First, explain graph neural networks in natural language.
[Introduction] GraphNarrator is the first tool developed by a research team at Emory University to generate natural language explanations for graph neural networks. By constructing and optimizing explanation pseudo-labels and then distilling these labels into an end-to-end model, the model can directly output high-quality natural language explanations, making the complex decision-making process of graph neural networks transparent and understandable. Its effectiveness has been verified on multiple real-world datasets.
Graph neural networks (GNNs) have become core tools for processing structured data and are widely used in scenarios such as social networks, drug design, and financial risk control.
However, the decision-making process of existing GNNs is highly complex and often lacks transparency: Why does the model make such predictions? Where are the key bases? This has become an important bottleneck hindering their large-scale implementation.
Most existing methods are based on "important subgraph extraction" or "node-edge attribution," such as GNNExplainer and PGExplainer. However, they can only output structural fragments, lack human readability, and have no ability to handle text-attributed nodes (e.g., literature graphs, product graphs).
The research team at Emory University proposed the first natural language explanation generator for graph neural networks, GraphNarrator, which for the first time generates high-quality natural language explanations from the input and output of GNNs, transforming graph neural networks from "black-box models" into "decision-making entities with justifications."
Paper link: https://arxiv.org/pdf/2410.15268
Code link: https://github.com/pb0316/GraphNarrator
GraphNarrator focuses on an important type of graph called Text-Attributed Graphs (TAGs), where node features are natural language texts (e.g., paper abstracts, product descriptions, disease descriptions).
The contributions of the paper include:
Proposing the first natural language explanation framework, extending the explanation of TAG graphs from the structural level to the language level;
Unifying structured and language information, bridging graph structure reasoning and the understanding ability of large language models (LLMs);
Open-sourcing a toolchain that provides a high-quality pseudo-label constructor and a self-supervised distillation method, facilitating migration to any GNN task.
The first author of the paper is Bo Pan, a Ph.D. student at Emory University, who has long been engaged in research in the fields of graph learning and explainable artificial intelligence.
The co-first authors are Zhen Xiong, a master's student at USC, and Guanchen Wu, a Ph.D. student at Emory University. The corresponding author is Liang Zhao, an associate professor in the Department of Computer Science at Emory University.
This research was accepted by the main conference of ACL 2025, proposing the first natural language explanation generator for graph neural networks, GraphNarrator.
Let GNNs Speak
GraphNarrator generally consists of three steps:
1. Constructing Explanation Pseudo-labels (Pseudo-label Generation)
Use saliency-based explanation methods to extract "important texts + key neighbor nodes" in the form of the importance of each feature (node, edge, token).
Convert these structures into structured prompts and input them into the GPT model together with the question and prediction to generate interpretable pseudo-labels.
2. Optimizing Pseudo-labels (Filtering via Expert-Designed Criteria)
Filter higher-quality pseudo-labels through two major criteria:
Faithfulness: Consistent with the model prediction. The researchers calculate the faithfulness between the generated textual explanations and the input and output through mutual information.
Conciseness: Information is condensed and highly readable, and shorter lengths are encouraged.
GraphNarrator optimizes these two objectives simultaneously through expert iteration to ensure that the teacher model generates high-quality explanations.
3. Distilling the Explainer (Training Final Explainer)
Distill the pseudo-labels into an end-to-end model (LlaMA 3.1 8B is used in the paper). By directly inputting the graph structure and text, the explanation statements can be automatically output.
Faithful, Readable, and More Appealing to Users!
Datasets
The researchers systematically evaluated GraphNarrator on multiple real-world Text-Attributed Graph (TAG) datasets, including:
Cora: A paper citation graph where nodes are papers and texts are abstracts.
DBLP: An author collaboration graph where texts are paper lists.
PubMed: A biomedical literature graph.
Comparison methods:
Zero-shot explanation generation by mainstream LLMs (LLaMA 3.1-8B, GPT‑3.5, GPT‑4o)
SMV: A saliency explanation template conversion method based on GPT‑4o
GraphNarrator (based on LLaMA 3.1-8B)
The evaluation objective is to test whether the natural language explanations generated by GraphNarrator are faithful, accurate, readable, and appealing to users.
Evaluation Results
The researchers evaluated the quality of the explanations generated by this method through both automatic and manual methods.
In the automatic evaluation, GraphNarrator led comprehensively in Simulatability (+8 - 10%), proving that the explanation content highly restored the GNN predictions;
The PMI - 10% coverage increased significantly (an average of +8.2%), indicating that it can capture the most important tokens; The Brevity (explanation length / input length) decreased by more than 13%, verifying its "concise and to the point" ability.
In the manual evaluation, reviewers with a computational linguistics background scored from four aspects: readability, insight, structural information, and semantic information (on a scale of 1 - 7).
The results show that it outperformed GPT‑4o and SMV in all aspects, especially with a significant advantage in structural understanding (+33%). The explanations are more fluent and logically clear, gaining higher trust from real users.
Reference materials:
https://arxiv.org/pdf/2410.15268
This article is from the WeChat official account "New Intelligence Yuan", edited by LRST. It is published by 36Kr with authorization.