Multi-label feature selection via spectral clustering-based label enhancement and manifold distribution consistency

Multi-label feature selection can effectively improve the performance and efficiency of subsequent learning tasks by selecting important features within multi-label data. However, for handling multiple labels, many approaches group them to gather insights for label fusion, but ignore the different importance of these label groups and treat them equally, which seems unfair to individual label groups and fails to consider their distinct significances. Moreover, for handling the relationship between features and labels, many multi-label feature selection methods efficiently achieve linear fitting of features and labels through manifold learning, but ignore fitting spatial distribution between feature space and label space. Motivated by these, this paper integrates label distribution learning and spectral clustering to evaluate the unique significance of each label group and construct an improved label space, which is then aligned with the feature space through manifold distribution consistency for multi-label feature selection. First, we propose a hypothetical model indicating the existence of a relationship among labels, wherein this relationship involves clustering subordinate labels around a central core label. On this basis, we employ spectral clustering to generate distinct label clusters by integrating density peaks, thereafter combining this with label distribution learning to assess the significance of each cluster. Then, we design a manifold distribution consistency evaluation, i.e., quantifying the structural disparity between feature space and the enhanced label space achieved through spectral clustering-based label enhancement strategy, so as to obtain a low-dimensional feature space and the optimal feature subset. Finally, experimental results showcase the superiority of our proposed multi-label feature selection algorithm when compared with five other algorithms, across several datasets from diverse domains.