Li, B., Song, B., He, T., & Ong, Y.-S. (2026). A Structural Knowledge Enhanced Re-ranking method with large language models for temporal knowledge graph prediction. Pattern Recognition, 113834. https://doi.org/10.1016/j.patcog.2026.113834
Abstract:
Temporal Knowledge Graph Reasoning (TKGR) aims to predict unknown facts at future timestamps by understanding the evolution of structured facts over time, which is critical for applications such as financial risk warning and policy impact forecasting. In this paper, we propose a Structural Knowledge Enhanced Re-ranking method (SKER) with large language models (LLMs) for temporal knowledge graph prediction. SKER is a novel reasoning method that combines the advantages of structural modeling through Graph Neural Networks (GNNs) and the reasoning capabilities of LLMs to enhance the performance in TKGR tasks. In SKER, we retrieve pivotal historical node information and contextual information relevant to reasoning tasks from temporal and structural dimensions by the designed temporal-weighted GNNs, respectively, thus providing accurate reasoning evidence for predictions within the limited input token length of LLMs. Furthermore, to address the challenges of uncontrollable format and content of LLMs, we formalize TKGR as re-ranking tasks of predictions and introduce a dual-identifier constraint policy to refine the LLM output, hence ensuring the generation of reliable predictions. Experimental results demonstrate that without fine-tuning the LLMs, SKER outperforms state-of-the-art methods on benchmark datasets that record real-world temporal events, showcasing its accuracy and reliability in utilizing LLMs for TKGR.
License type:
Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)
Funding Info:
This work has been supported by the China Scholarship Council Program (Project ID: 202406960051), the National Natural Science Foundation of China (No.62372357), and the Aviation Transformation Programme-Airport (ATP_Airport) of Singapore (Award No. ATP_AIRPORT/2026/ARES/FLASH).