Multi-label feature selection with local discriminant model and label correlations

Abstract

In multi-label learning, feature selection is an essential preprocessing module, which can be exploited a more compact and precise representation of instances. Most of existing multi-label feature selection methods are either converted into multiple single-label feature selection methods or directly utilize half-baked label information, thus it is difficult for them to obtain a discriminative feature subset across multiple labels. To tackle this problem, we propose multi-label feature selection with local discriminant model and label correlations. First, for each instance, a local clique comprising this instance and its neighboring instances is constructed, and a local discriminant model for each local clique is integrated globally to evaluate the clustering performance of all instances. Second, in terms of clustering results, we explore high-order label correlations to reduce the impact of half-baked label information. Finally, we combine l2,1-norm regularization to design the objective function to achieve multi-label feature selection. Comprehensive experiments are conducted on twelve real-world multi-label data sets, and results demonstrate the effectiveness of the proposed method in comparison with several representative methods.