Non-sparse label specific features selection for multi-label classification

In multi-label learning, extracting specific features for each label is a new strategy to construct discriminative classification models in recent years. Existing approaches usually consider that each label is only associated with a small subset of the original features. But in some applications this sparsity assumption does not hold, where non-sparse label specific features are more discriminative than sparse label specific features. In this paper, a novel feature selection-based approach is proposed to extract non-sparse label specific features. Firstly, we translate the logic labels to the numerical ones to convey more semantic information and embed the label correlations. Secondly, a linear regression is modeled to describe the discrimination of label specific features based on the numerical labels. To our best knowledge, it is one of the first attempts to utilize the numerical labels for extracting label specific features. Comprehensive experiments on several multi-label data sets clearly manifest that the superiority of our proposed algorithm. (C) 2019 Elsevier B.V. All rights reserved.