Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Preprints of the 9th International Symposium on Advanced Control of Chemical Processes The International Federation of Automatic Control June 7-10, 2015, Whistler, British Columbia, Canada MoM3.6 Nearest Correlation Louvain Method for Fast and Good Selection of Input Variables of Statistical Model ? Taku Uchimaru ∗ Koji Hazama ∗ Koichi Fujiwara ∗ Manabu Kano ∗ ∗ Department of Systems Science, Kyoto University, Kyoto, Japan (e-mail: [email protected]). Abstract: In the present work, a new input variable selection method for building linear regression models is proposed. The proposed method is referred to as nearest correlation Louvain method based variable selection (NCLM-VS). NCLM-VS is a correlation-based group-wise method; it constructs an affinity matrix of input variables by the nearest correlation (NC) method, partitions the affinity matrix by the Louvain method (LM), consequently clusters input variables into multiple variable classes, and finally selects variable classes according to their contribution to estimates. LM is very fast and optimizes the number of classes automatically unlike spectral clustering (SC). The advantage of NCLM-VS over conventional methods including NCSC-VS is demonstrated through their applications to soft-sensor design for an industrial chemical process and calibration modeling based on near-infrared (NIR) spectra. In particular, it is confirmed that NCLM-VS is significantly faster than the recently proposed NCSC-VS while NCLM-VS can achieve as good estimation performance as NCSC-VS. Keywords: Input variable selection, parameter estimation, prediction methods, modelling, spectral clustering, Louvain method, soft-sensor, near-infrared (NIR) spectroscopy 1. INTRODUCTION In the manufacturing industry, real-time control of product quality is crucial to supply high quality products. However, it is often difficult to measure the product quality in real time because of the high investment and maintenance cost of analyzers and also measurement delay. Thus, the methods of estimating difficult-to-measure product quality from easy-to-measure process variables have been investigated. For example, models to estimate product quality is known as soft-sensors in the chemical industry, where partial least squares (PLS) regression is widely used because PLS can avoid the problem of collinearity (Kano and Ogawa (2010); Kano and Fujiwara (2013)). In the semiconductor manufacturing industry, the models are called virtual metrology (VM), and PLS is regularly employed (Lynn et al. (2012)). The estimation accuracy of soft-sensors deteriorates when input variables which do not contribute toward estimating output variables are used. To construct a highly accurate soft-sensor, appropriate input variables must be selected. So far many input variable selection methods have been proposed. The simplest method is the R method, which is based on the correlation coefficients between input variables and the output variable. When PLS is adopted, one can use PLS-based variable selection methods such as the PLS-Beta method and the variable influence on projection (VIP) method. Lately, many other PLS-based ? This study is partially supported by Japan Society for the Promotion of Science (JSPS), Grant-in-Aid for Scientific Research (C) 24560940. Copyright © 2015 IFAC methods have been proposed: genetic algorithm combined with PLS regression (GA-PLS), sub-window permutation analysis coupled with PLS (SwPA-PLS), Sparse-PLS, and Powered PLS (PPLS) (Mehmood et al. (2012)). Recently, nearest correlation spectral clustering-based variable selection (NCSC-VS) was proposed (Fujiwara et al. (2012)). NCSC-VS clusters variables into some variable classes based on the correlation between variables by using NCSC (Fujiwara et al. (2010, 2011)), then each variable class is examined as to whether it should be used as input variables. It was demonstrated that NCSC-VS was superior to other conventional methods through applications to soft-sensor design problems. However, NCSC-VS requires heavy computational load to optimize