Choosing General Gaussian Kernel Parameters for Multiclass Pattern Classification

摘要

Kernel methods have been established as powerful tools in machine learning and data mining. However, the burden of choosing the appropriate kernel function and its parameters is left to the user. This paper presents an effective general Gaussian kernel parameter selection method for multiclass pattern classification. Unlike the standard single scale Gaussian kernel, the general Gaussian kernel adopts some linear transformations of the input space such that not only the scaling but also the rotation is adapted. We propose to optimize the general Gaussian kernel parameters by maximizing a newly presented kernel evaluation criterion named multiclass kernel polarization (MKP), and this problem is solved through a gradient-based optimization technique. MKP makes full use of all the training samples available and can assess the quality of a kernel in the multiclass classification scenario. The experiments with several UCI data sets demonstrate the effectiveness of the proposed method.