Introduction: The Challenge of Rumor Dynamics
The rapid spread of misinformation on social media presents a significant challenge for public decision-making. To intervene effectively, researchers must predict not just if a rumor will spread, but how and when it will evolve. This requires modeling Temporal Knowledge Graphs (TKGs) that capture the complex interplay between timing and network structure.
Existing methods for TKG completion often focus heavily on either the sequence of events (temporal) or the network hierarchy (structural), but rarely both. This paper addresses the challenge of modeling rumor dynamics across both scales simultaneously, which is critical for identifying the "peak periods" and "influential clusters" in a rumor's lifecycle.
The HierTKG Framework
The proposed HierTKG framework integrates three sophisticated components to bridge the gap between temporal dynamics and structural hierarchy.
- Temporal Graph Network (TGN): Serving as the model's "memory," the TGN uses a message-passing architecture with temporal encoding to track how individual entities (users, posts, events) interact over time.
- Hierarchical Pooling (DiffPool): To manage the noise of individual interactions, Differentiable Pooling clusters nodes into increasingly abstract representations. This allows the model to "zoom out" and observe how rumors move between entire communities or event categories.
- Attention-based Feature Fusion: Multi-head self-attention serves as the "glue" of the system, weighing the importance of temporal and structural embeddings to prioritize the most relevant features for predicting the next link.
Experimental Evaluation
The framework was evaluated on four diverse datasets to test its robustness:
- PHEME: A dataset of real-world Twitter rumors and non-rumors from major breaking news events (e.g., the Ferguson protests). This dataset is notably noisy and sparse.
- ICEWS14 & ICEWS18: Large-scale datasets from the Integrated Crisis Early Warning System, focusing on global political and conflict events.
- WikiData: A dataset derived from user interactions on Wikipedia, modeling how page popularity and connections evolve.
Key Results and Insights
HierTKG demonstrated superior performance across the board, particularly on structured datasets. On ICEWS14, it achieved a remarkable Mean Reciprocal Rank (MRR) of 0.9845, significantly outperforming baselines like EmtE (0.703) and SPA (0.658). On the noisier PHEME dataset, it maintained a competitive MRR of 0.8802.
Ablation studies confirmed that the combination of DiffPool + Attention was the winning formula. The attention mechanism acts as a noise filter, preventing the model from being overwhelmed by irrelevant interactions in sparse environments. This is particularly important for social media data, where the signal-to-noise ratio is low.
However, a performance gap remains between curated data (ICEWS) and organic social media data (PHEME), highlighting a fundamental challenge: even the best models struggle with the informal language and "missing links" typical of human social interaction.
Conclusion and Future Directions
While HierTKG sets a new benchmark for TKG link prediction by merging the "when" with the "where," challenges remain regarding computational complexity and sensitivity to data sparsity. Future work will focus on optimizing the algorithm for real-time deployment on massive live streams and exploring data augmentation techniques to help the model learn more effectively from sparse, real-world social media interactions.
Summary updated