Dynamic subgraph pruning and causal-aware knowledge distillation for temporal knowledge graphs

Temporal Knowledge Graph (TKG) reasoning has attracted attention for its ability to capture temporal evolution patterns and improve computational efficiency. However, existing methods still encounter challenges in entity and relation prediction tasks. To overcome these, we introduce DynTKG, a novel approach that combines dynamic subgraph pruning and causal-aware knowledge distillation. Using a time-decay Hawkes process, DynTKG filters historical events to reconstruct critical temporal subgraphs, effectively reducing redundant computations while maintaining essential dependencies. A gradient-sensitive graph attention mechanism alleviates semantic conflicts by adjusting node weights based on gradient norms, allowing the model to focus on conflict-free patterns. To optimize efficiency, DynTKG employs rule-guided contrastive knowledge distillation, transferring knowledge from a hybrid neural-symbolic teacher model to a lightweight student model, achieving significant compression with minimal performance loss. Extensive experiments on various real-world datasets demonstrate that DynTKG achieves notable improvements in entity prediction accuracy and inference speed. Its causal path visualizations enhance interpretability, while case studies in healthcare event prediction and financial risk forecasting further highlight its practical value in real-world applications.