Multi-label feature selection via maximum dynamic correlation change and minimum label redundancy

Information-theoretic measures have been commonly applied to evaluate the relevance and redundancy in multi-label feature selection. However, the current multi-label feature selection methods based on information-theoretic measures neglect the dynamic changes in the relevance of selected features and candidate features. Furthermore, they also do not fully consider the influence of label redundancy on the relevance of candidate features. In this paper, we first propose a new feature relevance term named Dynamic Correlation Change (DCC), which uses two conditional mutual information terms to evaluate the dynamic changes in the relevance of selected features and candidate features. We then introduce a new label redundancy term named Label Redundancy with Interaction Information (LRII), which more accurately quantifies the influence of label redundancy on the relevance of candidate features. On this basis, we design a new multi-label feature selection method, called Maximum Dynamic Correlation Change and Minimum Label Redundancy (MDCCMLR), by combining DCC and LRII. Finally, we conduct extensive experiments in order to verify the performance of our method by comparing it with some state-of-the-art multi-label feature selection methods based on information-theoretic measures in terms of six evaluation metrics. The experimental results show that the MDCCMLR method outperforms the other comparison methods on all six evaluation metrics.