Abstract:The recently proposed class-imbalanced semi-supervised learning (CISSL) algorithms achieved impressive performance by effectively leveraging unlabeled data. However, these algorithms often rely on a pre-defined fixed confidence threshold to filter unlabeled data during training, which overlooks the varying learning dynamics across different classes in class-imbalanced scenarios. Consequently, valuable data could be discarded, leading to degraded performance on minority classes. To tackle this issue, we introduce a novel method called Class-Adaptive Threshold (CAT), which dynamically defines and adjusts the confidence threshold based on the learning status of each class. The core idea of CAT is to iteratively update the thresholds for different classes at each time step, enabling us to fully exploit valuable information that would otherwise be ignored using fixed threshold algorithms. Importantly, CAT does not introduce any additional inference processes. In our experiments, the proposed algorithm achieves state-of-the-art performance on various class-imbalanced datasets. Furthermore, we show that CAT can be seamlessly integrated into the renowned CISSL algorithm, resulting in a remarkable boost in their performance.

Class-Adaptive Threshold for Class Imbalanced Semi-Supervised Learning