Support Vector Selection for Regression Machines

Lee Wan-Jui
Yang Chih-Cheng
Lee Shie-Jue
Abstract
In this paper, we propose a method to select support vectors to improve the performance of support vector regression machines. First, the orthogonal least-squares method is adopted to evaluate the support vectors based on their error reduction ratios. By selecting the representative support vectors, we can obtain a simpler model which helps avoid the over-fitting problem. Second, the simplified model is further refined by applying the gradient descent method to tune the parameters of the kernel functions. Learning rules for minimizing the regularized risk functional are derived. Experimental results have shown that our approach can improve effectively the generalization capability of support vector regressors.
Keywords
Orthogonal least-squares
over-fitting
gradient descent
learning rules
error reduction ratio
mean square error
ISSN
1883-3977
NCID
BB00577064