Precision through progression: Empowering temporal knowledge graph reasoning with knowledge-guided chain of thought

Abstract

Temporal Knowledge Graphs (TKGs) have emerged as a powerful paradigm for event forecasting, owing to their ability to dynamically represent the evolving relationships between entities over time. By effectively reasoning along the temporal dimension, TKGs help address real-world data incompleteness through inference of missing facts. Recent advances in large language models (LLMs) have led to their integration with TKG reasoning tasks. However, current LLM-based approaches face three critical challenges: (1) insufficient utilization of background knowledge, (2) inadequate modeling of the evolving temporal dynamics intrinsic to TKGs, and (3) difficulty in bridging the structural mismatch between the graph structure and the sequential operation mode of LLMs. To address these challenges, we propose EV-COT, a novel EVent-aware Chain-Of-Thought reasoning framework designed to explicitly model event evolution through structured, interpretable reasoning chains. EV-COT comprises three modular, plug-and-play components – knowledge module, perception module, and thinking module – that work collaboratively to extract essential event-related cues for enhanced reasoning. Specifically, the knowledge module generates high-quality contextual knowledge to enrich entity representation, and the perception module captures intricate structural and temporal patterns inherent in TKGs. Moreover, the thinking module extracts temporal logical rules, facilitating interpretable step-by-step reasoning. By effectively integrating these diverse contextual knowledge, EV-COT delivers more accurate predictions. Extensive evaluations on three datasets demonstrate that EV-COT consistently outperforms state-of-the-art methods, highlighting its effectiveness for precise event forecasting in TKGs.