LLM-Guided Explanation Generation for Knowledge Graph Embedding-Based Node Classification

ACL ARR 2025 February Submission3835 Authors

Abstract: Knowledge graph (KG) embedding models have achieved remarkable success in various tasks, particularly in node classification. However, their decision-making processes remain opaque, limiting interpretability and trustworthiness. To address this challenge, we propose the first method for explaining KG embedding-based node classification. It integrates large language models (LLMs) to generate both graph-structured and textual explanations, offering deeper insights into model reasoning. Specifically, we train a proxy model to approximate the behavior of the original KG embedding model. Leveraging the distilled knowledge from this proxy model, an LLM is finetuned to identify and reason about critical relation path patterns that significantly influence predictions. Guided by the selected patterns and proxy model, we design an efficient searching algorithm to extract the final set of critical triples with LLM-generated reasoning. Experiments on three representative KG embedding models across multiple benchmark datasets demonstrates the effectiveness and generalization of our method in explaining KG embedding-based node classification.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: explanation faithfulness, feature attribution

Contribution Types: Model analysis & interpretability

Languages Studied: English

Submission Number: 3835