FGBC: Flexible graph-based balanced classifier for class-imbalanced semi-supervised learning

•We propose a graph-based auxiliary balanced classifier head that is attached to the existing SSL to construct an end-2-end learning framework for efficient label propagation in CISSL.•We propose a simple yet effective flexible threshold adjustment strategy of leveraging different learning status of different classes for proper pseudo-label sieving.•We propose a class-aware feature MixUp (CFM) augmentation algorithm to adaptively augment training features for mitigating the over-fitting problem of tail classes.•Our FGBC framework shows favorable performance on the prevalent CISSL image classification datasets CIFAR10/100-LT, SVHN-LT, and Small ImageNet-127 with various levels of imbalance ratios and labeled ratios. The code will be available. Semi-supervised learning (SSL) has witnessed resounding success in many standard class-balanced benchmark datasets. However, real-world data often exhibit class-imbalanced distributions, which poses significant challenges for existing SSL algorithms. In general, fully supervised models trained on a class-imbalanced dataset are biased toward the majority classes, and this issue becomes more severe for class-imbalanced semi-supervised learning (CISSL) conditions. To address this issue, we put forward a novel CISSL framework dubbed FGBC by introducing a flexible graph-based balanced classifier with three innovations. Specifically, because the propagation of label information becomes difficult for tail classes, we propose a graph-based classifier head attached to the representation layer of the existing SSL framework for efficient pseudo-label propagation. Then, by considering that the learning status of different classes in CISSL may vary, we introduce a flexible threshold adjustment in pseudo-labeling to further select balanced samples to participate in training. Furthermore, to alleviate the risk of overfitting tail classes, we devised a class-aware feature MixUp (CFM) augmentation algorithm, which can further enhance the features of each class by considering their class sizes. Experimental results demonstrate that FGBC achieves state-of-the-art performance on datasets from CIFAR-10/100, SVHN and Small ImageNet-127 under various levels of CISSL conditions.