Pointwise fuzzy mutual information based multi-label feature selection via feature low-rank regularization

Feature selection is an effective solution to the dimensionality explosion of multi-label data. To assess the classification capability of features, many techniques have been developed from the view of granular computing. These methods can address the uncertainty in data but confront two primary challenges. Firstly, they rely on the predefined evaluation function and heuristic search strategies often converge to local optima. Secondly, redundant features remain difficult to be identified and eliminated. To tackle these challenges, we propose an embedded multi-label feature selection model by integrating pointwise fuzzy mutual information and feature low-rank regularization (PMILR). Redundant features tend to be highly correlated with other features, but contribute minimally to the labels. Highly correlated features exist in the same subspace, making the feature space low-rank. In this study, the low-rank structure of features is revealed to recognize potentially redundant features. Simultaneously, the pointwise fuzzy mutual information is formulated to capture the feature-label correlation. With the guidance of feature representation coefficients and feature-label correlation, a regularizer is properly designed to eliminate the effect of redundant features to labels. Theoretical analysis and experimental results validate the superiority of the developed method.