JFD-GNN: A Joint Graph Neural Network Approach for Detecting Fake News

Fake news is often designed to closely mimic legitimate news, making its detection based on content analysis alone unreliable. Hence, we propose a deeper exploration of the correlation between fake news and user profiles to understand its widespread. For this purpose, we propose a Joint Fake News Detection framework, called JFD-GNN, which consists of two distinct models designed to integrate user profiles, internal preferences, and external context. The first model, based on Graph Neural Networks (GNN), incorporates user preferences and leverages the tweet history of the post’s publisher as endogenous features. Additionally, it utilizes propagation graphs as exogenous features to derive a joint representation of user engagement. The second model focuses on user profile details and textual embeddings of comments as user features, constructing a GNN-based model for user profile analysis. We use BERT and Word2Vec as data encoders. To enhance the performance of the joint model, the fused output is passed through a fully connected layer to compute the final output and then compare it with the target. We conducted several experiments on two publicly available fake news datasets, Politifact and Gossipcop, where our approach achieved 97.99% accuracy on Gossipcop dataset and 95.16% accuracy on Politifact dataset. This illustrates the effectiveness of JFD-GNN for identifying fake news, and the proposed method reaches a comparatively high level by combining various information levels from different modalities.