parameters such as the number of variable classes. In the present work, to solve the problem of NCSCVS, a new input variable selection method is proposed. To reduce the computational load and to determine the number of variable classes automatically, the Louvain method (LM) is used instead of spectral clustering (SC). LM can partition an affinity matrix faster than SC and can optimize the number of clusters (Blondel et al. (2008)). The proposed method is called nearest correlation Louvain method-based variable selection (NCLM-VS). NCLM-VS clusters input variables into some variable classes based on NCLM, and selects fewer variable classes according to their contribution to estimates. Thus, NCLM-VS is a kind of correlation-based group-wise methods for selecting input variables. The usefulness of NCLM-VS is demonstrated through applications to soft-sensor design for an industrial 123 IFAC ADCHEM 2015 June 7-10, 2015, Whistler, British Columbia, Canada chemical process and calibration modeling based on nearinfrared (NIR) spectra. 2. CONVENTIONAL METHODS This section briefly explains conventional input variable selection methods. Two of them are based on PLS; therefore PLS is briefly explained at the beginning of this section. 2.1 Partial Least Squares (PLS) In PLS with one output variable, an input variable matrix X ∈ RN ×M and an output variable vector y ∈ RN are decomposed as follows. X =TPT + E (1) y =Tb + f (2) where T ∈ R is the latent variable score matrix whose columns are latent variables (score vectors) tk ∈ RN (k = 1, 2, · · · , K), P ∈ RM ×K is the loading matrix of X whose columns are loading vectors pk ∈ RM , and b ∈ RK is the loading vector of y whose values are bk (k = 1, 2, · · · , K). N and M are the numbers of samples and input variables, respectively. K denotes the number of adopted latent variables. E and f are errors. N ×K The latent variable t is a linear combination of the columns of X. kβselect k >µ (7) kβpls k is achieved. Here βselect denotes the vector of the regression coefficients corresponding to the selected variables. 2.4 Variable influence on projection (VIP) The VIP score expresses the contribution of each input variable to the output variable (Wold et al. (2001)). The VIP score of the jth input variable is defined as v u PK 2 2 2 uM k=1 bk ktk k (wjk /kwk k) Vj =t (8) PK 2 2 k=1 bk ktk k where wjk is the jth element of the weighting vector of the kth latent variable wk . The input variables satisfying Vj > µ are selected for modeling. 3. NEAREST CORRELATION SPECTRAL CLUSTERING (NCSC) NCSC was originally proposed for sample clustering based on the correlation among variables by not assuming any distribution (Fujiwara et al. (2010)). Later NCSC was used for variable clustering (Fujiwara et al. (2012)). t = Xw (3) M where the weighting vector w ∈ R is determined so that the covariance between y and t is maximized under the condition of kwk = 1. NCSC integrates the nearest correlation (NC) method, which detects samples whose correlation is similar to the query, and the spectral clustering (SC) method, which partitions a weighted graph into subgraphs. NC constructs the weighted graph that expresses the correlation-based similarities between samples and SC partitions the constructed graph. 2.2 R 3.1 Spectral clustering (SC) The R method selects input variables based on the absolute value of the correlation coefficient rm between each input variable and the output variable. SC is a clustering method based on the graph theory (Ding et al. (2001); Ng et al. (2002)). It partitions a weighted graph, whose weights express affinities between nodes, into subgraphs through cutting some of their arcs. Although some spectral clustering algorithms have been proposed, the max-min cut (Mcut) method is described in this subsection. s2x y rm = q m s2xm s2y (4) where s2xm y is the covariance of xm and y, and s2xm and s2y is the variance of xm and y, respectively. The input variables satisfying |rm | > µ are selected for modeling. 