Model Selection: An Empirical Study on Two Kernel Classifiers

This paper records our activities as a participant of the challenge in performance prediction for WCCI2006. We carry out model selection for two kernel classifiers, the support vector classifier and the Gaussian process classifier, on the five real-world data sets used in the challenge. K-fold cross validation is employed for the support vector classifier, while approximate Bayesian inference is used for the Gaussian process classifier. We give a detailed description of these model selection techniques and report the corresponding experimental results. The empirical study shows both techniques work well on these real-world applications.