Robust Feature Selection with Feature Correlation via Sparse Multi-Label Learning

Abstract

The multi-label feature selection that is regarded as a special case of multi-task learning has received much attention in recent years. In this paper, we propose a novel robust and pragmatic multi-label feature selection method, in which the joint l2,1-norm minimizations of loss function and regularization are emphasized. Specifically, the loss function based on the l2,1-norm is robust to outliers, and the l2,1-norm regularization selects features across all samples with joint sparsity. Besides, the feature information inherent in the data is used to construct the correlation matrix, which explores the correlation between features so as to remove the redundant features. An efficient algorithm based on the augmented Lagrangian multiplier method is proposed to solve the objective function. The extensive experiments compared with several state-of-the-art methods are performed on the multi-label datasets to show the effectiveness of the proposed method.