Dual Graph Networks with Synthetic Oversampling for Imbalanced Rumor Detection on Social Media

Rumor detection is to identify and mitigate potentially damaging falsehoods, thereby shielding the public from misleading information. However, existing methods fall short of tackling class imbalance, meaning rumor is less common than true messages, as they lack specific adaptation for the context of rumor dissemination. In this work, we propose Dual Graph Networks with Synthetic Oversampling (SynDGN), a novel method that can determine whether a claim made on social media is rumor or not in the presence of class imbalance. SynDGN properly utilizes dual graphs to integrate social media contexts and user characteristics to make accurate predictions. Experiments conducted on two well-known datasets verify that SynDGN consistently outperforms state-of-the-art models, regardless of whether the data is balanced or not.