Multi-label feature selection via joint label enhancement and pairwise label correlations

Multi-label feature selection(MFS) has gained in importance, and it is today confronted with the current need to process multi-semantic high-dimensional data. Recent studies usually figure out the MFS problems either simply assume that all associated labels are equally important for each instance; or that the labels are independent of each other. In many real-world applications, however, both cases may occur that the significance of each relevant label is generally different and label correlations are ubiquitous. Based on this observation, we propose a new algorithm, called FSEP, to perform MFS by considering label significance and pairwise label correlations. In FSEP, we first construct a label enhancement method that is able to obtain label distribution and further earn the information of label significance. Then, FSEP explores the influence mechanism of label correlations to features by using neighborhood mutual information and incorporates this influence into the process of feature evaluation. After that, a novel multi-label feature selection strategy, namely, Max-Relevance, Max-Contribution, and Min-Redundancy, is proposed, which achieves a favorable trade-off among feature relevance, the contribution of label correlations to features, and feature redundancy, simultaneously. Extensive experiments on both public and real-world datasets show that the proposed method achieves encouraging results compared with state-of-the-art MFS algorithms.