Robust multi-label feature selection with dual-graph regularization

Abstract

Multi-label learning is facing great challenges due to high-dimensional feature space, complex label correlations and noises in multi-label data. Feature selection techniques have attracted considerable attention to address the problems. In this paper, we design our method based on dual-graph regularization, i.e., feature graph regularization and label graph regularization. The feature graph regularization is used to preserve the geometric structure of features, while label graph regularization intends to explore the correlations of labels. Furthermore, the l2,1-norm is imposed on the loss function to enhance the robust of feature selection methods. As a result, a new feature selection method termed Robust Multi-label Feature Selection based on Dual-graph (DRMFS) is proposed. Particularly, only one unknown variable, feature weight matrix, is incorporated in our proposed method, which can reach global optimum. Additionally, we impose both l2,1-norm and non-negative constraints onto the feature weight matrix to enhance the property of row-sparse. Finally, we design an optimization scheme to solve the proposed method, and offer the convergence proof of the optimization scheme. Extensive experimental results demonstrate the superiority of the proposed method in comparison to the-state-of-art multi-label feature selection methods. Finally, some insightful discussions with respect to the convergence analysis, complexity analysis and parameter sensitivity analysis are presented.