L₁-Norm Least Squares Support Vector Regression via the Alternating Direction Method of Multipliers

抄録

In this study, we focus on the feature selection problem in regression, and propose a new version of L₁ support vector regression (L₁-SVR), known as L₁-norm least squares support vector regression (L₁-LSSVR). The alternating direction method of multipliers (ADMM), a method from the augmented Lagrangian family, is used to solve L₁-LSSVR. The sparse solution of L₁-LSSVR can realize feature selection effectively. Furthermore, L₁-LSSVR is decomposed into a sequence of simpler problems by the ADMM algorithm, resulting in faster training speed. The experimental results demonstrate that L₁-LSSVR is not only as effective as L₁-SVR, LSSVR, and SVR in both feature selection and regression, but also much faster than L₁-SVR and SVR.