Multi-label feature selection based on stable label relevance and label-specific features

Abstract

Multi-label feature selection can efficiently handle large amounts of multi-label data. However, two pressing issues remain in sparse learning for multi-label data. First, many methods explore label relevance through the original label matrix, which may introduce redundant and useless information, and they either consider local label relevance or global label relevance, which are complementary to obtaining potential label information. Second, most existing multi-label feature selection (MLFS) algorithms do not consider the characteristics of the labels themselves, i.e., label-specific features. To solve the above problems, an MLFS method based on stable label relevance and label-specific features are proposed in this study, which combines global and local label relevance to learn the label-specific features. A stable global label relevance strategy is designed for the self-representation model to avoid the effect of noise and outliers, and a manifold regularization term is adopted to explore the local label relevance. Moreover, ℓ1-norm is applied to the feature weight matrix and self-representation matrix to select the label-specific features. Then, a solution for optimizing the objective function is designed. Compared with eight related algorithms on sixteen public datasets. The results show that the proposed method can improve the classification performance of multi-label data.