2.3 PLS-Beta Given a wighted graph G and its affinity matrix (adjacency matrix) W , G is partitioned into two subgraphs A and B. The affinity between A and B is defined as X cut(A, B) ≡ W (A, B) = Wa,b (9) a∈A,b∈B The PLS-Beta method selects input variables based on the magnitude of its regression coefficients. The estimate ŷ derived by PLS is expressed by ŷ = T (T T T )−1 T y = Xβpls and the regression coefficients βpls is described as βpls = W (P T W )−1 (T T T )−1 y M ×K where W = [w1 , · · · , wK ] ∈ R (5) (6) . Input variables can be selected individually in descending order of the magnitude of regression coefficients until Copyright © 2015 IFAC W (A) ≡ W (A, A) (10) where a and b denote nodes of subgraphs A and B, and Wa,b is the sum of weights of arcs between a and b. That is, the affinity between subgraphs cut(A, B) is the sum of weights of arcs between subgraphs. The Mcut method searches subgraphs A and B that minimize cut(A, B) and maximize W (A) and W (B) simultaneously. cut(A, B) cut(A, B) + (11) W (A) W (B) Since the Mcut method partitions a weighted graph into two subgraphs, this procedure is repeated when three or min J = 124 IFAC ADCHEM 2015 June 7-10, 2015, Whistler, British Columbia, Canada x7 x3 x2 x8 l x4 x7 x5 x2 x6 V x1 more subgraphs are needed. In SC, the definition of an affinity is arbitrary and affects results. SC runs in O(n3 ) time complexity, where n is the number of nodes in G. 3.2 Nearest correlation (NC) method The NC method can detect samples whose correlation is similar to the query without any teacher signals (Fujiwara et al. (2010)). The concept of NC is as follows. Suppose that the affine subspace P in Fig. 1 (left) shows the correlation among variables and all the samples on P have the same correlation. Although x1 , x2 , · · · , x5 have the same correlation, x6 and x7 have a different correlation from the others. The NC method aims to detect samples whose correlation is similar to the query x1 . In this example, x2 , x3 , x4 , x5 on P should be detected. First, the whole space is translated so that the query becomes the origin as shown in Fig. 1 (right). That is, x1 is subtracted from all other samples xi (i = 2, 3, · · · , 7). Since the translated affine subspace contains the origin, it becomes the linear subspace V . Next, a line connecting each sample and the origin is drawn virtually. Then, it is checked whether another sample can be found on this line. In this case, x2 –x5 and x3 –x4 satisfy such a relationship. The correlation coefficients of these pairs must be 1 or −1. On the other hand, x6 and x7 that are not the elements of V cannot make such pairs. The pairs whose correlation coefficients are ±1 are thought to have the same correlation as x1 . In practice, the threshold of correlation coefficients γ (0 < γ ≤ 1) has to be used, since there are no pairs whose correlation coefficient is strictly ±1. Thus, pairs are selected when the absolute values of their correlation coefficients are larger than γ. The pairs whose correlation is similar to the query can be detected through the above procedure. NC method runs in O(n2 ) time complexity, where n is the number of samples. 3.3 Nearest correlation spectral clustering (NCSC) A correlation-based affinity matrix for spectral clustering can be constructed by using the NC method. Suppose that samples xn ∈ <M (n = 1, 2, · · · , N ) are stored in the database. The NCSC method is illustrated through a simple example (Fujiwara et al. (2010)). An objective data set consists of nine samples on the two dimensional space; x1 , x2 , x3 , x4 x2 x5 x3 Translated space Fig. 1. An example of the procedure of the NC method (Fujiwara et al. (2012)) Copyright © 2015 IFAC k x9 x6 x1 x6 Original space x7 x5 x3 x1 P x4 x4 Fig. 2. An objective data set consisting of two classes and one outlier (Fujiwara et al. (2012)) are on the line l and x5 , x6 , x7 , x8 are on the line k while x9 is an outlier as shown in Fig. 2. In addition, samples x2 , x7 and x9 are on one line by chance. First, the zero matrix S ∈ <9×9 is defined, and the correlations of all possible pairs of samples are checked by the NC method. For example, when x1 is the query, x2 –x3 , x2 –x4 and x3 –x4 are detected as pairs whose correlations can describe x1 . The detected pairs get one point; S2,3 = S3,2 = 1, S2,4 = S4,2 = 1 and S3,4 = S4,3 = 1. Similarly, since x1 –x3 , x1 –x4 , x3 –x4 and x7 –x9 are detected when x2 is the query, one is added to the elements of S corresponding to these pairs. This procedure is repeated until all samples become the query. Finally, the affinity matrix S becomes as follows. 