Disentangled Pseudo-Labeling and Classification for Class-Imbalanced Semi-Supervised Learning

Taemin Park, Heeyoung Kim

Keywords: Class-Imbalanced Semi-Supervised Learning, Pseudo-labeling

Abstract: Although significant improvements have been made in addressing class-imbalanced semi-supervised learning (CISSL), many algorithms still suffer from confirmation bias. Inaccurate pseudo-labels hinder the learning of the classifier, which in turn leads to further inaccurate pseudo-labels—creating a self-reinforcing loop that amplifies bias, particularly toward majority classes. This bias arises because the classifier that generates pseudo-labels is simultaneously trained on the unlabeled data it labels. To address this issue, we propose a novel CISSL algorithm, Disentangled Pseudo-Labeling and Classification (DPC). DPC introduces an auxiliary classifier, dedicated solely to generating pseudo-labels, called a pseudo-labeler, which is attached to the representation layer of the backbone semi-supervised learning algorithm. To prevent confirmation bias, the pseudo-labeler is trained exclusively on labeled data, ensuring that pseudo-label generation remains unaffected by noisy unlabeled samples. Furthermore, to mitigate imbalanced feature representations—which are often biased toward majority classes and exacerbate confirmation bias—DPC propagates the classifier’s training loss to the shared representation layer to encourage balanced feature learning. Benefiting from high-quality pseudo-labels and balanced feature representations, DPC achieves state-of-the-art classification performance on CISSL benchmark datasets.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 9190