Robust Clustered Support Vector Machine With Applications to Modeling of Practical Processes

Real datasets are often distributed nonlinearly. Although many least squares support vector machine (LS-SVM) methods have successfully modeled this kind of data using a divide-and-conquer strategy, they are often ineffective when nonlinear data are subject to noise due to a lack of robustness within each sub-model. In this paper, a robust clustered LS-SVM is proposed to model this type of data. First, the clustering method is used to divide the sample data into several sub-datasets. A local robust LS-SVM model is then developed to capture the local dynamics of the corresponding sub-dataset and to be robust to noise. Subsequently, a global regularization is constructed to intelligently coordinate all local models. These new features ensure that the global model is smooth and continuous and has a good generalization and robustness. Through the use of both artificial and real cases, the effectiveness of the proposed robust clustered LS-SVM is demonstrated.