Bridging graph structure and knowledge-guided editing for interpretable temporal knowledge graph reasoning.

Area of Science:

• Artificial Intelligence
• Data Science
• Knowledge Representation

Background:

• Existing Large Language Model (LLM)-based methods for temporal knowledge graph reasoning (TKGR) struggle with dynamic graph structures, prioritizing context over structural relations.
• This leads to difficulties in extracting relevant subgraphs, resulting in unstructured and hallucination-prone inferences, particularly with temporal inconsistencies.

Purpose of the Study:

• To propose IGETR (Integration of Graph and Editing-enhanced Temporal Reasoning), a hybrid framework combining Graph Neural Networks (GNNs) and LLMs for robust TKGR.
• To enhance the understanding of structural information in dynamic graphs and improve the accuracy and interpretability of future event predictions.

Main Methods:

• IGETR employs a three-stage pipeline: temporal GNN for identifying coherent paths, LLM-guided editing for refining paths, and integration for final predictions.
• The framework grounds reasoning in data via structurally and temporally coherent paths identified by a temporal GNN.
• LLM-guided path editing leverages external knowledge to correct logical and semantic inconsistencies in initial paths.

Main Results:

• IGETR achieves state-of-the-art performance on standard TKGR benchmarks, including the challenging ICEWS datasets.
• Relative improvements of up to 5.6% on Hits@1 and 8.1% on Hits@3 were observed compared to strong baselines.
• Ablation studies and additional analyses confirm the effectiveness of each component within the IGETR framework.

Conclusions:

• The proposed IGETR framework effectively addresses limitations in existing LLM-based TKGR methods by integrating GNNs and LLMs.
• IGETR demonstrates superior performance in predicting future events from dynamic knowledge graphs, offering accurate and interpretable results.
• The hybrid approach enhances structural information understanding and reduces inference-related hallucinations in temporal reasoning.