.png){kind=icon}
Return
Explainable extrapolation on temporal knowledge graphs via relation-driven and context-aware logical rules
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2026.116537.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Extrapolation on temporal knowledge graphs (TKGs) aims to predict future events based on historical facts, thereby supporting dynamic decision-making. Although neural network-based approaches have achieved promising results, they often suffer from limited interpretability, whereas rule-based approaches primarily focus on enhancing the explainability of predictions. However, existing rule-based methods face two critical limitations: (1) they neglect the dynamic evolution of relations and fail to capture semantic transitions; and (2) they overlook recurrent interaction patterns and the contextual dependencies between historical and future events. To address these challenges, we propose RCLR, a novel framework that integrates relation-driven and context-aware logical rules to improve predictive accuracy and interpretability jointly. RCLR introduces a relation-driven Markov transition matrix to guide temporal random walks, effectively capturing the dynamic evolution of relational semantics. Simultaneously, it extracts three categories of context-aware logical rules (Precursor, Bridging, and Consequent rules), enabling the discovery of expressive and contextually coherent reasoning patterns that transcend conventional rule mining constraints. Extensive experiments on six benchmark datasets demonstrate that RCLR consistently outperforms state-of-the-art methods, achieving up to a 13.13% improvement in Mean Reciprocal Rank (MRR). Moreover, RCLR provides interpretable reasoning paths for its predictions, offering transparent insights into the temporal extrapolation process.
Keywords:
Temporal knowledge graph
Explainable
Rule learning
Context-aware
Link prediction
Journal
K
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
Cited Papers
No cited papers available