A Spatio-Temporal Attention-Driven Graph Neural Network for Early Fake News Detection Using Multi-Stage Propagation Modeling

The spread of fake news has damaging social, political, and economic effects. It has become a major problem, particularly with the rapid spread of these news stories across social media platforms. Attempting to detect misinformation quickly is a problem because not much context is provided about the news in its early stages of dissemination. In this paper, we describe our Spatio-Temporal Attention-Driven Graph Neural Network (STAGNN), which utilizes multi-stage propagation to detect fake news in its early stages. This makes dynamic propagation graphs to model user and news content interactions to represent user interactions. Using a hierarchical multi-stage propagation method, we capture the broad, global interactions of news, and an attention-based aggregation method, we capture important news propagation. We also employ a unique hybrid loss function which includes a propagation regularization loss and a classification loss, which strengthens the model’s robustness and generalization while sparse early-stage propagation is modeled. Our STAGNN framework has been shown in multiple studies to surpass existing graph neural network models by at least 4% to 10% when measured with the F1 statistic and in other metrics for measured effectiveness. Our framework should be the go-to solution in real-world social media environments.