Integrating Danger Theory into Graph Neural Networks for Early Warning Fake News Detection

Abstract

The rapid propagation of fake news on social media poses a significant threat to society, demanding detection methods that are not only accurate but also timely. While Graph Neural Networks (GNNs) are powerful tools for modeling propagation cascades, they often struggle in early detection scenarios where structural information is scarce. This paper presents a bio-inspired hybrid architecture, the Hybrid Danger-inspired GNN (HybridDCAGNN). We propose a custom, trainable graph convolutional layer, the DCAConv, which learns to generate a “danger signal” for each node by analyzing the directional flow of information, inspired by the Danger Theory of the immune system. This signal enriches the node features fed into standard GNN backbones (GCN, SAGE, GIN, CHEB). Through extensive experiments on the GossipCop and Politifact datasets, we demonstrate that our hybrid approach provides a statistically significant advantage in early detection (at 10-40% of the cascade). Furthermore, initial explainability analysis suggests our model learns to highlight nodes that are influential in the propagation, aligning with the bio-inspired premise. While these findings are promising, we suggest that further quantitative analysis using network science metrics is needed to fully validate the mechanism. We conclude that integrating Danger Theory principles provides a robust framework for enhancing fake news detection, especially when it matters most: at the beginning of the spread.