CDRA: Class distribution-based re-balancing algorithm for class-imbalanced semi-supervised learning

Abstract

Following the rise of class-imbalanced semi-supervised learning(CISSL), numerous attempts have been made significant improvements. However, many CISSL algorithms have assumed that the class distribution of unlabeled data is the same as or similar to the class distribution of labeled data. Models based on such assumptions have a difficulty in dealing with data that have class distribution mismatch between labeled and unlabeled data, which is common situation in real-world CISSL settings. To address this issue, we suggest 'class distribution based re-balancing algorithm (CDRA)' that estimates unknown class distribution of unlabeled data and leverages this estimated distribution to alleviate class imbalance. CDRA uses predicted class probabilities for unlabeled samples in estimating class distribution of unlabeled data via Monte Carlo approximation. To embody our idea, we combine CDRA with an auxiliary balanced classifier (ABC) which employs a training loss rebalanced in accordance with the class distribution of the labeled data. By estimating class distribution of the unlabeled set, CDRA allows ABC to be trained in a balanced way even under severe class distribution mismatch. Furthermore, to mitigate imbalance in learning representations, we employ the previously proposed idea that mapping minority class samples into denser clusters. CDRA achieves state-of-the-art performance in all experimental settings, particularly excelling in scenarios with class distribution mismatches.