Robust Multi-Weight Vector Projection Support Vector Machine

Abstract

Recently proposed Multi-weight vector projection support vector machines (MVSVM) is an outstanding algorithm for binary classification. However, it measuring distance in the objective function by squared L2-norm, which is easy to find that the impact of outliers is exaggerated. To alleviate this, we propose an effective algorithm, termed as Robust MVSVM based on the L1-norm distance (L1-MVSVM). The distance in the objective of L1-MVSVM is measured by L1-norm. Besides, we design a powerful iterative algorithm to solve the optimal problem of L1-norm, whose convergence is theoretically ensured. Finally, the effectiveness of L1-MVSVM has been verified through extensive experiments.