FL-Evo: Jointly modeling fact and logic evolution patterns for temporal knowledge graph reasoning

Abstract

Temporal knowledge graphs (TKGs) extrapolation reasoning, intending to predict future events given the known KG sequence, benefits broad applications like policy-making and financial analysis. The key to this issue is to discern how knowledge evolves within these sequences. Currently, most works focus on modeling the evolution patterns through continuous sampling from TKGs, without ensuring the samples contain relevant facts or considering the knowledge beyond the samples. Faced with these challenges, we propose a novel model that performs prediction by capturing fact and logic knowledge evolution patterns (FL-Evo). For modeling fact evolution pattern, the fact knowledge is first distilled from large language models using designed prompts and subsequently refined with TKG. Then, entity-based subgraph sampling strategy extracts relevant facts from the TKG, capturing fact evolution patterns. Furthermore, logical knowledge mined from the TKG helps to derive the corresponding evolution pattern. Finally, the outputs of these two evolution patterns are integrated to realize the final prediction. Experimental results on five benchmark datasets demonstrate that FL-Evo outperforms existing temporal knowledge graph reasoning models, with improvements of up to 3.97 % in Hit@3 and 4.07 % in Hit@10. Notably, FL-Evo substantially enhances reasoning performance for unseen entities lacking prior records.