LLM-RDO: Large Language Model-Guided Rules Dynamic Optimization for Temporal Knowledge Graph Reasoning

Abstract

Temporal knowledge graph reasoning (TKGR) aims to predict future facts based on historical event facts. However, the traditional embedding-based methods lack interpretability and the rule-based methods are prone to falling into spurious correlation traps. Note that the recent large language model (LLM)-based methods are mostly static, open-loop generations lacking dynamic adaptability. This article proposes a novel hybrid reasoning framework, a so-called LLM-guided rule dynamic optimization (LLM-RDO), by integrating the structural robustness captured by graph neural networks (GNNs) into the discrete causal logic optimized by LLM. To extract logically coherent matching rules and the top-<inline-formula> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula> relevant relations, an association rule mining (ARM) algorithm is designed combined with temporal path matching. To generate higher-quality rules, an LLM-guided iterative dynamic feedback loop is constructed to optimize rules by integrating the semantic features of matching rules and top-<inline-formula> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula> relevant relations. Specifically, matching rules and top-<inline-formula> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula> relevant relations are input into LLM to generate rules, which are then interactively scored against current data. High-quality rules are used for subsequent result prediction, while low-quality rules are fed back into LLM for further optimization. Finally, to achieve higher prediction accuracy, a joint prediction mechanism combining rule-based and embedding-based approaches is introduced. Experiments on 4 benchmark datasets demonstrate that LLM-RDO achieves superior performance compared to existing state-of-the-art baseline methods.