GraphSAGE boosts fake news domain detection by 14% using network structure

A new arXiv paper (2608.02399) from researchers at the Max Planck Institute and collaborators tackles unreliable news detection by shifting focus from individual articles to source domains. Instead of relying solely on content—which low-reliability outlets increasingly mimic with credible journalism and AI-generated text—they leverage network structure derived from URL-sharing patterns in Telegram chats. Using a statistically validated domain co-sharing network, they found strong assortative mixing: low-reliability domains disproportionately share links with each other, and reliable domains do the same.

To exploit this structure, the team compared Graph Neural Networks (GNNs) against network-unaware baselines (Multi-Layer Perceptrons) using both content-aware features (multilingual text embeddings) and content-agnostic features (spreading dynamics). GraphSAGE, a popular GNN variant, delivered the best results in both settings: 0.63 accuracy with content and 0.53 without content, representing a 13-14% relative improvement over the MLP baseline on identical features. This demonstrates that network topology systematically improves domain reliability assessment—and remains effective even when content analysis is infeasible, offering a robust, harder-to-game signal for misinformation detection.