P r(z K T )) = P r(z K P r(z Ktar)) tar)log P r(z K T ) dz. (7) Theore. The distance between K tar and K T in Equation (7) is equivalent to: d(k tar K T ) = 2 (tr(k T K tar) log K tar + log K T ). (8) To capture the inforation of K T, the optial A is searched by iniizing the distance d(k tar K T ) in Equation (8). Therefore, the geoetry of unlabeled X tar can be preserved in the learnt distance A: A = arg in A 0 d(ktar KT ) = arg in A 0 tr(k T X T tarax tar) log X T tarax tar. (9) 20

4 The reaining issue is to define the ideal kernel K T for geoetry preservation.. Constructing a k-nearest neighbor graph: let G t denote a directed graph containing a set of nodes V t nubered to and a set of edges E t. Two nodes i and j are connected by an edge (i.e., (i, j) E t ) if x t i is one of the k nearest neighbor of x t j. 2. Choosing weights: let M t refer to the adjacency atrix of G t, and it is given by: { M t exp( d ij ) if (i, j) E t (i, j) = 2σ 2 (0) 0 otherwise, where d ij = x t i xt j 2 and σ is the width. 3. Defining a kernel function K T on G t : specific kernel functions (Kondor and Lafferty 2002; Sola and Kondor 2003) on G t induced by the weights can give a useful and ore global sense of siilarity between instances. Let D t be an diagonal atrix with D t ii = j Mt ij. The Laplacian of G t is L t = D t M t, and the Noralized Laplacian is L t = (D t ) 2 L(D t ) 2. The eigenvalues and eigenvectors of L t are denoted as λ t i and φt i, i.e., L t = i λt i (φt i )(φt i )T. In this paper, we investigate the diffusion kernel (Kondor and Lafferty 2002) which is proven to be a generalization of Gaussian kernel to graphs: K T = exp( σd/2λ 2 t i)(φ t i)(φ t i) T, () i= where K T 0 since all the eigenvalues are positive (i.e., exp( σd 2/2λt i ) > 0). Label Inforation Utilization of X src A linear kernel K src is constructed for WX src : K src = (WX src ) T (WX src ) = X T srcax src. Label inforation is critical for classification tasks and encourages the siilarities between two points if and only if they belong to the sae class. Geoetry preservation is an iportant coponent for generalization ability (Weinberger and Saul 2009; Yang et al. 2006). By incorporating these two sources of inforation, an ideal kernel K S is defined for X src based on two idealizations: ) siilarities between points with different labels will be penalized; 2) siilarities between points in the sae class will be encouraged according to the neighborhood structure.. Constructing a within class graph: let G s denote a directed graph which consists of a set of nodes V s nubered to n and a set of edges E s. Two nodes i and j are connected by an edge (i.e., (i, j) E s ) if yi s = ys j. 2. Choosing the adjacency atrix M s of G s : M s (i, j) = exp( dij 2σ ) if (i, j) E s, otherwise M s (i, j) = Defining 2 a diffusion kernel function K S on G s : K S = n i= exp( σ2 d /2λs i )(φs i )(φs i )T, where (λ s i, φs i ) are eigenvalues and eigenvectors of the Noralized Laplacian. 4. Miniizing d(k src K S ): the optial A is searched by iniizing the distance d(k src K S ) derived fro Equation (8). Therefore, the learnt distance A axially aligns the geoetry of X src with its label inforation: A = arg in A 0 tr(k S X T srcax src) log X T srcax src. (2) The Cost Function CDML ais at searching the optial distance etric A by iniizing Equation (2), Equation (9) and Equation (2) siultaneously. This cobination effectively transfers the discriinating power gained fro the labeled source doain data to the unlabeled target doain data. The overall cost function is as follows: A = arg in tr(x(k + A 0 µl)xt A) log X T tarax tar (3) log X T srcax src, where µ > 0 is a tradeoff and K 0 = ( K S 0 0 K T Proposition. The (n + ) (n + ) atrix L in Equation(2) and Equation (3) is positive sei-definite. Proof. For any colun vector z R n+, we have z T Lz = ( a b ) ( ) ( ) P R a T R T Q b T ). (4) where a = (z,..., z n ), b = (z n+,..., z n+ ), P R n n with [P] ij = /n 2, Q R with [Q] ij = / 2 and R R n with [R] ij = /n. z T Lz in Equation (4) is equal to: apa T + bqb T + 2aRb T = n n i= j= z i z j n n + i= j= z n+i z n+j 2 n i= j= = ( z n z n n z n+... z n+ )2 0 Therefore, L 0. The proposition follows. z i z n+j n Based on Proposition, we can obtain the closed-for solution of CDML in the following proposition. Proposition 2. The optial solution to Equation (3) is: A = 2(X(K + µl)x T ) (5) Proof. The derivative of Equation (3) w.r.t. A is: X(K + µl)x T 2A. (6) Since K 0 and L 0, then (K + µl) 0. Proposition 2 now follows by setting the derivative to 0. Low Diensional Projections The Mahalanobis distance etric A learnt in CDML is of full rank. If A has the rank r < d, we can represent it in the for: A = W T r W r, where W r R r d projects the original data to an r-diensional space for diension reduction. To copute W r, a straightforward solution is to optiize Equation (3) with a constraint rank(a) = r. However, rank constraints on atrices are not convex (Boyd and Vandenberghe 2004). In this paper, the projection atrix W r is coputed by a substitute approach (Globerson and Roweis 2006) as follows: ) eigenvalues and eigenvectors of full-rank A in Equation (5) are calculated: A = d i= λ iu i u T i, where λ λ 2... λ d ; 2) W r = diag( λ,..., λ r )[u T ;... ; u T r ]. The eigen spectru of A usually rapidly decays and any eigenvalues are very sall, suggesting this solution is close to the optial one returned by iniizing the rank constrained optiization. 202

