Bridging Graph Structure and Knowledge-Guided Editing for Interpretable Temporal Knowledge Graph Reasoning

TL;DR

IGETR addresses the challenge of temporally coherent and interpretable TKGR by combining a semantic-aware temporal GNN with an LLM-driven path-editing module and a temporal graph Transformer for evidence aggregation. The three-stage pipeline grounds reasoning in structural data, refines candidate paths with external knowledge, and fuses multi-hop evidence while preserving temporal causality. Empirical results on ICEWS benchmarks show state-of-the-art performance and robustness to hallucination, supported by ablations and case studies that demonstrate the value of explicit path editing. The work advances trustworthy TKGR by enabling controllable, knowledge-guided reasoning grounded in reliable graph evidence, with practical implications for high-stakes temporal inference.

Abstract

Temporal knowledge graph reasoning (TKGR) aims to predict future events by inferring missing entities with dynamic knowledge structures. Existing LLM-based reasoning methods prioritize contextual over structural relations, struggling to extract relevant subgraphs from dynamic graphs. This limits structural information understanding, leading to unstructured, hallucination-prone inferences especially with temporal inconsistencies. To address this problem, we propose IGETR (Integration of Graph and Editing-enhanced Temporal Reasoning), a hybrid reasoning framework that combines the structured temporal modeling capabilities of Graph Neural Networks (GNNs) with the contextual understanding of LLMs. IGETR operates through a three-stage pipeline. The first stage aims to ground the reasoning process in the actual data by identifying structurally and temporally coherent candidate paths through a temporal GNN, ensuring that inference starts from reliable graph-based evidence. The second stage introduces LLM-guided path editing to address logical and semantic inconsistencies, leveraging external knowledge to refine and enhance the initial paths. The final stage focuses on integrating the refined reasoning paths to produce predictions that are both accurate and interpretable. Experiments on standard TKG benchmarks show that IGETR achieves state-of-the-art performance, outperforming strong baselines with relative improvements of up to 5.6% on Hits@1 and 8.1% on Hits@3 on the challenging ICEWS datasets. Additionally, we execute ablation studies and additional analyses confirm the effectiveness of each component.

Figures (6)

• Figure 1: The architecture of IGETR. The model consists of three key stages: (A) LLM-empowered GNN reasoning, where a sampling-based GNN extracts candidate entities and reasoning paths from the Temporal Knowledge Graph (TKG) with attention-based scoring; (B) LLM-guided path refinement, where the extracted paths are iteratively revised and optimized to improve logical consistency and relevance; and (C) Graph Transformer-based learning, which processes the refined paths to enhance interpretability and structural coherence. By integrating GNNs for structured data representation and LLMs for contextual reasoning, IGETR ensures both accurate and explainable TKGR.
• Figure 2: Prompt design for semantic embedding initialization. The prompt template guides the LLM to generate textual descriptions for entities and relations, as well as passive forms for reverse relations, enhancing semantic understanding for subsequent embedding and reasoning steps.
• Figure 3: Prompt design for path optimization. The structured prompt clearly instructs the LLM to revise, keep, or remove segments of extracted inference paths according to specified logical constraints and processing rules, thus ensuring optimized relevance, accuracy, and interpretability of the reasoning paths.
• Figure 4: We conducted experiments on the ICEWS14 and ICEWS0515 datasets to demonstrate the effectiveness of the LLM-based path editing module. The results indicate a performance improvement across evaluation metrics when incorporating the LLM-based path editing.
• Figure 5: Case study demonstrating LLM hallucination in dynamic temporal reasoning.