0 2 2 2 S= 0 0 0 0 2 0 2 2 0 0 1 0 0 1 2 2 0 2 0 0 0 0 0 2 2 2 0 0 0 0 0 0 0 0 0 0 0 2 2 2 0 0 0 0 0 2 0 2 2 0 0 1 0 0 2 2 0 2 1 0 0 0 0 2 2 2 0 0 0 1 0 0 0 . 0 1 0 (12) 0 By applying SC to S, the nodes (samples) are classified into {x1 , x2 , x3 , x4 }, {x5 , x6 , x7 , x8 } and {x9 }. This example clearly shows that NCSC can classify samples on the basis of the correlation among variables. 3.4 Nearest correlation spectral clustering-based variable selection (NCSC-VS) Although NCSC was originally proposed for sample clustering, it can be used for variable clustering (Fujiwara et al. (2012)). NCSC-VS clusters input variables into some variable classes based on their correlation by using NCSC, and evaluates each variable class for variable selection. First, NCSC-VS clusters input variables into J variable classes vj = {xm | m ⊂ Vj } (j = 1, 2, · · · , J) based on their correlation, P where Vj is the jth subset of variable indexes, and V = j Vj . To classify input variables by NCSC, an affinity matrix is constructed from the transposed variable matrix X T by using NC. Then, the jth PLS model fj is constructed from the jth variable class matrix Xj , whose columns consist of variables belonging to the jth variable class, and the validity of the constructed PLS model is examined on the basis of the contribution ratio defined as 125 IFAC ADCHEM 2015 June 7-10, 2015, Whistler, British Columbia, Canada kŷj,k k2 (13) kyk2 where ŷj,k denotes the estimates by fj with k latent variables. Finally, variable classes are selected in descending order of Cj,k so that the estimation performance is maximized. A statistical model can be constructed by using the variables belonging to the selected variable classes. Cj,k = 1 − 4. NEAREST CORRELATION LOUVAIN METHOD (NCLM) In the present work, nearest correlation Louvain (NCLM) method is proposed by integrating the NC method and the Louvain method (LM). LM can partition the affinity matrix much faster than SC, and also it can optimize the number of classes in its algorithm. Thus, NCLM can divide input variables into classes based on the correlation among variables more efficiently than NCSC. 4.1 Louvain method (LM) The Louvain method (LM) was originally developed in the field of community detection (Blondel et al. (2008)). The objective of community detection is to densely connect nodes within the same community and to sparsely connect nodes between different communities. Among indexes to evaluate created clusters (communities), modularity has received a lot of attention. The modularity is a scalar value between −1 and 1, and it measures the density of links inside communities as compared to links between communities. In the case of weighted networks, the modularity score is defined as 1 X ki kj Q= Aij − δ(ci , cj ) (14) 2m i,j 2m ki = X Aij , m = j 1X Aij 2 i,j (15) where Aij represents the weight of the link between nodes i and j, ki is the sum of weights of links attached to node i, and ci is the community to which node i is assigned. The δ-function δ(u, v) is 1 if u = v and 0 otherwise. Exact modularity optimization is computationally hard to solve. Thus, several approximation algorithms to optimize modularity faster have been proposed. Among such algorithms, LM is a well-known method that can optimize modularity in short time. The algorithm of LM consists of two phases, which are repeated iteratively. Assume that a weighted network has N nodes and each node is assigned to different communities. For each node i, the gain of modularity is evaluated by removing node i from its community and by placing it in the community of node j. Here node j is a neighboring node of i. Then, the node i is placed in the community for which the gain of modularity is