A class possibility based kernel to increase classification accuracy for small data sets using support vector machines

摘要

Appropriate choice of kernels is the most important task when using kernel-based learning methods such as support vector machines. The current widely used kernels (such as polynomial kernel, Gaussian kernel, two-layer perceptron kernel, and so on) are all functional kernels for general purposes. Currently, there is no kernel proposed in a data-driven way. This paper proposes a new kernel generating method dependent on classifying related properties of the data structure itself. The new kernel concentrates on the similarity of paired data in classes, where the calculation of similarity is based on fuzzy theories. The experimental results with four medical data sets show that the proposed kernel has superior classification performance than polynomial and Gaussian kernels.