Selection of Kernels for Regression Models

Abstract

This thesis presents a new kernel selection strategy for regression models. For the approximation of continuous target functions, regression models with kernel functions can be trained for the given samples generated from the target functions. These regression models are usually trained by the well known training method such as the support vector regression (SVR) or relevance vector regression (RVR). In this training, the critical factor for the performance of regression is the shape of kernel functions, for example, the centers and width of the kernels. One of the methods in the classification problems, in which the outputs are class labels rather than the values, is using the measure of kernel alignment. In this thesis, this concept of kernel alignment, is applied to the regression problems using the correlation measure and the contribution of each kernel to the regression models is analyzed.