maximum only if the gain is positive. If there is no positive gain, node i stays in its original community. This process is applied to all nodes and repeated until no further improvement is achieved. In this first phase, the gain of modularity ∆Q, which is derived by moving node i into community C, is calculated as follows. Copyright © 2015 IFAC "P P 2 # +ki,in tot +ki ∆Q = − 2m 2m "P P 2 2 # ki tot in − − − 2m 2m 2m in (16) P P where in is the sum of weights of links inside C, tot is the sum of weights of links attached to nodes in C, and ki,in is the sum of weights of links from node i to nodes in C. In the second phase, a new network is generated. The network consists of nodes that are communities generated in the first phase. Weights between new nodes are defined as the sum of weights of links between nodes in the corresponding two communities. These two phases are iterated while significant improvement of the network modularity is obtained. LM runs in O(n log n) time complexity, where n is the number of nodes in weighted networks. LM is extremely fast because the gain of modularity is easy to compute and the number of communities decrease drastically after a few iterations. As a result, LM is faster than SC, which needs to compute eigenvalue decomposition. By using LM instead of SC, NCLM can execute clustering significantly faster than NCSC. In the following example and case studies, the generalized Louvain method is used (De Meo et al. (2011); Jutla et al. (2011)). 4.2 Numerical example The discrimination performance of NCLM was compared with that of the k-means method (KM) and NCSC through a numerical example. The data set consists of three classes of samples, which have different correlation and intersected at the origin, as shown in Fig. 3. For each class, 100 samples were generated through x = sai + n (i = 1, 2, 3) (17) T where x = [x1 x2 ] is measurements of two variables and T n = [n1 n2 ] is measurement noise. s ∼ N (0, 1) and ni ∼ N (0, 0.01) are variables following the normal distribution N (m, σ) with mean m and standard deviation σ. The T T coefficient vectors are a1 = [ 1 3 ] , a2 = [ 2 2 ] and T a3 = [ 2 1 ] . Before clustering, x1 and x2 are normalized (auto-scaled). In NCSC and NCLM, the parameter γ of the NC method was determined by trial and error and set as γ = 0.999. The clustering results of KM, NCSC, and NCLM are shown in Fig. 3. The conventional distance-based method, KM, did not classify samples correctly, but both NCSC and NCLM successfully classified them except around the origin. Table 1 compares the computational time of each method. It was confirmed that NCLM required less computational time than NCSC. 4.3 NCLM-based variable selection The present work aims to develop a new correlationbased group-wise method for selecting input variables of 126 IFAC ADCHEM 2015 June 7-10, 2015, Whistler, British Columbia, Canada Table 2. Comparison of input variable selection methods: estimation of the concentration of unreacted NaOH in the alkali washing tower True groups ALL R PLS-Beta VIP NCSC-VS NCLM-VS k-means 0.4 0.2 0 x2 -0.2 -0.4 -0.6 0.4 0.2 0 x2 -0.2 -0.4 -0.6 -1 -0.5 x 0 1 0.5 -1 NCSC -0.5 x1 0 0.5 NCLM 0.4 0.2 x2 0 -0.2 -0.4 -0.6 0.4 0.2 0 x2 -0.2 -0.4 -0.6 -1 -0.5 x1 0 0.5 -1 -0.5 x1 0 0.5 Fig. 3. Classification results of numerical example when three classes intersect at the origin: true classes (topleft), k-means (top-right), NCSC (bottom-left), and NCLM (bottom-right) Table 1. Comparison of clustering methods in computational time Method k-means NCSC NCLM Time [s] 0.0056 0.1554 0.0959 statistical models; in particular, the proposed method is expected to be faster than NCSC-VS. This goal would be achieved by using NCLM for variable selection. This method is called NCLM-VS. NCLM-VS clusters input variables into variable classes based on their correlation and evaluates the contribution of each variable class to estimates. Unlike NCSC-VS, NCLMVS optimizes the number of classes automatically and it does not require heavy computational load. 5. APPLICATIONS In this section, the proposed NCLM-VS is validated through two case