Artificial intelligence and machine learning models often struggle when trained on real-world datasets, which frequently exhibit severe class imbalance. A research team has proposed a groundbreaking framework called DALE that effectively addresses the "confirmation bias" in class-imbalanced semi-supervised learning, significantly improving model performance on skewed data. The findings were published on [DATE] in Frontiers of Computer Science.
Existing semi-supervised learning methods usually rely on the classifier itself to generate pseudo-labels for unlabeled data. However, in class-imbalanced scenarios, this creates an inherent objective conflict. The classifier tries to optimize performance uniformly across all classes, while the pseudo-labeling process emphasizes the accurate annotation of unlabeled data that naturally follows a skewed distribution. This misalignment causes a confirmation bias, where errors propagate and reinforce the imbalance in the pseudo-labeled pool, leading to suboptimal performance.
To solve these problems, a research team from Southeast University published their new research in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
The team proposed a novel framework named DALE (DebiAsed class-imbalanced Label Enhancement learning). Unlike conventional approaches, DALE explicitly disentangles pseudo-label generation from the classification task, recognizing that the two tasks pursue inherently different objectives. The key motivation is to design a dedicated pseudo-label generator that can align the class distributions between the labeled and unlabeled data before generating pseudo-labels.
In this framework, the pseudo-label generator and the predictive model are trained alternately. The pseudo-label generator is trained on an instance-level debiased label enhancement objective to capture instance-level noise patterns. The predictive model then leverages these high-quality, enhanced pseudo-labels along with class-level debiasing to maintain balanced classification. By employing this decoupled design and using corrected loss, the framework guarantees improved performance.
The detection method was extensively verified on imbalanced benchmark datasets, including CIFAR-10-LT, CIFAR-100-LT, STL-10-LT, and Small-ImageNet127. The experimental data shows that, compared with existing research results, the proposed DALE framework successfully mitigates class-level bias and achieves state-of-the-art performance across most settings.
Future work can focus on relaxing the assumption that the imbalance ratio between labeled and unlabeled data is strictly consistent, broadening the settings to better reflect even more complex real-world scenarios.

DOI
10.1007/s11704-026-51784-8