A Comparative Study of Kernels for the Multi-class Support Vector Machine

Support Vector Machine (SVM) is a powerful classification technique based on the idea of structural risk minimization. Use of a kernel function enables the curse of dimensionality to be addressed. However, a proper kernel function for a certain problem is dependent on the specific dataset and as such there is no good method on how to choose a kernel function. In this paper, the choice of the kernel function is studied empirically and optimal results are achieved for multiclass SVMs combining several binary classifiers. The performance of the Multi-class SVM is illustrated by extensive experimental results which indicate that with suitable Kernel and parameters better classification accuracy can be achieved as compared to other methods. The experimental results of the four datasets show that Gaussian Kernel is not always the best choice to achieve high generalization of classifier although it is often the default choice.

1. INTRODUCTION

In the past two decades valuable work has been carried out in the area of Text Categorization [10], [11], Optical Character Recognition [13], Intrusion Detection [14], Speech Recognition [18], Handwritten Recognition [20] etc. All such real-world applications are essentially multi-class classification problems. Multi -class classification is intrinsically harder than binary classification problem because the classification has to learn to construct a greater number of separation boundaries or relations. Classification error rate is greater in multi -class problem than that of binary as there can be error in determination of anyone of the decision boundaries or relations.