studies. One is the soft-sensor design for an industrial chemical process, and the other is the calibration modeling from near-infrared (NIR) spectroscopy data. 5.1 Soft-sensor design for chemical process NCLM-VS and conventional methods were compared through their applications to soft-sensor design. The objective was to estimate the concentration of unreacted NaOH in an alkali washing tower. By taking account of the process dynamics, 50 process variables such as pressure, temperature, and flow rate are used as input variables. Copyright © 2015 IFAC RMSE 0.87 0.70 0.91 0.99 0.65 0.64 r 0.71 0.78 0.72 0.63 0.79 0.79 #Var 50 43 14 36 27 34 #LV 4 4 7 4 7 4 Time [s] — 0.74 0.83 1.14 9.51 0.21 Samples were divided into three sets; 150 samples were used for constructing models (Model set), 19 samples were used for tuning parameters (Parameter set), and 40 samples were used for evaluating the estimation accuracy of PLS models (Test set). Parameters were optimized so that root mean square error (RMSE) for the Parameter set was minimized. RMSE is defined as v u N u1 X RMSE = t (yn − ŷn )2 (18) N n=1 where N is the number of samples in the data set, yn is the nth reference value of the output variable, and ŷn is the estimated value of yn . The number of latent variables of the final PLS model was determined by cross validation. Finally, the estimation accuracy was evaluated by using the Test set. The threshold of NC was set at 0.8 in both NCSC-VS and NCLM-VS, which was determined by trial and error. The number of classes J of SC was optimized within the range from 1 to 11; this setting contributes toward reducing the computational time of NCSC-VS. The specifications of the computer used in this work is as follows; OS: Windows 7 Enterprise, CPU: Intel(R) Core(TM) i7-2600K 3.40GHz, Memory: 8GB. In this case study, six variable selection methods were compared: ALL (no selection), R, PLS-Beta, VIP, NCSCVS, and NCLM-VS. Table 2 shows RMSE for the Test set, the correlation coefficient between the measured NaOH concentration and the corresponding estimates (r), the number of selected input variables (#Var), the number of adopted latent variables (#LV), and the computational time required to optimize parameters and select input variables. The conventional methods, R, slightly improved the estimation accuracy compared with ALL. However, the other conventional method such as VIP and PLS-Beta did not improved the estimation accuracy compared with ALL. The correlation-based group-wise methods, i.e, NCSCVS and NCLM-VS, achieved the significantly better performance than the others. In particular, NCLM-VS was the best in the estimation accuracy among all methods. Furthermore, the computational load of NCLM-VS was drastically lighter than that of NCSC-VS as well as the other methods. The optimal number of generated variable classes was 11 in NCSC-VS and 15 in NCLM-VS. On the basis of the contribution ratio, both NCSC-VS and NCLMVS adopted only one variable class. Here it is noted that the number of generated variable classes in NCSC- 127 IFAC ADCHEM 2015 June 7-10, 2015, Whistler, British Columbia, Canada Table 3. Comparison of input variable selection methods: estimation of the freezing point of diesel fuel ALL R PLS-Beta VIP NCSC-VS NCLM-VS RMSE 2.47 2.12 2.48 2.27 2.22 2.15 r 0.80 0.84 0.78 0.81 0.82 0.84 #Var 401 204 5 155 60 195 #LV 6 7 5 6 7 7 Time [s] — 59 59 59 226 3.0 VS was optimized within the range from 1 to 11. The computational time of NCSC-VS strongly depends on the maximum number of generated variable classes. The computational time of R, PLS-Beta, and VIP was very long because they have 50 classes (input variables). These methods add each variable increamentally for model construction. Then, each of these methods finds the optimal number of selected input variables within the range from 1 to 50 while NCSC-VS and NCLM-VS need to find the optimal number of selected variable classes within the range from 1 to 11 and from 1 to 15, respectively. Moreover, NCLM-VS optimizes the number of generated variable classes without any limitation. Hence, the computational efficiency of NCLM-VS is remarkable. 