A safe screening rule for accelerating weighted twin support vector machine

Abstract

Weighted twin support vector machine with local information (WLTSVM) is a novel algorithm for binary classification problems. It can exploit as much underlying correlation information as possible. Unfortunately, it remains challenging to apply WLTSVM into large-scale problems directly. Motivated by the sparse solution of WLTSVM, in this paper, a safe screening rule is proposed for WLTSVM, termed as SSR-WLTSVM. The SSR-WLTSVM can delete a majority of training samples before actually solving it to reduce the scale of WLTSVM. Therefore, computation time can be reduced greatly. More importantly, our screening rule is safe in the sense that the reduced problem can derive an identical optimal solution as the original one. Besides, a different neighbor k having a different effect on the performance of SSR-WLTSVM is further elaborated, that is, a bigger k will achieve a greater speedup. Sequential versions of SSR-WLTSVM are further introduced to substantially accelerate the parameter tuning process. And a fast algorithm clipDCD is introduced in this paper to handle large-scale datasets. In addition, Friedman test and paired-sample t test are used to verify the effectiveness of SSR-WLTSVM. Experimental results on 30 benchmark datasets confirm the efficiency of our proposed algorithm.