Feature subset selection for support vector machine through separability assessment in Kernel-defined feature space

In biomedical applications, high-dimensional feature vectors impose a high computational cost, as well as the risk of "overfitting" and the decrease in gene visualization. Feature selection addresses the dimensionality reduction problem by selecting a subset of available features that is most essential for classification. In this paper, we present a novel feature selection method for support vector machine (SVM), utilizing a new class separability defined in kernel-defined feature space. The key idea of our method is that the feature whose removal downgrades the class separability the most in a kernel-defined feature space is relevant to the classification. We also present a kernel parameter selection method based on maximizing the variance of pattern similarity in feature space. Experiments on linear and nonlinear synthetic problems and real world data sets from UCI repository have been carried out to demonstrate the effectiveness of this method. In comparison with conventional filter methods and SVM-RFE, our method eliminates feature redundancy automatically and yields better classification performance.