Get started
Work

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

Year: 2026
Type: article
Abstract: Rumours on social media propagate through conversational cascades whose structural characteristics influence patterns of information diffusion across online communities. Graph Neural Networks (GNNs) have been applied to rumour detection using the structural information available in social media conversations. However, it remains important to examine how these models perform when the test conversations belong to events that were not represented during training. In this study, we analyse 4,072 conversational cascades from four events in the PHEME dataset: Charlie Hebdo, Ferguson, Germanwings Crash, and Sydney Siege. For each discussion thread, we use the available reply relationships to construct a directed graph and examine its propagation structure. Of the 4,072 retained cascades, 1,950 are classified as star-like, 1,661 as hybrid, and 461 as chain-like. The statistical comparisons identify differences between rumour and non-rumour cascades in several structural measures, but these differences are generally small and do not provide a clear distinction between the two classes. To investigate cross-event generalisation, five graph-based models-GCN, GAT, GraphSAGE, Hybrid GCN + GAT, and Hybrid GCN+GraphSAGE-are assessed through a strict Leave-One-Event-Out (LOEO) protocol. The evaluation is organised into four folds, with each PHEME event serving once as the test event and the remaining three events providing the training data. We also evaluate an adapted Bi-GCN using the same event partitions and investigate the effect of replacing TF-IDF features with contextual embeddings obtained from a frozen BERT encoder. The results confirm that cross-event rumour detection remains challenging and strongly event-dependent. For GCN, replacing TF-IDF with frozen BERT embeddings increases the mean LOEO Macro-F1 from 0.5264 ± 0.0546 to 0.5928 ± 0.1550, although the improvement is not consistent across all held-out events. Bi-GCN with TF-IDF obtains a mean Macro-F1 of 0.5308 ± 0.0737. These findings show that contextual representations can improve average cross-event performance, while their higher event-to-event variability highlights the continuing difficulty of generalisation to unseen events. Overall, the study demonstrates the importance of strict event-separated evaluation and shows that both propagation structure and textual representation influence the robustness of graph-based rumour detection.
DOI: 10.1007/s41109-026-00848-6
Language: English
Open Access status: gold
Estimated APC paid: $1,690
OpenAlex ID: W7225826731
FWCI:
Cites:
Cited by:
Related to: