Graph Contrastive Adversarial Learning for Rumor Detection via Similarity-Preserving

Jing Dai, Jiachen Ma, Wei Zhang, Yong Liu

Abstract: Graph contrastive learning (GCL) has been widely applied in rumor detection due to its ability to effectively capture propagation structure information while reducing dependency on labels. Existing methods primarily combine GCL with adversarial training (AT) to enhance the robustness of GCL-based rumor detection models. However, applying AT to GCL-based models forces the node representations in the original propagation structure to become closer to those in the attacked propagation structure, leading to nodes with similar features not having similar representations. This paper introduces a novel adversarial graph contrastive learning framework that incorporates a similarity-preserving mechanism to compare augmented propagation structures with their similarity-preserving counterparts (SPC). The SPC component maintains node-level feature similarity by deriving self-supervised signals from the original node attributes. Furthermore, to enhance the utilization of node feature information, we integrate an Adversarial Training Module (ATM), designed to generate contradictory negative instances that assist in identifying robust rumor-related features. Empirical evaluations on publicly available datasets confirm that our method achieves superior performance compared to existing state-of-the-art approaches.

External IDs:dblp:conf/icic/DaiMZL25