5 Experients In this section, we evaluate the proposed ethod in two etric learning related applications: ) face recognition and 2) text classification. Data Preparation Face Data Sets FERET (Phillips et al. 2000) and YALE (Belhueur, Hespanha, and Kriegan 997) are two public face data sets. FERET data set contains 3,539 face iages fro,565 individuals with different sizes, poses, illuinations and facial expressions. YALE data set has 65 iages fro 5 individuals with different expressions or configurations. Soe exaple face iages are shown in Figure. As in the previous work (Si, Tao, and Geng 200), we construct two cross-doain data sets: ) Y vs F: the source doain set is YALE, and the target doain set consists of 00 individuals randoly selected fro FERET. 2) F vs Y: the source set contains 00 individuals randoly selected fro FERET, and the target set is YALE. Figure : Iage exaples in (a) FERET data set and (b) YALE data set. Text Data Sets 20-Newsgroups and Reuters-2578 are two benchark text data sets widely used for evaluating the transfer learning algoriths (Dai et al. 2007b; Li, Jin, and Long 202; Pan et al. 20). 20-Newsgroups consists of nearly 20,000 docuents partitioned into 20 different subcategories. The corpus has four top categories and each top category has four subcategories as shown in Table 2. Following the work (Dai et al. 2007b), we construct six crossdoain data sets for binary text classification: cop vs rec, cop vs sci, cop vs talk, rec vs sci, rec vs talk and sci vs talk. Specifically, for each data set (e.g., cop vs rec), one top category (i.e., cop) is selected as the positive class and the other category (i.e., rec) is the negative class. Then two subcategories under the positive and the negative classes respectively are selected to for the source doain, the other two subcategories are used to for the target doain. Table 2: Top categories and their subcategories. Top Category Subcategory Exaples cop cop.graphics, cop.sys.ac.hardware, cop.os.s-windows.isc, cop.sys.ib.pc.hardware 3870 rec rec.autos, rec.otorcycles, rec.sport.baseball, rec.sport.hokey 3968 sci sci.crypt, sci.electronics, sci.ed, sci.space 3945 talk talk.politics.guns, talk.politics.ideast, talk.politics.isc, talk.religion.isc 3250 Reuters-2578 has three biggest top categories: orgs, people and places. The preprocessed version of Reuters-2578 on the web site ( is used which contains three cross-doain data sets: orgs vs people, orgs vs place and people vs place. Baseline Methods We systeatically copare CDML with three state-ofthe-art etric learning ethods, i.e., Inforation-Theoretic Metric Learning (ITML) (Davis et al. 2007); Inforation Geoetry Metric Learning (IGML) (Wang and Jin 2009); Large Margin Nearest Neighbor (LMNN) (Weinberger and Saul 2009); and three feature-based transfer learning ethods, i.e., Joint Distribution Adaption (JDA) (Long et al. 203); Seisupervised Transfer Coponent Analysis (SSTCA) (Pan et al. 20); Transferred Fisher s Linear Discriinant Analysis (TFLDA) (Si, Tao, and Geng 200); For the six coparison ethods, the paraeters spaces are epirically searched using their own optial paraeter settings and the best results are reported. CDML involves four paraeters: σ d, σ, µ and k. Specifically, we set σ d by searching the values aong {0.,, 0}, σ aong {0.,, 0} and µ aong {0.0, 0.,, 0}. The neighborhood size k for CDML is 3. In general, CDML is found to be robust to these paraeters. The experients are carried out on a single achine with Intel Core Ghz and 0 GB of RAM running 64-bit Windows 7. Experiental Results Results of Face Recognition In this section, we evaluate the ability of CDML to separate different classes in target doain. For Y vs F and F vs Y, one rando point for each target doain class is selected as the reference data set (Si, Tao, and Geng 200). The diensionality of each iage is reduced to 00 by PCA. All the ethods are trained as a etric learning procedure without the labels of target doain data. At the testing stage, the distance between a target point and every reference point is calculated using the learnt distance etric, then the label of the testing point is predicted as that of the nearest reference point. Since FERET and YALE has different class nubers, JDA is not suitable for this task which requires that source and target doain should share the sae class nuber. TFLDA can find at ost c eaningful diensions, where c is the class nuber of source doain. Figure 2 shows the classification error rates across different diensions. Soe observations can be concluded. The first general trend is that conventional etric learning algoriths (i.e., ITML, IGML and LMNN) show their liits on these cross-doain data sets. The etrics learnt only fro the source doain data fail to separate different classes in target doain. The second general trend is that SSTCA shows good classification perforance. SSTCA tries to learn a kernel atrix across doains such that the label dependence is axiized and the anifold structure is preserved. However, CDML consistently provides uch higher accuracy than SSTCA. A possible reason is that CDML focuses on keeping the data points in the sae class close together while ensuring those fro different classes far apart. The third general trend is that although TFLDA works 203

6 (a) (b) (c) Figure 2: Coparison of ITML, IGML, LMNN, SSTCA, TFLDA and CDML on the face data sets. (a) Classification error rates on Y-F data set. (b) Classification error rates on F-Y data set. (c) Running tie coparison. Method ITML IGML LMNN JDA SSTCA CDML Table 3: -NN classification errors (in percent) of the applied ethods. # Di orgs vs people orgs vs place people vs place Data Set cop vs rec cop vs sci cop vs talk rec vs sci rec vs talk sci vs talk quite well, it can just find at ost c eaningful diensions. By contrast, CDML alost achieves the optial error rate across all the diensions which illustrates its effective perforance in separating different target classes. To test the efficiency of CDML, we report the average training tie in Figure 2(c). ITML, LMNN and TFLDA are coputationally expensive since they forulate an alternative optiization proble. Even worse, TFLDA is nonconvex and ay be trapped in local solutions. Although IGML is fast due to the closed-for solution, it shows high classification error on these cross-doain data sets. We find CDML and SSTCA run quite efficiently, while CDML outperfors SSTCA in ters of classification accuracy. Results of Text Classification In this section, we evaluate the ability of CDML for text classification and a siple easureent is used: isclassification rate by -nearest neighbor classifier (-NN) without paraeters tuning. The unlabeled target instances are copared to the points in the labeled source doain using the learnt distance etric. We copare our proposed CDML with ITML, IGML, LMNN, JDA, SSTCA for this binary task. The classification results across different diensions are shown in Table 3. Soe advantages can be concluded fro the results. First, the results of non-transfer etric learning ethods are better than that of the transfer algoriths on cop vs rec and rec vs talk. A possible explanation is that on these two data sets, the distributions of source and target data are not significantly varied. But we would like to ention that the transfer ethods always perfor well on other cross-doain data sets. Second, JDA provides better results on people vs place and sci vs talk. The possible explanation is two-fold. ) Besides reducing the arginal distribution difference, the conditional distribution difference is also exploited in JDA. 2) The coon assuption in transferring learning that reducing the difference of arginal distributions will draw close the conditional distributions is not always valid. Third, CDML achieves the inial error rate on ost of the data sets, which illustrates the reliable and effective perforance of CDML for doain adaption. Conclusion In this paper, we have proposed a novel etric learning algorith to address transfer learning proble based on inforation theory. It learns a shared Mahalanobis distance across doains to transfer the discriinating power gained fro the source doain to the target doain. Based on the learnt distance, a standard classification odel trained only in the source doain can correctly classify the target doain data. Experients deonstrate the effectiveness of our proposed ethod. In future work, it is iportant and proising to explore an online algorith for cross-doain etric learning and the nonlinear version needs to be investigated. 204

