Multi-Label Feature Selection via Non-Additive Fusion of Embedded and Spectral Evidence

Multi-label feature selection (MLFS) is a critical task for handling high-dimensional datasets where labels exhibit complex interdependencies. While embedded methods effectively identify discriminative features by optimizing predictive performance through structured sparsity, they often overlook the underlying topological structure of the label space. Conversely, spectral methods capture label correlations via similarity structures but lack direct optimization for prediction. To bridge this gap, this paper introduces CFIFS (Choquet Fuzzy Integral Feature Selection), a framework that models MLFS as a non-additive, score-level information fusion problem. The proposed pipeline integrates two complementary perspectives: a group-sparse embedded score for predictive relevance and a spectral score---combining Dirichlet smoothness and the Hilbert-Schmidt Independence Criterion (HSIC)---for topological consistency. Rather than relying on traditional additive aggregation, CFIFS employs a two-source Choquet fuzzy integral to fuse these rankings. The fuzzy capacity is learned through an automated data-driven procedure, allowing the framework to adaptively weigh the synergy or redundancy between input scores. Furthermore, the model provides an interpretable interaction index that quantifies the relationship between the predictive and topological evidence. Extensive experiments on 20 benchmark datasets demonstrate that CFIFS significantly outperforms seven state-of-the-art MLFS baselines. Rigorous statistical validation via Friedman and Holm tests confirms the robustness and superior generalization of the proposed fusion approach across diverse application domains.