Semantic-aware contrastive learning for graph classification

Abstract

Currently, the performance of vanilla graph-level contrastive learning is limited by traditional data augmentation strategies and size imbalance. First, adding noise in the embedding space for data augmentation may cause the graph-level representation to exceed the class decision boundary and alter its original semantic information. Second, transferring information from head graphs to tail graphs to alleviate size imbalance without distinguishing the semantic information of the graphs may lead to suboptimal model performance. To address these issues, we propose a semantic-aware graph-level contrastive learning method named SAGCL. Specifically, SAGCL achieves controllable data augmentation by adjusting the closeness between class center features and augmented features, preserving the inherent structure and semantic information of the graph. Meanwhile, the intra-cluster variance is used as a regularization term to maintain the uniformity of the feature distribution. In addition, SAGCL employs a confidence-weighted approach to obtain the semantic prototypes of head graphs and tail graphs within each cluster. Then, the rich semantic information from the head graphs is transferred to the tail graphs, effectively enhancing the model’s ability to distinguish tail graphs. Experiments on graph classification tasks on eight imbalanced datasets demonstrate that SAGCL outperforms existing state-of-the-art methods.