Kernel parameter selection for support vector machine classification

Abstract

Parameter selection for kernel functions is important to the robust classification performance of a support vector machine (SVM). This paper introduces a parameter selection method for kernel functions in SVM. The proposed method tries to estimate the class separability by cosine similarity in the kernel space. The optimal parameter is defined as the one that can maximize the between-class separability and minimize the within-class separability. The experiments for several kernel functions are conducted on eight benchmark datasets. The results demonstrate that our method is much faster than grid search with comparable classification accuracy. We also found that the proposed method is an extension of a reported method in reference [2].