Scaling the kernel function based on the separating boundary in input space: A data-dependent way for improving the performance of kernel methods

摘要

The performance of a kernel method often depends mainly on the appropriate choice of a kernel function. In this study, we present a data-dependent method for scaling the kernel function so as to optimize the classification performance of kernel methods. Instead of finding the support vectors in feature space, we first find the region around the separating boundary in input space, and subsequently scale the kernel function correspondingly. It is worth noting that the proposed method does not require a training step to enable a specified classification algorithm to find the boundary and can be applied to various classification methods. Experimental results using both artificial and real-world data are provided to demonstrate the robustness and validity of the proposed method.