A Heuristic Kernel Combination Approach Based on Kernel Fisher Criterion

Empirical success of kernel-based learning methods is very much dependent on the kernel used. Instead of using a single fixed kernel, Multiple Kernel Learning (MKL) algorithms learn a combination of different kernels in order to better match the underlying problem. In this paper, we propose an effective kernel combination approach that is unified for both binary and multiclass classification problems. The key property of the proposed approach is that it adopts the Kernel Fisher Criterion (KFC) as evaluation criterion to measure the goodness of the base kernel. More specifically, aiming at determining weights for convex combination of multiple kernels, we develop a heuristic rule based on KFC to directly assign a weight to each base kernel. The proposed approach is demonstrated with some UCI machine learning benchmark examples. Copyright © 2013 Binary Information Press.