Learning General Gaussian Kernels by Optimizing Kernel Polarization

The problem of model selection for Support vector machines (SVM) with general Gaussian kernels is considered. Unlike the conventional standard singlescale Gaussian kernels, where all the basis functions have acommon kernel width, the general Gaussian kernels adoptsome linear transformations of input space such that notonly the scaling but also the rotation is adapted. We proposed a gradient-based method for learning the optimalgeneral Gaussian kernels by optimizing kernel polarization.This method can find a more powerful kernel for a givenclassification problem without designing any classifier. Experiments on both synthetic and real data sets demonstratethat tuning of the scaling and rotation of Gaussian kernelsusing our method can yield better generalization performance of support vector machines.