A new principle for measuring the generalization performance of SVMs

摘要

A new method for estimating the VC dimension in advance is proposed, which can be taken as a principle for measuring the generalization performance of SVMs. Our method adopts the two-order statistic of training samples and maintains consistency with the method for estimating the VC dimension in statistical learning theory. Our method can be applied to the problem of model selection and samples pre-processing. Simulation results show the feasibility and practicable of our method.