7 Acknowledgents This work is supported by Natural Science Foundation of China (630364) and Beijing Natural Science Foundation (944037). References Belhueur, P. N.; Hespanha, J. P.; and Kriegan, D. J Eigenfaces versus fisherfaces: recognition using class specific linear projection. IEEE Transactions on Pattern Analysis and Machine Intelligence 9(7): Boyd, S., and Vandenberghe, L Convex Optiization. Cabridge University Press, Cabridge. Caruana, R Multitask learning, Machine Learning 28():4-75. Dai, W.; Yang, Q.; Xue, G.; and Yu, Y Boosting for transfer learning. In Proceedings of the 24th International Conference on Machine Learning (ICML), Dai, W.; Xue, G.-R.; Yang, Q.; and Yu, Y Coclustering based classification for out-of-doain docuents. In Proceedings of the 3th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD). Davis, J. V.; Kulis, B.; Jain, P.; Sra, S.; and Dhillon, I. S Inforation-theoretic etric learning. In Proceedings of the 24th International Conference on Machine Learning (ICML), Fisher, R The use of ultiple easureents in taxonoic probles. Annals of Huan Genetics 7(2): Geng, B.; Tao, D.; and Xu, C. 20. DAML: Doain adaptation etric learning. IEEE Transactions on Iage Process 20(0): Globerson, A., and Roweis S Metric learning by collapsing classes. In Proceedings of the 20th Annual Conference on Advances in Neural Inforation Processing Systes (NIPS), Gretton, A.; Borgwardt, K. M.; Rasch, M. J.; Scholkopf, B.; and Sola, A. J A kernel ethod for the two-saple proble. In Proceedings of the 6th Annual Conference on Advances in Neural Inforation Processing Systes (NIPS). Jin, R.; Wang, S.; and Zhou, Y Regularized distance etric learning:theory and algorith. In Proceedings of the 23rd Annual Conference on Advances in Neural Inforation Processing Systes (NIPS), Jolliffe, I Principal Coponent Analysis. Springer- Verlag. Kondor, R. S., and Lafferty, J Diffusion kernels on graphs and other discrete input spaces. In Proceedings of the 9th International Conference on Machine Learning (ICML), Li, L.; Jin, X.; and Long, M Topic correlation analysis for cross-doain text classification. In Proceedings of the 26th AAAI Conference on Artificial Intelligence (AAAI). Long, M.; Wang, J.; Ding, G.; Sun, J.; and Yu, P. S Transfer Feature Learning with Joint Distribution Adaptation. In Proceedings of the 4th IEEE International Conference on Coputer Vision (ICCV). Long, M.; Wang, J.; Ding, G.; Shen, D.; and Yang, Q Transfer learning with graph co-regularization. In Proceedings of the 26th AAAI Conference on Artificial Intelligence (AAAI). Pan, S. J.; Kwok, J. T.; and Yang, Q Transfer learning via diensionality reduction. In Proceedings of the 23rd AAAI Conference on Artificial Intelligence (AAAI). Pan, S. J., and Yang, Q A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering 22: Pan, S. J.; Tsang, I. W.; Kwok, J. T.; and Yang, Q. 20. Doain adaptation via transfer coponent analysis. IEEE Transactions on Neural Networks 22(2): Phillips, J. P.; Moon, H.; Rizvi, S. A.; and Rauss, P. J The FERET evaluation ethodology for face-recognition algoriths. IEEE Transactions on Pattern Analysis and Machine Intelligence 22(0): Si, S.; Tao, D.; and Geng, B Bregan divergencebased regularization for transfer subspace learning. IEEE Transactions on Knowledge and Data Engineering 22(7): Sola, A., and Kondor, R Kernels and regularization on graphs. In Proceedings of the 6th Annual Conference on Learning Theory (COLT), Wang, C., and Mahadevan, S. 20. Heterogeneous doain adaptation using anifold alignent. In Proceedings of the 25th AAAI Conference on Artificial Intelligence (AAAI). Wang, S., and Jin, R An inforation geoetry approach for distance etric learning. In Proceedings of the 2nd International Conference on Artificial Intelligence and Statistics (AISTATS), Weinberger, K. Q.; Sha, F.; and Saul, L. K Learning a kernel atrix for nonlinear diensionality reduction. In Proceedings of the 2th International Conference on Machine Learning (ICML), Weinberger, K. Q., and Saul, L. K Distance etric learning for large argin nearest neighbor classification. Journal of Machine Learning Research 0: Xing, E. P.; Ng, A. Y.; Jordan, M. I.; and Russell, S. J Distance etric learning, with application to clustering with side-inforation. In Proceedings of the 6th Annual Conference on Advances in Neural Inforation Processing Systes (NIPS), Yang, L.; Jin, R.; Sukthankar, R.; and Liu, Y An efficient algorith for local distance etric learning. In Proceedings of the 2st AAAI Conference on Artificial Intelligence (AAAI),

Online Bagging and Boosting

Online Bagging and Boosting Abstract Bagging and boosting are two of the ost well-known enseble learning ethods due to their theoretical perforance guarantees and strong experiental results. However, these algoriths have been used

More information

Extended-Horizon Analysis of Pressure Sensitivities for Leak Detection in Water Distribution Networks: Application to the Barcelona Network

Extended-Horizon Analysis of Pressure Sensitivities for Leak Detection in Water Distribution Networks: Application to the Barcelona Network 2013 European Control Conference (ECC) July 17-19, 2013, Zürich, Switzerland. Extended-Horizon Analysis of Pressure Sensitivities for Leak Detection in Water Distribution Networks: Application to the Barcelona

More information

Image restoration for a rectangular poor-pixels detector

Image restoration for a rectangular poor-pixels detector Iage restoration for a rectangular poor-pixels detector Pengcheng Wen 1, Xiangjun Wang 1, Hong Wei 2 1 State Key Laboratory of Precision Measuring Technology and Instruents, Tianjin University, China 2

More information

Applying Multiple Neural Networks on Large Scale Data

Applying Multiple Neural Networks on Large Scale Data 0 International Conference on Inforation and Electronics Engineering IPCSIT vol6 (0) (0) IACSIT Press, Singapore Applying Multiple Neural Networks on Large Scale Data Kritsanatt Boonkiatpong and Sukree

More information

Machine Learning Applications in Grid Computing

Machine Learning Applications in Grid Computing Machine Learning Applications in Grid Coputing George Cybenko, Guofei Jiang and Daniel Bilar Thayer School of Engineering Dartouth College Hanover, NH 03755, USA gvc@dartouth.edu, guofei.jiang@dartouth.edu

More information

Media Adaptation Framework in Biofeedback System for Stroke Patient Rehabilitation

Media Adaptation Framework in Biofeedback System for Stroke Patient Rehabilitation Media Adaptation Fraework in Biofeedback Syste for Stroke Patient Rehabilitation Yinpeng Chen, Weiwei Xu, Hari Sundara, Thanassis Rikakis, Sheng-Min Liu Arts, Media and Engineering Progra Arizona State

More information

High Performance Chinese/English Mixed OCR with Character Level Language Identification

High Performance Chinese/English Mixed OCR with Character Level Language Identification 2009 0th International Conference on Docuent Analysis and Recognition High Perforance Chinese/English Mixed OCR with Character Level Language Identification Kai Wang Institute of Machine Intelligence,

More information

Stable Learning in Coding Space for Multi-Class Decoding and Its Extension for Multi-Class Hypothesis Transfer Learning

Stable Learning in Coding Space for Multi-Class Decoding and Its Extension for Multi-Class Hypothesis Transfer Learning Stable Learning in Coding Space for Multi-Class Decoding and Its Extension for Multi-Class Hypothesis Transfer Learning Bang Zhang, Yi Wang 2, Yang Wang, Fang Chen 2 National ICT Australia 2 School of

More information

CRM FACTORS ASSESSMENT USING ANALYTIC HIERARCHY PROCESS

CRM FACTORS ASSESSMENT USING ANALYTIC HIERARCHY PROCESS 641 CRM FACTORS ASSESSMENT USING ANALYTIC HIERARCHY PROCESS Marketa Zajarosova 1* *Ph.D. VSB - Technical University of Ostrava, THE CZECH REPUBLIC arketa.zajarosova@vsb.cz Abstract Custoer relationship

More information

The Research of Measuring Approach and Energy Efficiency for Hadoop Periodic Jobs

The Research of Measuring Approach and Energy Efficiency for Hadoop Periodic Jobs Send Orders for Reprints to reprints@benthascience.ae 206 The Open Fuels & Energy Science Journal, 2015, 8, 206-210 Open Access The Research of Measuring Approach and Energy Efficiency for Hadoop Periodic