5.2 Calibration modeling for NIR spectroscopy In this case study, calibration models were constructed by using NIR spectral data. The output variable is the freezing point of diesel fuel, and the input variables are absorbances at 401 wavelengths of NIR spectroscopy. Samples were divided into three sets; 133 samples were used for constructing models (Model set), 56 samples were used for tuning parameters (Parameter set), and 56 samples were used for evaluating the estimation accuracy of PLS models (Test set). The procedure and parameter settings were the same as those in the previous case study except that the number of generated variable classes in NCSC-VS was optimized within the range from 1 to 40. The results are summarized in Table 3. The difference of RMSE was not significant, but it is confirmed that NCLMVS achieved better estimation accuracy than NCSCVS and reduced the computational time significantly. The computational time of NCSC-VS is extremely long mainly because the maximum number of generated variable classes was 40 (the number was 11 in the previous case study). The case studies clearly demonstrate the advantage of the proposed NCLM-VS method over the conventional variable selection methods in the computational load and the estimation performance. REFERENCES Blondel, V., Guillaume, J.L., Lambiotte, R., and Lefebvre, E. Fast unfolding of communities in large networks. Journal of Statistical Mechanics: Theory and Experiment, P10008, 2008. De Meo, P., Ferrara, E., Fiumara, G., and Provetti, A. Generalized louvain method for community detection in large networks. International Conference on Intelligent Systems Design and Applications, ISDA, 88–93, 2011. Ding, C., He, X., Zha, H., Gu, M., and Simon, H.. A min-max cut algorithm for graph partitioning and data clustering. In Proceedings of IEEE International Conference on Data Mining (ICDM), 107–114, 2001. Fujiwara, K., Kano, M., and Hasebe, S. Development of correlation-based clustering method and its application to software sensing. Chemometrics and Intelligent Laboratory Systems, 101, 130–138, 2010. Fujiwara, K., Kano, M., and Hasebe, S. Correlationbased spectral clustering for flexible process monitoring. Journal of Process Control, 21, 1438–1448, 2011. Fujiwara, K., Sawada, H., and Kano, M. Input variable selection for PLS modeling using nearest correlation spectral clustering. Chemometrics and Intelligent Laboratory Systems, 118, 109–119, 2012. Jutla, I.S., Jeub, L.G., and Mucha, P.J. A generalized louvain method for community detection implemented in MATLAB. http://netwiki.amath.unc.edu/GenLouvain, 2011. Kano, M. and Fujiwara, K. Virtual sensing technology in process industries: Trends and challenges revealed by recent industrial applications. Journal of Chemical Engineering of Japan, 46, 1–17, 2013. Kano, M. and Ogawa, M. The state of the art in chemical process control in japan: Good practice and questionnaire survey. Journal of Process Control, 20(9), 969–982, 2010. Lynn, S., Ringwood, J., and MacGearailt, N. Global and local virtual metrology models for a plasma etch process. IEEE Transactions on Semiconductor Manufacturing, 25, 94–103, 2012. Mehmood, T., Liland, K. H., Snipen, L., and Sb, S. A review of variable selection methods in partial least squares regression. Chemometrics and Intelligent Laboratory Systems, 118, 62–69, 2012. Ng, A., Jordan, M., and Weiss, Y. On spectral clustering: Analysis and an algorithm. Advances in Neural Information Processing Systems, 2002. Wold, S., Sjöström, M., and Eriksson, L. PLS-regression: A basic tool of chemometrics. Chemometrics and Intelligent Laboratory Systems, 58, 109–130, 2001. 6. CONCLUSION To construct highly accurate statistical models, a new input variable selection method, NCLM-VS, was proposed. NCLM-VS is a correlation-based group-wise method integrating the nearest correlation method and the Louvain method. NCLM-VS is remarkably faster than the conventional methods including NCSC-VS. The advantage of NCLM-VS over conventional methods were demonstrated through two case studies. Copyright © 2015 IFAC 128