CGAR: A Contrastive Graph Attention Residual Network for Enhanced Fake News Detection

The detection and mitigation of fake news are critical to bolstering online security, upholding accurate knowledge, and safeguarding freedom of speech. Recent efforts have also focused on the structural properties of news propagation beyond the analysis of textual features for fake news detection. Although graph models employing cross-entropy loss can rapidly identify salient fake news, they often suffer from poor generalization, especially when faced with class imbalances. This study introduces a novel Contrastive Graph Attention Residual Network (CGAR) designed to tackle these complexities. The proposed CGAR model integrates a Propagation and Dispersion Graph Neural Network, merging a Graph Convolutional Network (GCN) with a Graph Attention Network (GAT) via residual connections, thereby improving the model’s ability to extract local graph features and identify long-range dependencies. Additionally, the integration of contrastive learning into the loss function enables the model to explicitly differentiate between conversational threads of identical and distinct classes, thereby addressing the challenge of class imbalance by emphasizing sample similarities. Empirical evaluations on two public benchmark datasets reveal that CGAR surpasses competing state-of-the-art models in performance.