Novel multi-label feature selection via label enhancement and relative maximal discernibility pairs

Multi-label feature selection is an effective solution to the multi-label data dimensionality disaster problem. However, there are few studies on multi-label feature selection considering label enhancement methods. Meanwhile, most existing label enhancement methods neglect the relative importance of labels, which can degrade the classification performance of the model. To address this issue, we propose a novel multi-label feature selection algorithm based on label enhancement and relative maximal discernibility pairs. Firstly, we propose the label importance weight based on relative discernibility pairs and design the concept of soft relevance between objects and labels via fuzzy rough sets. Secondly, we propose a novel label enhancement algorithm by combining the soft relevance and the label importance weight. Thirdly, we define a relative maximal discernibility pair model for evaluating features in label distribution information systems. Additionally, based on the relative maximal discriminative pair model and label enhancement, we present a multi-label feature selection algorithm which can continuously reduce the universe of object pairs in the selection process. Finally, to validate the effectiveness and stability of our algorithm, we conduct extensive comparison experiments with 7 representative multi-label feature selection algorithms on 13 datasets. Experimental results show that our algorithm performs better than the compared 7 algorithms in 5 evaluation metrics.