An Optimization-Based Framework for Feature Selection and Parameters Determination of SVMs

— In this paper, feature selection and parameters\uddetermination in SVM are cast as an energy minimization\udprocedure. The problem of feature selection and parameters\uddetermination is a very difficult problem where the number of\udfeature is very large and where the features are highly correlated.\udWe define the problem of feature selection and parameters\uddetermination in SVM as a combinatorial problem and we use a\udstochastic method that, theoretically, guarantees to reach the\udglobal optimum. Several public datasets are employed to\udevaluate the performance of our approach. Also, we propose to\uduse the DNA Microarray Datasets which are characterized by\udthe large number of features. To validate our approach, we\udapply it to image classification. The feature descriptors of the\udimages were extracted by using the Pyramid Histogram of\udOriented Gradients. The proposed approach was compared with\udtwenty feature selection methods. Experimental results indicate\udthat the classification accuracy rates of the proposed approach\udexceed those of other approaches