Multi-label Robust Feature Selection via Subspace-Sparsity Learning

Multi-label feature selection is crucial for managing feature redundancy and irrelevance in high-dimensional datasets. Existing methods reduce information redundancy through subspace dimensionality reduction but often suffer from instability due to high degrees of freedom and lack flexibility. This is because of the assumption of a shared subspace for features and labels, which leads to reduced performance. To address these problems, we introduce a novel multi-label feature selection approach. Specifically, we propose a dual subspace learning approach to capture both label correlations and feature correlations for feature selection. Therefore, our method can mitigate the adverse effects of noise, redundancy and imperfect features in the quest for discriminative features. Additionally, it reduces the sensitivity of the constructed model to the noise and outliers present in the data. Empirical experiments conducted on real-world datasets illustrate the efficiency and superiority of our proposed approach.