C-Score Evaluates Semi-Supervised Learning Robustness in Open-World Settings.

Semi-supervised learning (SSL) methods, particularly those relying on pseudo-labeling, have shown strong performance but typically assume that unlabeled data comes from the same distribution as labeled data. In real-world scenarios, unlabeled datasets often contain out-of-distribution (OOD) samples, which can lead to significant issues. These OOD samples can be incorrectly assigned high-confidence pseudo-labels, corrupting the training process even if standard accuracy metrics appear stable. To address this critical evaluation gap, researchers have developed C-Score, a comprehensive diagnostic framework. C-Score moves beyond simple accuracy by analyzing the internal learning dynamics of SSL models across three distinct dimensions: prediction behavior, feature representation, and optimization alignment. It includes metrics like PLE and CCI for unlabeled prediction, Sem-Drift for semantic deviation, and Grad-Align for the compatibility between labeled and unlabeled optimization. Experiments demonstrate that C-Score effectively uncovers "hidden collapse" in SSL models, where internal degradation occurs despite seemingly stable clean test accuracy. For instance, under OOD contamination, C-Score metrics showed significant changes while accuracy remained largely unaffected, highlighting its ability to provide a more reliable assessment of SSL robustness in open-world environments.