Robust Low-Rank Subspace Learning for Multi-Label Feature Selection with Global-Local Correlation Modeling

Multi-label classification faces significant challenges from high-dimensional features and complex label dependencies. Traditional feature selection methods often fail to capture these dependencies effectively or suffer from high computational costs. This paper proposes a novel Robust Low-Rank Subspace Learning (RLRSL) framework for multi-label feature selection. Our method integrates global label correlations and local feature structures within a unified objective function, utilizing Schatten-p norm for low-rank subspace learning, l_(2,1),-norm for joint feature sparsity, and manifold regularization for local geometry preservation. We develop an efficient optimization algorithm to solve the resulting non-convex problem. Comprehensive experiments on seven benchmark datasets demonstrate that RLRSL consistently outperforms state-of-the-art methods across multiple evaluation metrics, including ranking loss, multi-label accuracy, and F1-score, using both ML-*k* NN and SVM classifiers. The results confirm the robustness, efficiency, and superior generalization capability of our proposed approach.