More information

Reliability Constrained Packet-sizing for Linear Multi-hop Wireless Networks

Reliability Constrained Packet-sizing for Linear Multi-hop Wireless Networks Reliability Constrained acket-sizing for inear Multi-hop Wireless Networks Ning Wen, and Randall A. Berry Departent of Electrical Engineering and Coputer Science Northwestern University, Evanston, Illinois

More information

Software Quality Characteristics Tested For Mobile Application Development

Software Quality Characteristics Tested For Mobile Application Development Thesis no: MGSE-2015-02 Software Quality Characteristics Tested For Mobile Application Developent Literature Review and Epirical Survey WALEED ANWAR Faculty of Coputing Blekinge Institute of Technology

More information

Analyzing Spatiotemporal Characteristics of Education Network Traffic with Flexible Multiscale Entropy

Analyzing Spatiotemporal Characteristics of Education Network Traffic with Flexible Multiscale Entropy Vol. 9, No. 5 (2016), pp.303-312 http://dx.doi.org/10.14257/ijgdc.2016.9.5.26 Analyzing Spatioteporal Characteristics of Education Network Traffic with Flexible Multiscale Entropy Chen Yang, Renjie Zhou

More information

PERFORMANCE METRICS FOR THE IT SERVICES PORTFOLIO

PERFORMANCE METRICS FOR THE IT SERVICES PORTFOLIO Bulletin of the Transilvania University of Braşov Series I: Engineering Sciences Vol. 4 (53) No. - 0 PERFORMANCE METRICS FOR THE IT SERVICES PORTFOLIO V. CAZACU I. SZÉKELY F. SANDU 3 T. BĂLAN Abstract:

More information

Performance Evaluation of Machine Learning Techniques using Software Cost Drivers

Performance Evaluation of Machine Learning Techniques using Software Cost Drivers Perforance Evaluation of Machine Learning Techniques using Software Cost Drivers Manas Gaur Departent of Coputer Engineering, Delhi Technological University Delhi, India ABSTRACT There is a treendous rise

More information

Multi-Class Deep Boosting

Multi-Class Deep Boosting Multi-Class Deep Boosting Vitaly Kuznetsov Courant Institute 25 Mercer Street New York, NY 002 vitaly@cis.nyu.edu Mehryar Mohri Courant Institute & Google Research 25 Mercer Street New York, NY 002 ohri@cis.nyu.edu

More information

An Integrated Approach for Monitoring Service Level Parameters of Software-Defined Networking

An Integrated Approach for Monitoring Service Level Parameters of Software-Defined Networking International Journal of Future Generation Counication and Networking Vol. 8, No. 6 (15), pp. 197-4 http://d.doi.org/1.1457/ijfgcn.15.8.6.19 An Integrated Approach for Monitoring Service Level Paraeters

More information

ASIC Design Project Management Supported by Multi Agent Simulation

ASIC Design Project Management Supported by Multi Agent Simulation ASIC Design Project Manageent Supported by Multi Agent Siulation Jana Blaschke, Christian Sebeke, Wolfgang Rosenstiel Abstract The coplexity of Application Specific Integrated Circuits (ASICs) is continuously

More information

An improved TF-IDF approach for text classification *

An improved TF-IDF approach for text classification * Zhang et al. / J Zheiang Univ SCI 2005 6A(1:49-55 49 Journal of Zheiang University SCIECE ISS 1009-3095 http://www.zu.edu.cn/zus E-ail: zus@zu.edu.cn An iproved TF-IDF approach for text classification

More information

Managing Complex Network Operation with Predictive Analytics

Managing Complex Network Operation with Predictive Analytics Managing Coplex Network Operation with Predictive Analytics Zhenyu Huang, Pak Chung Wong, Patrick Mackey, Yousu Chen, Jian Ma, Kevin Schneider, and Frank L. Greitzer Pacific Northwest National Laboratory

More information

An Innovate Dynamic Load Balancing Algorithm Based on Task

An Innovate Dynamic Load Balancing Algorithm Based on Task An Innovate Dynaic Load Balancing Algorith Based on Task Classification Hong-bin Wang,,a, Zhi-yi Fang, b, Guan-nan Qu,*,c, Xiao-dan Ren,d College of Coputer Science and Technology, Jilin University, Changchun

More information

AUC Optimization vs. Error Rate Minimization

AUC Optimization vs. Error Rate Minimization AUC Optiization vs. Error Rate Miniization Corinna Cortes and Mehryar Mohri AT&T Labs Research 180 Park Avenue, Florha Park, NJ 0793, USA {corinna, ohri}@research.att.co Abstract The area under an ROC

More information

AN ALGORITHM FOR REDUCING THE DIMENSION AND SIZE OF A SAMPLE FOR DATA EXPLORATION PROCEDURES

AN ALGORITHM FOR REDUCING THE DIMENSION AND SIZE OF A SAMPLE FOR DATA EXPLORATION PROCEDURES Int. J. Appl. Math. Coput. Sci., 2014, Vol. 24, No. 1, 133 149 DOI: 10.2478/acs-2014-0011 AN ALGORITHM FOR REDUCING THE DIMENSION AND SIZE OF A SAMPLE FOR DATA EXPLORATION PROCEDURES PIOTR KULCZYCKI,,

More information

arxiv:0805.1434v1 [math.pr] 9 May 2008

arxiv:0805.1434v1 [math.pr] 9 May 2008 Degree-distribution stability of scale-free networs Zhenting Hou, Xiangxing Kong, Dinghua Shi,2, and Guanrong Chen 3 School of Matheatics, Central South University, Changsha 40083, China 2 Departent of

More information

Preference-based Search and Multi-criteria Optimization

Preference-based Search and Multi-criteria Optimization Fro: AAAI-02 Proceedings. Copyright 2002, AAAI (www.aaai.org). All rights reserved. Preference-based Search and Multi-criteria Optiization Ulrich Junker ILOG 1681, route des Dolines F-06560 Valbonne ujunker@ilog.fr

More information

RECURSIVE DYNAMIC PROGRAMMING: HEURISTIC RULES, BOUNDING AND STATE SPACE REDUCTION. Henrik Kure

RECURSIVE DYNAMIC PROGRAMMING: HEURISTIC RULES, BOUNDING AND STATE SPACE REDUCTION. Henrik Kure RECURSIVE DYNAMIC PROGRAMMING: HEURISTIC RULES, BOUNDING AND STATE SPACE REDUCTION Henrik Kure Dina, Danish Inforatics Network In the Agricultural Sciences Royal Veterinary and Agricultural University

More information

Dynamic Placement for Clustered Web Applications

Dynamic Placement for Clustered Web Applications Dynaic laceent for Clustered Web Applications A. Karve, T. Kibrel, G. acifici, M. Spreitzer, M. Steinder, M. Sviridenko, and A. Tantawi IBM T.J. Watson Research Center {karve,kibrel,giovanni,spreitz,steinder,sviri,tantawi}@us.ib.co

More information

Factored Models for Probabilistic Modal Logic

Factored Models for Probabilistic Modal Logic Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence (2008 Factored Models for Probabilistic Modal Logic Afsaneh Shirazi and Eyal Air Coputer Science Departent, University of Illinois

More information

Searching strategy for multi-target discovery in wireless networks

Searching strategy for multi-target discovery in wireless networks Searching strategy for ulti-target discovery in wireless networks Zhao Cheng, Wendi B. Heinzelan Departent of Electrical and Coputer Engineering University of Rochester Rochester, NY 467 (585) 75-{878,

