Multi-label feature selection based on multi-granulation separability

Abstract

Multi-label feature selection plays a critical role in reducing data dimensionality while preserving discriminative information. As an effective feature evaluation criterion, the class separability criterion has been extensively employed in single-label data. However, the existing class separability-based methods cannot directly process multi-label data. To address this issue, we introduce a new concept of granulation separability, which seamlessly integrates the principles of data granulation and discriminative separability. Based on this innovation, we design a novel multi-label feature selection method MSMFS. Specifically, we first design a multi-level granulation strategy to divide the data into multiple granules at different granularity levels, which captures multi-level discriminative patterns. Second, we define a novel feature evaluation criterion called multi-granulation separability score, which reflects the ability of features to separate samples from multi-level perspective. Third, we construct a greedy mechanism to iteratively select a feature subset, which can effectively reduce redundant information in the selected feature subset. Extensive experiments demonstrate that MSMFS outperforms nine state-of-the-art methods across fourteen benchmark datasets, achieving the highest Macro-F1 score on 79% of datasets while completing computations within 10 s on 93% of datasets.