Robust twin support vector regression via second-order cone programming

Abstract

Twin Support Vector Regression is an effective machine learning strategy, which splits the predictive task into two small problems, gaining in both efficiency and predictive performance. In this paper, a novel extension for twin Support Vector Regression is presented. The proposal is based on robust optimization, conferring robustness to the predictive task by dealing effectively with uncertainty. The method is first developed as a linear one, and then, subsequently extended to a kernel-based formulation. Our approach accomplishes the best performance on benchmark datasets compared to alternative methods, such as linear regression, support vector regression, and twin support vector regression. This gain in performance demonstrates the virtues of robust optimization on reducing the risk of overfitting, and generalizing the training patterns well with reduced complexity.