More information

Binary Embedding: Fundamental Limits and Fast Algorithm

Binary Embedding: Fundamental Limits and Fast Algorithm Binary Ebedding: Fundaental Liits and Fast Algorith Xinyang Yi The University of Texas at Austin yixy@utexas.edu Eric Price The University of Texas at Austin ecprice@cs.utexas.edu Constantine Caraanis

More information

International Journal of Management & Information Systems First Quarter 2012 Volume 16, Number 1

International Journal of Management & Information Systems First Quarter 2012 Volume 16, Number 1 International Journal of Manageent & Inforation Systes First Quarter 2012 Volue 16, Nuber 1 Proposal And Effectiveness Of A Highly Copelling Direct Mail Method - Establishent And Deployent Of PMOS-DM Hisatoshi

More information

Evaluating Inventory Management Performance: a Preliminary Desk-Simulation Study Based on IOC Model

Evaluating Inventory Management Performance: a Preliminary Desk-Simulation Study Based on IOC Model Evaluating Inventory Manageent Perforance: a Preliinary Desk-Siulation Study Based on IOC Model Flora Bernardel, Roberto Panizzolo, and Davide Martinazzo Abstract The focus of this study is on preliinary

More information

A Scalable Application Placement Controller for Enterprise Data Centers

A Scalable Application Placement Controller for Enterprise Data Centers W WWW 7 / Track: Perforance and Scalability A Scalable Application Placeent Controller for Enterprise Data Centers Chunqiang Tang, Malgorzata Steinder, Michael Spreitzer, and Giovanni Pacifici IBM T.J.

More information

Use of extrapolation to forecast the working capital in the mechanical engineering companies

Use of extrapolation to forecast the working capital in the mechanical engineering companies ECONTECHMOD. AN INTERNATIONAL QUARTERLY JOURNAL 2014. Vol. 1. No. 1. 23 28 Use of extrapolation to forecast the working capital in the echanical engineering copanies A. Cherep, Y. Shvets Departent of finance

More information

Cooperative Caching for Adaptive Bit Rate Streaming in Content Delivery Networks

Cooperative Caching for Adaptive Bit Rate Streaming in Content Delivery Networks Cooperative Caching for Adaptive Bit Rate Streaing in Content Delivery Networs Phuong Luu Vo Departent of Coputer Science and Engineering, International University - VNUHCM, Vietna vtlphuong@hciu.edu.vn

More information

6. Time (or Space) Series Analysis

6. Time (or Space) Series Analysis ATM 55 otes: Tie Series Analysis - Section 6a Page 8 6. Tie (or Space) Series Analysis In this chapter we will consider soe coon aspects of tie series analysis including autocorrelation, statistical prediction,

More information

Real Time Target Tracking with Binary Sensor Networks and Parallel Computing

Real Time Target Tracking with Binary Sensor Networks and Parallel Computing Real Tie Target Tracking with Binary Sensor Networks and Parallel Coputing Hong Lin, John Rushing, Sara J. Graves, Steve Tanner, and Evans Criswell Abstract A parallel real tie data fusion and target tracking

More information

Comment on On Discriminative vs. Generative Classifiers: A Comparison of Logistic Regression and Naive Bayes

Comment on On Discriminative vs. Generative Classifiers: A Comparison of Logistic Regression and Naive Bayes Coent on On Discriinative vs. Generative Classifiers: A Coparison of Logistic Regression and Naive Bayes Jing-Hao Xue (jinghao@stats.gla.ac.uk) and D. Michael Titterington (ike@stats.gla.ac.uk) Departent

More information

INTEGRATED ENVIRONMENT FOR STORING AND HANDLING INFORMATION IN TASKS OF INDUCTIVE MODELLING FOR BUSINESS INTELLIGENCE SYSTEMS

INTEGRATED ENVIRONMENT FOR STORING AND HANDLING INFORMATION IN TASKS OF INDUCTIVE MODELLING FOR BUSINESS INTELLIGENCE SYSTEMS Artificial Intelligence Methods and Techniques for Business and Engineering Applications 210 INTEGRATED ENVIRONMENT FOR STORING AND HANDLING INFORMATION IN TASKS OF INDUCTIVE MODELLING FOR BUSINESS INTELLIGENCE

More information

Work Travel and Decision Probling in the Network Marketing World

Work Travel and Decision Probling in the Network Marketing World TRB Paper No. 03-4348 WORK TRAVEL MODE CHOICE MODELING USING DATA MINING: DECISION TREES AND NEURAL NETWORKS Chi Xie Research Assistant Departent of Civil and Environental Engineering University of Massachusetts,

More information

Markovian inventory policy with application to the paper industry

Markovian inventory policy with application to the paper industry Coputers and Cheical Engineering 26 (2002) 1399 1413 www.elsevier.co/locate/copcheeng Markovian inventory policy with application to the paper industry K. Karen Yin a, *, Hu Liu a,1, Neil E. Johnson b,2

More information

This paper studies a rental firm that offers reusable products to price- and quality-of-service sensitive

This paper studies a rental firm that offers reusable products to price- and quality-of-service sensitive MANUFACTURING & SERVICE OPERATIONS MANAGEMENT Vol., No. 3, Suer 28, pp. 429 447 issn 523-464 eissn 526-5498 8 3 429 infors doi.287/so.7.8 28 INFORMS INFORMS holds copyright to this article and distributed

More information

Partitioning Data on Features or Samples in Communication-Efficient Distributed Optimization?

Partitioning Data on Features or Samples in Communication-Efficient Distributed Optimization? Partitioning Data on Features or Saples in Counication-Efficient Distributed Optiization? Chenxin Ma Industrial and Systes Engineering Lehigh University, USA ch54@lehigh.edu Martin Taáč Industrial and

More information

Fuzzy Sets in HR Management

Fuzzy Sets in HR Management Acta Polytechnica Hungarica Vol. 8, No. 3, 2011 Fuzzy Sets in HR Manageent Blanka Zeková AXIOM SW, s.r.o., 760 01 Zlín, Czech Republic blanka.zekova@sezna.cz Jana Talašová Faculty of Science, Palacký Univerzity,

More information

Support Vector Machine Soft Margin Classifiers: Error Analysis

Support Vector Machine Soft Margin Classifiers: Error Analysis Journal of Machine Learning Research? (2004)?-?? Subitted 9/03; Published??/04 Support Vector Machine Soft Margin Classifiers: Error Analysis Di-Rong Chen Departent of Applied Matheatics Beijing University

More information

Partitioned Elias-Fano Indexes

Partitioned Elias-Fano Indexes Partitioned Elias-ano Indexes Giuseppe Ottaviano ISTI-CNR, Pisa giuseppe.ottaviano@isti.cnr.it Rossano Venturini Dept. of Coputer Science, University of Pisa rossano@di.unipi.it ABSTRACT The Elias-ano

More information

Resource Allocation in Wireless Networks with Multiple Relays

Resource Allocation in Wireless Networks with Multiple Relays Resource Allocation in Wireless Networks with Multiple Relays Kağan Bakanoğlu, Stefano Toasin, Elza Erkip Departent of Electrical and Coputer Engineering, Polytechnic Institute of NYU, Brooklyn, NY, 0

