Social media rumor detection using multi-relational temporal-semantic augmented dynamic graph neural networks

The rapid dissemination of rumors on social media has caused tremendous effects in both the online and offline world. Although graph neural networks have shown promise in modeling propagation structures, many existing dynamic approaches rely on fixed-interval snapshots and a restricted number of relational views. This rigid segmentation impairs the continuous capture of critical dynamics during rapid outbreaks and introduces substantial noise and computational redundancy during dormant periods, thereby resulting in feature dilution. Furthermore, these methods often overlook implicit semantic correlations between posts that lack direct structural connections. To address these limitations, we propose a Multi-Relational Temporal-Semantic Augmented Dynamic Graph Neural Network (MRTS). First, we introduce an Adaptive Temporal Segmentation strategy that partitions the event timeline based on propagation velocity. This approach enables the precise capture of fine-grained burstiness while reducing computational redundancy. Second, we construct a unified Triple-View Dynamic Graph by integrating user interaction, propagation, and semantic similarity views to comprehensively model heterogeneous dependencies. Finally, to ensure robust view fusion, we design a Spatio-Temporal Feature Learning Module. This module comprises a Cross-Modal Synergistic Attention mechanism for non-linear filtering and a Relation-Aware Time-Decay Fusion Unit for adaptive temporal evolution modeling. Experimental results on two public real-world datasets indicate that MRTS performs favorably against current state-of-the-art baselines.