Promoting Rumor Detection by Adaptive Graph Augmentation Based Contrastive Learning

Abstract

The widespread proliferation of social media has brought convenience to individuals, which has also intensified the propagation of rumors. Existing graph-based approaches primarily concentrate on identifying the structural patterns of rumor propagation. However, the dissemination structures of rumors are highly susceptible to interference noises from various sources, which undermines the robustness of the models. Inspired by the success of the adaptive learning mechanism, we propose a novel Graph Adaptive Enhanced Contrastive Learning (GAECL) method to enhance the robustness of the rumor detector by introducing graph perturbation. Specifically, we design an adaptive graph augmentation strategy by dynamic edge perturbation and attribute masking. These strategies not only preserve its intrinsic structure and properties but also enrich the diversities of topological structures. Simultaneously, graph contrastive learning is further employed to improve the rumor-discriminating ability, where samples with consistent labels are considered as positive pairs and those with different labels as negative pairs, effectively discerning the perturbed and original graphs. Experimental results on three public datasets demonstrate that our proposed GAECL outperforms state-of-the-art methods in enhancing rumor detection robustness.

Keywords: Rumor Detection, Social Media, Graph Contrastive Learning, Adaptive Graph Augmentation