Global-Guided Label-Aware Adaptive Multilabel Feature Selection via Fuzzy Mutual Information

In recent years, embedded-based multi-label feature selection (MFS) has received increasing attention due to its ability to perform both feature selection and model training simultaneously. However, most existing methods tend to overlook the varying importance of different labels and typically rely on a weight matrix to implicitly conduct feature selection, which often leads to unstable results and feature redundancy. Moreover, these methods usually fail to fully exploit the global feature importance information inherent in the original data, thereby limiting further improvements in model performance. To address these issues, this paper proposes a novel global-guided label-aware adaptive MFS method based on fuzzy mutual information, called GLA-FMI. First, a convergent recursive propagation algorithm based on label correlations is designed to stably estimate label importance. Then, an adaptive feature activation strategy is designed, which explicitly selects features using binary activation vectors. Additionally, three complementary regularization terms are constructed to ensure that the model identifies an informative, compact, and interpretable subset of features. Specifically, the redundancy-aware regularization term prevents the simultaneous selection of highly redundant features, the feature sparsity regularization term enforces consistency between the binary activation vector and the weight matrix, and the global feature importance-based regularization term guides the model to prioritize more discriminative features. In comparative experiments, GLA-FMI ranks first in 77.33% of the 75 cases (15 datasets × 5 evaluation metrics), significantly outperforming state-of-the-art MFS methods. The code is available athttps://github.com/xyy0717/GLA-FMI.