Temporal knowledge graphs reasoning with iterative guidance by temporal logical rules

Abstract

Reasoning is essential for the development of large temporal knowledge graphs, which aim to infer new facts based on existing ones. Recent temporal knowledge graph reasoning methods mainly embed timestamps into low-dimensional spaces. These methods focus on entity reasoning, which cannot obtain the specific reasoning paths. More importantly, they ignore the logic and explanation of reasoning paths in temporal knowledge graphs (TKGs). To overcome this limitation, we propose a novel Temporal Logical reasoning Model, denoted as TLmod. This model represents a reasoning process that works through iterative guidance by temporal logical rules. More importantly, we propose two principles of temporal logical rules and define five types of temporal logical rules. Meanwhile, considering the diversity of temporal logical rules, we propose a pruning strategy for obtaining them and calculating the confidence score by combining traversing and random selection. Experimental results show that our model outperforms most metrics compared to prior state-of-the-art baselines across two benchmarks. In addition, analysis of the ablation experiment reveals the validity and importance of temporal logical rules in TKGs.