NETWORK INFORMATION ENHANCES UNRELIABLE NEWS DOMAIN DETECTION

As generative AI and sophisticated mimicry make content-based detection of unreliable news increasingly difficult, this study investigates whether the structural patterns of how news is shared can serve as a more robust signal for reliability. The authors move beyond article-level analysis to a domain-level approach, constructing a statistically validated domain co-sharing network based on URL-sharing patterns within Telegram chat groups. By applying the Bipartite Configuration Model (BiCM), they filter out spurious connections to create a robust network backbone. They then compare various Graph Neural Networks (GNNs)—specifically GCN, GAT, and GraphSAGE—against a network-unaware Multi-Layer Perceptron (MLP) baseline to predict the reliability of news domains.

The researchers identify a clear assortative mixing pattern in the Telegram co-sharing network: reliable domains tend to be shared alongside other reliable domains, while unreliable domains cluster together. This structural signal is highly predictive of source quality. The study demonstrates that GNNs consistently outperform the MLP baseline, with the GraphSAGE architecture achieving the highest accuracy (0.63 with content features, 0.53 without). Notably, the network-aware models maintain superior performance even when textual content analysis is entirely excluded, relying solely on spreading dynamics and metadata. This suggests that the relational topology of information sharing provides a powerful, complementary, and sometimes sufficient signal for identifying low-reliability news sources.

This research provides a scalable alternative to manual fact-checking and content-heavy detection methods. As platforms become more decentralized and content becomes harder to verify through text alone, leveraging the collective intelligence embedded in user sharing behavior offers a promising path for automated misinformation detection. The findings highlight that the social context of information dissemination is as critical as the information itself for assessing credibility.