PGR: Pseudo Graph Regularization for Semi-Supervised Classification

• Published in: IEEE transactions on artificial intelligence pp. 1 - 14
• Main Authors: Hu, Cong, Song, Jiangtao, Wu, Xiao-Jun
• Format: Journal Article
• Language: English
• Published: IEEE 2025
• Subjects: Adaptation models; Artificial intelligence; Computational modeling; Contrastive learning; Data augmentation; Data models; Feature extraction; Graph regularization; Perturbation methods; Predictive models; Semi-supervised learning; Training
• ISSN: 2691-4581; 2691-4581
• DOI: 10.1109/TAI.2025.3585095

Semi-supervised learning (SSL) is gaining attention for its intrinsic ability to extract valuable information from labeled and unlabeled data with improved performance. Recently, consistency regularization methods have gained interest due to their efficient learning procedures. However, they are confined to pseudo-label or feature representation-level perturbations, negating the benefit of having both forms in a single framework. This leads to the model remaining robust to either the pseudo-label or the feature representation. To this end, we propose Pseudo Graph Regularization (PGR) for Semi-Supervised Classification, which leverages graph-based contrastive learning to unify pseudo-labels and feature embeddings in a single semi-supervised framework. The model imposes graph regularization on both pseudo-labels and feature embeddings of unlabeled data to retain the intrinsic geometric structure. Feature embeddings into the model impose constraints on the class probability, forcing the class probability distributions of unlabeled data subject to different perturbations to be consistent. The pseudo-labels regularly optimize the embedding space's structure through graph-based contrastive learning, which allows data with similar pseudo-labels to have similar feature embeddings in latent space. PGR unifies pseudo-label and feature representation of unlabeled data to improve the ability of model to resist noise interference and generalization ability. Extensive experiments on four benchmark datasets demonstrate that PGR can generate higher quality pseudo-labels for unlabeled data, and is superior to the state-of-the-art (SOTA) methods. The code is available at https://github.com/song-leap/PGR .