IVS: interpretative variable selection via perfect bipartite matching

Abstract

Feature selection is a fundamental process in machine learning to identify the most relevant subset of features for a given problem. Among the various feature selection approaches, filter methods stand out for their simplicity and efficiency. However, these methods lack interpretability regarding the relationships between the selected and unselected features. To address this challenge, we propose a novel pairwise feature selection method based on Perfect Bipartite Matching, which establishes optimized linear relationships between features, thus facilitating the interpretation of feature connections. We also demonstrate how to incorporate domain knowledge, allowing users to exclude/include desirable patterns (e.g., pre-select specific features). Empirical evaluations using 17 datasets demonstrate the effectiveness of our approach compared to baseline methods. Furthermore, we present a case study on Chagas disease, showcasing detailed interpretation results and the significance of selected features in sudden cardiac death prevention.