More information

Information Processing Letters

Information Processing Letters Inforation Processing Letters 111 2011) 178 183 Contents lists available at ScienceDirect Inforation Processing Letters www.elsevier.co/locate/ipl Offline file assignents for online load balancing Paul

More information

An Optimal Task Allocation Model for System Cost Analysis in Heterogeneous Distributed Computing Systems: A Heuristic Approach

An Optimal Task Allocation Model for System Cost Analysis in Heterogeneous Distributed Computing Systems: A Heuristic Approach An Optial Tas Allocation Model for Syste Cost Analysis in Heterogeneous Distributed Coputing Systes: A Heuristic Approach P. K. Yadav Central Building Research Institute, Rooree- 247667, Uttarahand (INDIA)

More information

Audio Engineering Society. Convention Paper. Presented at the 119th Convention 2005 October 7 10 New York, New York USA

Audio Engineering Society. Convention Paper. Presented at the 119th Convention 2005 October 7 10 New York, New York USA Audio Engineering Society Convention Paper Presented at the 119th Convention 2005 October 7 10 New York, New York USA This convention paper has been reproduced fro the authors advance anuscript, without

More information

Energy Proportionality for Disk Storage Using Replication

Energy Proportionality for Disk Storage Using Replication Energy Proportionality for Disk Storage Using Replication Jinoh Ki and Doron Rote Lawrence Berkeley National Laboratory University of California, Berkeley, CA 94720 {jinohki,d rote}@lbl.gov Abstract Energy

More information

CLOSED-LOOP SUPPLY CHAIN NETWORK OPTIMIZATION FOR HONG KONG CARTRIDGE RECYCLING INDUSTRY

CLOSED-LOOP SUPPLY CHAIN NETWORK OPTIMIZATION FOR HONG KONG CARTRIDGE RECYCLING INDUSTRY CLOSED-LOOP SUPPLY CHAIN NETWORK OPTIMIZATION FOR HONG KONG CARTRIDGE RECYCLING INDUSTRY Y. T. Chen Departent of Industrial and Systes Engineering Hong Kong Polytechnic University, Hong Kong yongtong.chen@connect.polyu.hk

More information

Online Community Detection for Large Complex Networks

Online Community Detection for Large Complex Networks Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence Online Counity Detection for Large Coplex Networks Wangsheng Zhang, Gang Pan, Zhaohui Wu, Shijian Li Departent

More information

Optimal Resource-Constraint Project Scheduling with Overlapping Modes

Optimal Resource-Constraint Project Scheduling with Overlapping Modes Optial Resource-Constraint Proect Scheduling with Overlapping Modes François Berthaut Lucas Grèze Robert Pellerin Nathalie Perrier Adnène Hai February 20 CIRRELT-20-09 Bureaux de Montréal : Bureaux de

More information

Capacity of Multiple-Antenna Systems With Both Receiver and Transmitter Channel State Information

Capacity of Multiple-Antenna Systems With Both Receiver and Transmitter Channel State Information IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 49, NO., OCTOBER 23 2697 Capacity of Multiple-Antenna Systes With Both Receiver and Transitter Channel State Inforation Sudharan K. Jayaweera, Student Meber,

More information

Modeling Parallel Applications Performance on Heterogeneous Systems

Modeling Parallel Applications Performance on Heterogeneous Systems Modeling Parallel Applications Perforance on Heterogeneous Systes Jaeela Al-Jaroodi, Nader Mohaed, Hong Jiang and David Swanson Departent of Coputer Science and Engineering University of Nebraska Lincoln

More information

2. FINDING A SOLUTION

2. FINDING A SOLUTION The 7 th Balan Conference on Operational Research BACOR 5 Constanta, May 5, Roania OPTIMAL TIME AND SPACE COMPLEXITY ALGORITHM FOR CONSTRUCTION OF ALL BINARY TREES FROM PRE-ORDER AND POST-ORDER TRAVERSALS

More information

Trading Regret for Efficiency: Online Convex Optimization with Long Term Constraints

Trading Regret for Efficiency: Online Convex Optimization with Long Term Constraints Journal of Machine Learning Research 13 2012) 2503-2528 Subitted 8/11; Revised 3/12; Published 9/12 rading Regret for Efficiency: Online Convex Optiization with Long er Constraints Mehrdad Mahdavi Rong

More information

Energy Efficient VM Scheduling for Cloud Data Centers: Exact allocation and migration algorithms

Energy Efficient VM Scheduling for Cloud Data Centers: Exact allocation and migration algorithms Energy Efficient VM Scheduling for Cloud Data Centers: Exact allocation and igration algoriths Chaia Ghribi, Makhlouf Hadji and Djaal Zeghlache Institut Mines-Téléco, Téléco SudParis UMR CNRS 5157 9, Rue

More information

Exploiting Hardware Heterogeneity within the Same Instance Type of Amazon EC2

Exploiting Hardware Heterogeneity within the Same Instance Type of Amazon EC2 Exploiting Hardware Heterogeneity within the Sae Instance Type of Aazon EC2 Zhonghong Ou, Hao Zhuang, Jukka K. Nurinen, Antti Ylä-Jääski, Pan Hui Aalto University, Finland; Deutsch Teleko Laboratories,

More information

Airline Yield Management with Overbooking, Cancellations, and No-Shows JANAKIRAM SUBRAMANIAN

Airline Yield Management with Overbooking, Cancellations, and No-Shows JANAKIRAM SUBRAMANIAN Airline Yield Manageent with Overbooking, Cancellations, and No-Shows JANAKIRAM SUBRAMANIAN Integral Developent Corporation, 301 University Avenue, Suite 200, Palo Alto, California 94301 SHALER STIDHAM

More information

Botnets Detection Based on IRC-Community

Botnets Detection Based on IRC-Community Botnets Detection Based on IRC-Counity Wei Lu and Ali A. Ghorbani Network Security Laboratory, Faculty of Coputer Science University of New Brunswick, Fredericton, NB E3B 5A3, Canada {wlu, ghorbani}@unb.ca

More information

Data Set Generation for Rectangular Placement Problems

Data Set Generation for Rectangular Placement Problems Data Set Generation for Rectangular Placeent Probles Christine L. Valenzuela (Muford) Pearl Y. Wang School of Coputer Science & Inforatics Departent of Coputer Science MS 4A5 Cardiff University George

More information

Reconnect 04 Solving Integer Programs with Branch and Bound (and Branch and Cut)

Reconnect 04 Solving Integer Programs with Branch and Bound (and Branch and Cut) Sandia is a ultiprogra laboratory operated by Sandia Corporation, a Lockheed Martin Copany, Reconnect 04 Solving Integer Progras with Branch and Bound (and Branch and Cut) Cynthia Phillips (Sandia National

More information

Quality evaluation of the model-based forecasts of implied volatility index

Quality evaluation of the model-based forecasts of implied volatility index Quality evaluation of the odel-based forecasts of iplied volatility index Katarzyna Łęczycka 1 Abstract Influence of volatility on financial arket forecasts is very high. It appears as a specific factor

More information

A framework for performance monitoring, load balancing, adaptive timeouts and quality of service in digital libraries

A framework for performance monitoring, load balancing, adaptive timeouts and quality of service in digital libraries Int J Digit Libr (2000) 3: 9 35 INTERNATIONAL JOURNAL ON Digital Libraries Springer-Verlag 2000 A fraework for perforance onitoring, load balancing, adaptive tieouts and quality of service in digital libraries

