Feature Selection Method for L1-norm Twin Support Vector Regression

Abstract

L1-norm twin support vector regression is a sparse regression algorithm with certain feature selection ability, but its feature selection is inefficient and cannot be applied to nonlinear problems. To solve this problem, a feature selection method for L1-norm twin support vector regression (L1-FTSVR) is proposed to automatically select significant features. Feature selection is implemented in L1-FTSVR by introducing a diagonal matrix whose diagonal element is 0 or 1. The feature selection matrix is used to convert the regression upper and lower bound functions into a multi-objective hybrid programming problem, and then the alternate iterative method is used to solve the multi-objective programming problem. Experimental results on several UCI datasets show that the proposed algorithm not only has good regression performance, but also effectively improves feature selection ability.