Multi-label feature selection via nonlinear mapping and manifold regularization

Feature selection has become a critical research focus within the field of multi-label learning, where traditional methods largely rely on the assumption of linearity between features and target labels. Nevertheless, in many real-world scenarios, features and labels are often associated with each other through complex nonlinear relationships. As a result, there is a urgent need to incorporate nonlinear mappings to accurately capture the underlying connections between features and labels. To address this, we propose a novel multi-label feature selection method that leverages nonlinear mapping and manifold regularization (NMMFS). Specifically, we first analyze the semantic similarity and correlations among labels, refining semantic labels using matrix decomposition techniques. Next, we construct a nonlinear mapping from the original features to the semantic labels through the sigmoid function. This ensures that the output values of the nonlinear mapping remain within the [0,1] interval, better aligning with the distribution of semantic labels. To maintain data structure consistency during this transformation, we apply manifold learning as a regularization technique. The experimental results show that the proposed algorithm greatly surpasses existing mainstream ones in terms of performance metrics, validating its theoretical feasibility and technical advantages.