Enhanced temporal knowledge graph reasoning through TKG-LPEHD: LLM-driven dynamic fusion of graph and rule-based methods

Temporal Knowledge Graphs are essential for understanding dynamic, time-evolving real-world scenarios. However, the effective modeling of temporal validity during reasoning remains a critical and often overlooked challenge. Conventional methods relying on static representations or predefined temporal windows struggle to capture the context-dependent nature of rule applicability over time, leading to inaccurate predictions and a constrained understanding of evolving relationships. To address these issues, this study proposes TKG-LPEHD, a novel framework that deeply integrates rule-based reasoning, graph-based reasoning and large language models to dynamically model the temporal validity of rules. A key feature is the LLM-guided mechanism for dynamic rule extraction and dynamic weight fusion reasoning. The framework employs an LLM-guided three-dimensional relational path sampling strategy with LLM assisted validity assessment, and fuses rule-based and graph-based reasoning through dynamic weights to ensure both high quality and temporal relevance of inference rules. The design further optimizes Neo4j storage architecture, temporal SPARQL query engineering and systematically tuned prompt engineering to facilitate real-world deployment. Evaluations on public benchmarks and a confidential Chinese temporal knowledge dataset demonstrate that the proposed approach significantly outperforms state-of-the-art methods and achieves outstanding performance on long-range temporal reasoning tasks. This research not only validates the paradigm of explicitly incorporating temporal validity as a core reasoning dimension but also provides a more explainable and time-aware solution for high-stakes applications including financial forecasting and geopolitical analysis.