Predicting future events based on historical data—such as forecasting international conflicts or economic trends—is a critical capability of Artificial Intelligence. However, existing methods often struggle because they ingest too much irrelevant historical information, creating "noise" that reduces accuracy. To address this, a research team led by Fuzhen Zhuang from Beihang University published their new research in Frontiers of Computer Science, co-published by Higher Education Press and Springer Nature.
Temporal Knowledge Graphs (TKGs) are essential for organizing dynamic data to predict future facts. However, current reasoning models typically aggregate all historical facts indiscriminately. This information redundancy introduces outdated or irrelevant interactions that mislead the reasoning process. For instance, when predicting a political interaction between two specific countries, historical trade data involving a third, unrelated country can act as noise, hindering the model's ability to make accurate forecasts.
To solve this problem, the team proposed a Temporal knowledge graph reasoning model via Multi-granularity Knowledge Refinement (T-MKR). This novel approach is designed to "prune" the historical data, selectively removing irrelevant edges and unnecessary nodes to focus on the most valuable information. The model functions through a unique multi-granularity mechanism. It employs a History-based Edge Selector to filter specific relationships relevant to the current query, removing cyclic or redundant facts, while simultaneously using a History-based Node Selector to identify relevant entities and find high-value hidden connections. By integrating these refined subgraphs using a gating mechanism and contrastive learning, the model ensures that the reasoning process is driven by the most informative historical patterns.
The researchers validated T-MKR on six widely used benchmark datasets, including ICEWS, GDELT, and WIKI. The experimental results demonstrated that T-MKR significantly outperforms state-of-the-art baselines in accuracy. Furthermore, the model is highly efficient; it achieved the shortest training time across most datasets, offering a speedup of at least 2x compared to other leading models.
DOI:10.1007/s11704-026-51857-8