More information

Equivalent Tapped Delay Line Channel Responses with Reduced Taps

Equivalent Tapped Delay Line Channel Responses with Reduced Taps Equivalent Tapped Delay Line Channel Responses with Reduced Taps Shweta Sagari, Wade Trappe, Larry Greenstein {shsagari, trappe, ljg}@winlab.rutgers.edu WINLAB, Rutgers University, North Brunswick, NJ

More information

Adaptive Modulation and Coding for Unmanned Aerial Vehicle (UAV) Radio Channel

Adaptive Modulation and Coding for Unmanned Aerial Vehicle (UAV) Radio Channel Recent Advances in Counications Adaptive odulation and Coding for Unanned Aerial Vehicle (UAV) Radio Channel Airhossein Fereidountabar,Gian Carlo Cardarilli, Rocco Fazzolari,Luca Di Nunzio Abstract In

More information

Incorporating Complex Substitution Patterns and Variance Scaling in Long Distance Travel Choice Behavior

Incorporating Complex Substitution Patterns and Variance Scaling in Long Distance Travel Choice Behavior Incorporating Coplex Substitution Patterns and Variance Scaling in Long Distance Travel Choice Behavior Frank S. Koppelan Professor of Civil Engineering and Transportation Northwestern University Evanston,

More information

An Approach to Combating Free-riding in Peer-to-Peer Networks

An Approach to Combating Free-riding in Peer-to-Peer Networks An Approach to Cobating Free-riding in Peer-to-Peer Networks Victor Ponce, Jie Wu, and Xiuqi Li Departent of Coputer Science and Engineering Florida Atlantic University Boca Raton, FL 33431 April 7, 2008

More information

Impact of Processing Costs on Service Chain Placement in Network Functions Virtualization

Impact of Processing Costs on Service Chain Placement in Network Functions Virtualization Ipact of Processing Costs on Service Chain Placeent in Network Functions Virtualization Marco Savi, Massio Tornatore, Giacoo Verticale Dipartiento di Elettronica, Inforazione e Bioingegneria, Politecnico

More information

Efficient Key Management for Secure Group Communications with Bursty Behavior

Efficient Key Management for Secure Group Communications with Bursty Behavior Efficient Key Manageent for Secure Group Counications with Bursty Behavior Xukai Zou, Byrav Raaurthy Departent of Coputer Science and Engineering University of Nebraska-Lincoln Lincoln, NE68588, USA Eail:

More information

Budget-optimal Crowdsourcing using Low-rank Matrix Approximations

Budget-optimal Crowdsourcing using Low-rank Matrix Approximations Budget-optial Crowdsourcing using Low-rank Matrix Approxiations David R. Karger, Sewoong Oh, and Devavrat Shah Departent of EECS, Massachusetts Institute of Technology Eail: {karger, swoh, devavrat}@it.edu

More information

Method of supply chain optimization in E-commerce

Method of supply chain optimization in E-commerce MPRA Munich Personal RePEc Archive Method of supply chain optiization in E-coerce Petr Suchánek and Robert Bucki Silesian University - School of Business Adinistration, The College of Inforatics and Manageent

More information

Subspace Analysis and Optimization for AAM Based Face Alignment

Subspace Analysis and Optimization for AAM Based Face Alignment Subspace Analysis and Optimization for AAM Based Face Alignment Ming Zhao Chun Chen College of Computer Science Zhejiang University Hangzhou, 310027, P.R.China zhaoming1999@zju.edu.cn Stan Z. Li Microsoft

More information

Entity Search Engine: Towards Agile Best-Effort Information Integration over the Web

Entity Search Engine: Towards Agile Best-Effort Information Integration over the Web Entity Search Engine: Towards Agile Best-Effort Inforation Integration over the Web Tao Cheng, Kevin Chen-Chuan Chang University of Illinois at Urbana-Chapaign {tcheng3, kcchang}@cs.uiuc.edu. INTRODUCTION

More information

PREDICTION OF POSSIBLE CONGESTIONS IN SLA CREATION PROCESS

PREDICTION OF POSSIBLE CONGESTIONS IN SLA CREATION PROCESS PREDICTIO OF POSSIBLE COGESTIOS I SLA CREATIO PROCESS Srećko Krile University of Dubrovnik Departent of Electrical Engineering and Coputing Cira Carica 4, 20000 Dubrovnik, Croatia Tel +385 20 445-739,

More information

The Application of Bandwidth Optimization Technique in SLA Negotiation Process

The Application of Bandwidth Optimization Technique in SLA Negotiation Process The Application of Bandwidth Optiization Technique in SLA egotiation Process Srecko Krile University of Dubrovnik Departent of Electrical Engineering and Coputing Cira Carica 4, 20000 Dubrovnik, Croatia

More information

Evaluating the Effectiveness of Task Overlapping as a Risk Response Strategy in Engineering Projects

Evaluating the Effectiveness of Task Overlapping as a Risk Response Strategy in Engineering Projects Evaluating the Effectiveness of Task Overlapping as a Risk Response Strategy in Engineering Projects Lucas Grèze Robert Pellerin Nathalie Perrier Patrice Leclaire February 2011 CIRRELT-2011-11 Bureaux

More information

Modeling Nurse Scheduling Problem Using 0-1 Goal Programming: A Case Study Of Tafo Government Hospital, Kumasi-Ghana

Modeling Nurse Scheduling Problem Using 0-1 Goal Programming: A Case Study Of Tafo Government Hospital, Kumasi-Ghana Modeling Nurse Scheduling Proble Using 0-1 Goal Prograing: A Case Study Of Tafo Governent Hospital, Kuasi-Ghana Wallace Agyei, Willia Obeng-Denteh, Eanuel A. Andaa Abstract: The proble of scheduling nurses

More information

Online Methods for Multi-Domain Learning and Adaptation

Online Methods for Multi-Domain Learning and Adaptation Online Methods for Multi-Doain Learning and Adaptation Mark Dredze and Koby Craer Departent of Coputer and Inforation Science University of Pennsylvania Philadelphia, PA 19104 USA {dredze,craer}@cis.upenn.edu

More information

CPU Animation. Introduction. CPU skinning. CPUSkin Scalar:

CPU Animation. Introduction. CPU skinning. CPUSkin Scalar: CPU Aniation Introduction The iportance of real-tie character aniation has greatly increased in odern gaes. Aniating eshes ia 'skinning' can be perfored on both a general purpose CPU and a ore specialized

More information

Standards and Protocols for the Collection and Dissemination of Graduating Student Initial Career Outcomes Information For Undergraduates

Standards and Protocols for the Collection and Dissemination of Graduating Student Initial Career Outcomes Information For Undergraduates National Association of Colleges and Eployers Standards and Protocols for the Collection and Disseination of Graduating Student Initial Career Outcoes Inforation For Undergraduates Developed by the NACE

More information

Evaluating Software Quality of Vendors using Fuzzy Analytic Hierarchy Process

Evaluating Software Quality of Vendors using Fuzzy Analytic Hierarchy Process IMECS 2008 9-2 March 2008 Hong Kong Evaluating Software Quality of Vendors using Fuzzy Analytic Hierarchy Process Kevin K.F. Yuen* Henry C.W. au Abstract This paper proposes a fuzzy Analytic Hierarchy

