A Multimodal Graph Framework for News Credibility Assessment and Propagation-Level Prediction
TL;DR
A multimodal framework that produces a credibility classification and a propagation-level prediction within a single processing pipeline is evaluated that combines frozen RoBERTa features, engineered interaction statistics, a standard GCN, and Transformer self-attention over modality representations.
Abstract
The rapid dissemination of misinformation through online social networks creates challenges for information reliability and network stability. We evaluate a multimodal framework that produces a credibility classification and a propagation-level prediction within a single processing pipeline. The framework combines frozen RoBERTa features, engineered interaction statistics, a standard GCN, and Transformer self-attention over modality representations. These are established components; the purpose of the framework is to integrate the available modalities and return both outputs from one model, rather than to introduce a new encoder or attention mechanism. On 12,701 MCFEND news items, the framework obtains 94.77% accuracy and 95.79% F1-score for credibility classification, together with an MAE of 0.34, RMSE of 0.44, and R2 of 0.96 for propagation-level prediction. The small variation across five matched seeds supports the stability of these means under the fixed protocol. The credibility F1 difference from modality-matched MLP-Fusion is not significant after Holm correction (adjusted p = 0.0810), and five pairs provide limited power for detecting small differences; the two implementations are therefore interpreted as having close performance. RandomForest also obtains lower propagation errors than the evaluated framework. The implemented retrospective fractional-observation protocol observes a fraction defined by final cascade size and uses snapshot engagement values with a fixed random split. In one diagnostic, replacing the fractional-observation rule with fixed K = 15 preserves classification F1 at a similar level but reduces the framework’s propagation R2 from 0.963 to 0.896. In a separate target-definition diagnostic conducted with the original observation setting, predicting residual future interactions yields an R2 of 0.870. The findings therefore characterize retrospective within-dataset prediction and do not establish leakage-free early forecasting. The two outputs are interpreted separately because the present experiments evaluated one shared dual-output configuration rather than comparing it with two independently optimized systems.