GLRE: Low-Rank Regularized Label Enhancement with Manifold Constraints

Abstract Label Distribution Learning (LDL) effectively handles label ambiguity but is often constrained by the difficulty of acquiring ground-truth distributions. Label Enhancement (LE) addresses this by recovering latent distributions from logical labels; however, existing graph-based methods typically overlook global semantic correlations among labels and remain sensitive to annotation noise . To bridge this gap, this paper proposes Global-Local Regularized Enhancement (GLRE), a unified optimization framework that integrates graph Laplacian regularization with nuclear norm minimization. This dual-constraint mechanism allows GLRE to simultaneously capture the local geometry of the feature manifold and the global low-rank structure of the label space . Furthermore, to tackle the scalability challenge, we develop an efficient ADMM-based solver accelerated by the Conjugate Gradient (CG) method . Extensive experiments on four benchmark datasets demonstrate that GLRE yields competitive performance, achieving KL divergence reductions of approximately 30%-60%, particularly in small-sample and high-noise scenarios compared to existing representative baselines.