More information

Automatic Name-Face Alignment to Enable Cross-Media News Retrieval

Automatic Name-Face Alignment to Enable Cross-Media News Retrieval Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence Autoatic ae-face Alignent to Enable Cross-Media ews Retrieval Yueie Zhang*, Wei Wu*, Yang Li*, Cheng Jin*, Xiangyang

More information

Data Streaming Algorithms for Estimating Entropy of Network Traffic

Data Streaming Algorithms for Estimating Entropy of Network Traffic Data Streaing Algoriths for Estiating Entropy of Network Traffic Ashwin Lall University of Rochester Vyas Sekar Carnegie Mellon University Mitsunori Ogihara University of Rochester Jun (Ji) Xu Georgia

More information

The individual neurons are complicated. They have a myriad of parts, subsystems and control mechanisms. They convey information via a host of

The individual neurons are complicated. They have a myriad of parts, subsystems and control mechanisms. They convey information via a host of CHAPTER 4 ARTIFICIAL NEURAL NETWORKS 4. INTRODUCTION Artificial Neural Networks (ANNs) are relatively crude electronic odels based on the neural structure of the brain. The brain learns fro experience.

More information

A short-term, pattern-based model for water-demand forecasting

A short-term, pattern-based model for water-demand forecasting 39 Q IWA Publishing 2007 Journal of Hydroinforatics 09.1 2007 A short-ter, pattern-based odel for water-deand forecasting Stefano Alvisi, Marco Franchini and Alberto Marinelli ABSTRACT The short-ter, deand-forecasting

More information

ABSTRACT KEYWORDS. Comonotonicity, dependence, correlation, concordance, copula, multivariate. 1. INTRODUCTION

ABSTRACT KEYWORDS. Comonotonicity, dependence, correlation, concordance, copula, multivariate. 1. INTRODUCTION MEASURING COMONOTONICITY IN M-DIMENSIONAL VECTORS BY INGE KOCH AND ANN DE SCHEPPER ABSTRACT In this contribution, a new easure of coonotonicity for -diensional vectors is introduced, with values between

More information

COMBINING CRASH RECORDER AND PAIRED COMPARISON TECHNIQUE: INJURY RISK FUNCTIONS IN FRONTAL AND REAR IMPACTS WITH SPECIAL REFERENCE TO NECK INJURIES

COMBINING CRASH RECORDER AND PAIRED COMPARISON TECHNIQUE: INJURY RISK FUNCTIONS IN FRONTAL AND REAR IMPACTS WITH SPECIAL REFERENCE TO NECK INJURIES COMBINING CRASH RECORDER AND AIRED COMARISON TECHNIQUE: INJURY RISK FUNCTIONS IN FRONTAL AND REAR IMACTS WITH SECIAL REFERENCE TO NECK INJURIES Anders Kullgren, Maria Krafft Folksa Research, 66 Stockhol,

More information

REQUIREMENTS FOR A COMPUTER SCIENCE CURRICULUM EMPHASIZING INFORMATION TECHNOLOGY SUBJECT AREA: CURRICULUM ISSUES

REQUIREMENTS FOR A COMPUTER SCIENCE CURRICULUM EMPHASIZING INFORMATION TECHNOLOGY SUBJECT AREA: CURRICULUM ISSUES REQUIREMENTS FOR A COMPUTER SCIENCE CURRICULUM EMPHASIZING INFORMATION TECHNOLOGY SUBJECT AREA: CURRICULUM ISSUES Charles Reynolds Christopher Fox reynolds @cs.ju.edu fox@cs.ju.edu Departent of Coputer

More information

The AGA Evaluating Model of Customer Loyalty Based on E-commerce Environment

The AGA Evaluating Model of Customer Loyalty Based on E-commerce Environment 6 JOURNAL OF SOFTWARE, VOL. 4, NO. 3, MAY 009 The AGA Evaluating Model of Custoer Loyalty Based on E-coerce Environent Shaoei Yang Econoics and Manageent Departent, North China Electric Power University,

More information

Visualizing Similarity Data with a Mixture of Maps

Visualizing Similarity Data with a Mixture of Maps Visualizing Siilarity Data with a Mixture of Maps Jaes Cook, Ilya Sutskever, Andriy Mnih and Geoffrey Hinton Departent of Coputer Science University of Toronto Toronto, Ontario M5S 3G4 located at: /h/71/hinton/papers/a/a1.tex

More information

Considerations on Distributed Load Balancing for Fully Heterogeneous Machines: Two Particular Cases

Considerations on Distributed Load Balancing for Fully Heterogeneous Machines: Two Particular Cases Considerations on Distributed Load Balancing for Fully Heterogeneous Machines: Two Particular Cases Nathanaël Cheriere Departent of Coputer Science ENS Rennes Rennes, France nathanael.cheriere@ens-rennes.fr

More information

On Computing Nearest Neighbors with Applications to Decoding of Binary Linear Codes

On Computing Nearest Neighbors with Applications to Decoding of Binary Linear Codes On Coputing Nearest Neighbors with Applications to Decoding of Binary Linear Codes Alexander May and Ilya Ozerov Horst Görtz Institute for IT-Security Ruhr-University Bochu, Gerany Faculty of Matheatics

More information

Modeling operational risk data reported above a time-varying threshold

Modeling operational risk data reported above a time-varying threshold Modeling operational risk data reported above a tie-varying threshold Pavel V. Shevchenko CSIRO Matheatical and Inforation Sciences, Sydney, Locked bag 7, North Ryde, NSW, 670, Australia. e-ail: Pavel.Shevchenko@csiro.au

More information

ESTIMATING LIQUIDITY PREMIA IN THE SPANISH GOVERNMENT SECURITIES MARKET

ESTIMATING LIQUIDITY PREMIA IN THE SPANISH GOVERNMENT SECURITIES MARKET ESTIMATING LIQUIDITY PREMIA IN THE SPANISH GOVERNMENT SECURITIES MARKET Francisco Alonso, Roberto Blanco, Ana del Río and Alicia Sanchis Banco de España Banco de España Servicio de Estudios Docuento de

More information

Lecture L26-3D Rigid Body Dynamics: The Inertia Tensor

Lecture L26-3D Rigid Body Dynamics: The Inertia Tensor J. Peraire, S. Widnall 16.07 Dynaics Fall 008 Lecture L6-3D Rigid Body Dynaics: The Inertia Tensor Version.1 In this lecture, we will derive an expression for the angular oentu of a 3D rigid body. We shall

More information

Towards Change Management Capability Assessment Model for Contractors in Building Project

Towards Change Management Capability Assessment Model for Contractors in Building Project Middle-East Journal of Scientific Research 23 (7): 1327-1333, 2015 ISSN 1990-9233 IDOSI Publications, 2015 DOI: 10.5829/idosi.ejsr.2015.23.07.120 Towards Change Manageent Capability Assessent Model for

More information

A CHAOS MODEL OF SUBHARMONIC OSCILLATIONS IN CURRENT MODE PWM BOOST CONVERTERS

A CHAOS MODEL OF SUBHARMONIC OSCILLATIONS IN CURRENT MODE PWM BOOST CONVERTERS A CHAOS MODEL OF SUBHARMONIC OSCILLATIONS IN CURRENT MODE PWM BOOST CONVERTERS Isaac Zafrany and Sa BenYaakov Departent of Electrical and Coputer Engineering BenGurion University of the Negev P. O. Box

More information