Identifying label noise via the divergence between label-induced and feature distributions

Identifying label noise in class-overlapping regions remains a significant challenge in the field of label noise learning. Existing methods primarily rely on loss-based or confidence-based detection mechanisms, but they inherently struggle to distinguish between annotation errors and class ambiguities near decision boundaries. This limitation frequently results in misclassification of borderline instances, thereby degrading noise detection performance. This study proposes a Stein-score-based noise identification and correction (SS-NIC) method, which leverages the divergence between label-induced and feature distributions to identify label noise. Specifically, the Stein score function, estimated by score-based diffusion model, guides each instance toward distinct regional centroids within both label-induced and feature distributions, respectively. This process is iterated until convergence, generating dual-movement trajectories from each instance to its corresponding centroids in the two distribution spaces. By analyzing the directional and magnitude divergences between the two movement trajectories, the method can identify noisy instances in class-overlapping regions that are typically difficult to detect. Experiments on benchmarks with symmetric, asymmetric, and real-world label noise demonstrate that SS-NIC consistently outperforms baseline approaches.