University of Houston

High classification accuracy results in kernel methods are the reason of their increasing popularity in the machine learning community. Since the introduction of Support Vector Machines several works have been proposed to learn the optimal kernel. Recently the most successful technique for kernel methods is Multiple Kernel Learning, where several kernels are combined into a general compound kernel. To achieve this, each kernel is assigned a weight that is fixed by an optimization technique such as Convex Optimization, Lasso and Group Lasso.

However despite the good results they produce, and their increasing popularity in the community, zero to no effort has been put toward understanding the behavior of the kernel itself. In this work we are presenting an empirical study that aims to extract abstract kernel properties that allow us to understand the kernel transformations and therefore the performance enhancement that different kernels produce. By applying this knowledge the kernel selection process will be an informed one, rather than a blind search in infinity of solutions, like the optimization case.