Randomized and Dimension Reduced Kernel Generation for Support Vector Machine

Abstract

Generally, kernel support vector machines have several hyper-parameters which have to be determined by appropriate estimation methods. Usually, such hyper-parameters are fixed to same values for all learning samples. In this paper, we propose a novel kernel function based on usual radial basis function. We choose seeds of kernel function from learning samples at random, and we set randomized hyper-parameters for each kernel seed individually.