Feature relevance term variation for multi-label feature selection

Multi-label feature selection is a critical dimension reduction technique in multi-label learning. In conventional multi-label feature selection methods based on information theory, feature relevance is evaluated by mutual information between candidate features and each label. However, previous methods ignore two issues: the influence of the already-selected features on the feature relevance and the influence of the correlations among labels on the feature relevance. To address these two issues, we design a new feature relevance term named Double Conditional Relevance (DCR) that employs two conditional mutual information terms to take the already-selected features and the correlations among labels into account. Finally, a novel multi-label feature selection method combining the new feature relevance term with feature redundancy term is proposed, the proposed method is named Double Conditional Relevance-Multi-label Feature Selection (DCR-MFS). Additionally, an extended method DCR-MFSarg is designed to avoid the shortcoming of the inconsistency of magnitude in DCR-MFS method. The experimental results together with theoretical analysis demonstrate the superiority of the proposed methods.