GA-Clip: Semantic-Aware Graph Augmentation for Contrastive Learning

Shuaiqi Lu, Yi Guo, Zhenlin An, Yan Zhu, Ning Huang

Abstract: Recent advancements in Text-Attributed Graphs (TAGs) have attracted significant attention for their wide-ranging applications in domains such as social networks, academics, and e-commerce. The powerful text-processing capabilities of pre-trained language models offer a promising avenue for effectively integrating textual attributes with graph structures. However, existing methods exhibit two key limitations: (1) reliance on rigid graph construction processes that fail to capture a comprehensive view of the text-attributed graph data; (2) insufficient fusion of textual semantics and graph topology, leading to information loss, unstable training, and limited generalization across diverse downstream tasks. In this work, we propose GA-Clip, a novel semantic-aware graph augmentation contrastive learning model. We leverage the pre-trained language model to generate semantic edges that extract the fine-grained topology within the text feature space to augment the graph structure. We then separately employ the graph and text encoders to sufficiently fuse the different modalities through a modified self-supervised contrastive learning approach. This augmentation mitigates the dependency on cumbersome graph construction processes and integrates information from different modalities, which jointly enables scalable graph learning on coarse-grained, large-scale source data. Experimental results demonstrated that our approach achieved a 2-4% improvement over the SOTA methods in accuracy across multiple datasets, validating the effectiveness of the proposed method.

External IDs:dblp:conf/icmcs/LuGAZH25