A Heterogeneous Graph Attention Network with Pre-trained Language Models for Multi-Modal Fake News Detection

Abstract

The spreading of fake news via online social media has emerged as one of the major issues in the modern information systems. The existing fake news detection techniques mainly rely on text-based classification methods and are ineffective in incorporating multimodal features, which can be used for fake news detection and classification. This study proposes HGAT-BERT, where a pretrained BERT-Large encoder is coupled with a Heterogeneous Graph Attention Network (HGAT) to learn joint representations for textual, social network and external knowledge graph features. Cross-modal attention and gated residual connections facilitate the integration of these heterogeneous feature streams into a unified representation. On the FakeNewsNet dataset split on PolitiFact and GossipCop data, the research reports an accuracy of 93.7% and an F1 score of 93.2%, an improvement of around 3.5% over the existing methods. The ablation analysis shows that all components represent meaningful contributions to the overall model performance, with cross-modal attention being responsible for the largest marginal contribution.