Cross-Event Rumour Detection Using Graph Neural Networks Remains Strongly Event Dependent

Cross-event rumour detection on social media using graph neural networks remains heavily event-dependent and difficult when models encounter unseen situations, according to new research published in Applied Network Science. Social media rumours propagate through conversational cascades where structural patterns shape how information diffuses across online communities.

To evaluate the structural properties of online rumour diffusion and model generalisation, researchers Avijit Bose, Pradyut Sarkar, and Premananda Jana analysed 4,072 conversational cascades drawn from four distinct real-world crisis events in the PHEME dataset.

The authors determined that 1,950 cascades exhibited a star-like structure, 1,661 displayed a hybrid configuration, and 461 were classified as chain-like propagation trees. When comparing rumour cascades directly against non-rumour cascades, statistical analyses identified measurable differences across several structural graph metrics. However, the researchers observed that these structural variations between rumour and non-rumour threads were generally small, indicating that structural measures alone do not offer a clear or definitive distinction between verified information and unverified rumours.

Because cascade structure alone provided insufficient separation, the investigation rigorously tested the cross-event generalisation capacity of multiple graph-based machine learning architectures. The researchers evaluated five primary graph models: Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), GraphSAGE, a hybrid combining GCN and GAT, and a hybrid architecture joining GCN with GraphSAGE.

Alongside the primary architectures, the authors evaluated an adapted bidirectional graph convolutional network known as Bi-GCN under the identical four-fold event partitions. The experimental design specifically explored the impact of replacing traditional term frequency-inverse document frequency (TF-IDF) feature vectors with contextual sentence embeddings generated by a frozen BERT encoder. Under the standard TF-IDF representation, the Bi-GCN architecture achieved a mean Leave-One-Event-Out Macro-F1 score of 0.5308 ± 0.0737.

Integrating contextual embeddings produced notable yet variable performance shifts across the evaluated benchmarks. For the standard GCN model, substituting TF-IDF features with frozen BERT embeddings raised the mean Leave-One-Event-Out Macro-F1 score from 0.5264 ± 0.0546 to 0.5928 ± 0.1550. Despite this overall increase in average detection accuracy, the authors observed that the performance improvement was not uniform across all four held-out events. The substantially wider standard deviation underscored heightened event-to-event variability, demonstrating that while rich textual context assists graph architectures, overcoming event-specific idiosyncrasies remains a major barrier in automated rumour detection.

The findings collectively emphasize the necessity of rigorous, event-separated benchmarking standards for automated content verification systems. The study concludes that both conversational propagation structures and textual representations directly govern the overall robustness of graph-based rumour detection models.

Subject of Research: Cross-event rumour detection using graph neural networks in social media cascades

Article Title: Structural patterns of rumour propagation and hybrid graph learning in multi-event social networks

Article References: Bose, A., Sarkar, P., & Jana, P. (2026). Structural patterns of rumour propagation and hybrid graph learning in multi-event social networks. Applied Network Science. https://doi.org/10.1007/s41109-026-00848-6

Image Credits: AI Generated

DOI: 10.1007/s41109-026-00848-6

Keywords: rumour detection, social media cascades, graph neural networks, GNN, PHEME dataset, conversational cascades, cross-event generalisation, leave-one-event-out, propagation structure, star-like cascades, chain-like cascades, hybrid cascades, BERT embeddings, TF-IDF, Bi-GCN, GraphSAGE, graph attention networks, macro-F1