WSEAS Transactions on Mathematics
Print ISSN: 1109-2769, E-ISSN: 2224-2880
Volume 12, 2013
Learning Optimal Kernel for Pattern Classification
Authors: ,
Abstract: Kernel methods provide an efficient mechanism to derive nonlinear algorithms. Using a kernel function, original data can be implicitly mapped to a very high or even infinite dimensional feature space where the data is approximately linearly separable. For it, a main challenge is to select an appropriate kernel. In this paper, we optimize combinative weight coefficients and combination kernel is constructed by two methods. one method is learning optimal kernel for kernel fisher discriminant analysis (KFDA) for finding optimally combinative weight coefficients. In this method, we treat optimizing combinative weight coefficients as optimization problem over the convex set of finitely many basic kernels. Besides, in order to solve the optimization problem, we use a new iterative method. Another method is a feature space based class separability measure which is introduced in order to further show the efficacy of combination kernel .With this measure, the weight coefficients of combination kernel were optimized. Experiments on five real-words data sets are performed to test and evaluate the efficacy of combination kernel on classification accuracy. The results show that the efficacy of combination kernel is very significant.
Search Articles
Keywords: Fisher discriminant analysis, kernel function, support vector machines, combination kernel, kernel optimization, iterative method