Abstract In multi-label applications like text classification, feature selection is crucial for reducing complexity and enhancing interpretability. Existing methods often inadequately exploit label space discriminative information and rely excessively on pairwise feature correlations for redundancy reduction. To address these limitations, this paper proposes an efficient multi-label feature selection method based on label Orthogonal embedding and feature Self-representation (OSMFS). In the feature space, a feature self-representation matrix, learned by feature-wise local linear embedding, is used to construct a feature manifold regularization term for suppressing feature redundancy. In the label space, the original logical label matrix is orthogonally embedded into a latent label matrix and a dynamic label manifold regularizer is introduced to align this latent space with the feature representation space. A unified sparse regression model is constructed by incorporating the two manifold regularizers, enabling it to simultaneously explore label correlations, suppress feature redundancy, and enforce cross-space structural consistency. Finally, an optimization scheme is proposed to solve the model, along with a rigorous convergence analysis. Extensive experiments on 18 multi-label datasets, using three different base classifiers (ML-KNN, BR-SVM, and CC-SVM), consistently demonstrate that OSMFS significantly outperforms seven state-of-the-art methods across all four evaluation metrics, achieving the best average ranks on Hamming Loss, Average Precision, Macro-F1, and Micro-F1. The code is available at https://github.com/xxwang714/OSMFS .
Wang et al. (Sat,) studied this question.