Title: Implicit Regularization for Multi-label Feature Selection

Abstract: This study investigates feature selection within the framework of multi-label learning, introducing a novel estimator that leverages both implicit regularization and label embedding. In contrast to conventional sparse feature selection techniques that rely on penalized estimators with explicit regularization components—such as the $l_{2,1}$-norm, MCP, or SCAD—we offer a streamlined alternative utilizing Hadamard product parameterization. To steer the feature selection mechanism, we employ a latent semantic approach for multi-label information, which serves as the label embedding. Benchmarked against established datasets, experimental findings indicate that our proposed estimator incurs significantly reduced extra bias and has the potential to facilitate benign overfitting.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC