Fake News Detection Using BERT + Graph Neural Networks

The rapid dissemination of misinformation and fake news through online platforms has become a major challenge, impacting public perception, social stability, and decision- making. Existing transformer-based models, such as BERT, demonstrate strong textual understanding but fail to capture the relational and propagation dynamics of news across social networks. Conversely, graph-based models effectively model relationships between users, posts, and comments but lack semantic comprehension of textual content. To overcome these limitations, Zhang et al. (2024) introduced GBCA (Graph-BERT Co- Attention), which integrates Graph Convolutional Networks (GCNs) with BERT through a co-attention mechanism. While GBCA achieves significant performance gains, it still exhibits several drawbacks — it models homogeneous graphs, ignores temporal propagation, and underperforms in noisy or sparse network environments. This research proposes an enhanced hybrid framework, Heterogeneous Temporal Graph- BERT (HTGBERT), that extends GBCA by introducing heterogeneous graph modelling, temporal learning, and contrastive pretraining for robust fake news detection. The proposed model encodes textual semantics using BERT, constructs a heterogeneous social graph incorporating posts, users, comments, and entities, and applies a Temporal Graph Neural Network (TGAT/TGN) to learn propagation dynamics over time. A cross-modal contrastive learning module is employed to align text and graph representations, improving generalization and robustness to sparse or noisy data. Experiments will be conducted on benchmark datasets including Fake Newsnet, Twitter15/16, and PHEME, comparing the proposed model against baselines such as BERT- only, GNN-only, and GBCA. Performance will be evaluated using accuracy, F1-score, AUC, and time-to-detection metrics under event-separated evaluation protocols to ensure realistic generalization. The proposed HTGBERT framework is expected to achieve earlier, more accurate, and explainable detection of fake news by integrating semantic, structural, and temporal dimensions of information dissemination. This research not only advances hybrid fake news detection techniques but also contributes a reproducible, explainable, and temporally aware framework for real-world misinformation mitigation.