Model selection for Support Vector Machines based on kernel density estimation

This paper, aiming at overcoming the obstacles in selection of optimal kernel and its parameters, proposes a model selection approach of kernel parameters for Support Vector Machine based on kernel density estimation. By investigating kernel density estimation theory, the kernel density function based on inner product relationship of data distribution in high dimensional feature space is constructed, at the same time an evaluation function on performance of kernel mapping is established. Experimental results on both synthetic dataset and practical dataset show that proposed method is able to effectively avoid such limitations as high computational cost and process complexity in traditional model selection, and capable of optimizing kernel parameters as well as keeping better classification accuracy. So, our approach is feasible and effective.