A Comparative Study and Choice of an Appropriate Kernel for Support Vector Machines

Abstract

Support Vector machine (SVM) has become an optimistic method for data mining and machine learning. The exploit of SVM gave rise to the development of a new class of theoretically refined learning machines, which uses a central concept of kernels and the associated reproducing kernel Hilbert space. The performance of SVM largely depends on the kernel. However, there is no premise about how to choose a good kernel function for a particular domain. This paper focuses in this issue i.e. the choice of the Kernel Function is studied empirically and optimal results are achieved for binary class SVMs. The performance of the Binary class SVM is illustrated by extensive experimental results. The experimental results of the datasets show that RBF Kernel or any other kernels is not always the best choice to achieve high generalization of classifier although it is often the default choice. © Springer-Verlag Berlin Heidelberg 2010.