GraphXAIN: Narratives to Explain Graph Neural Networks

Natural language explanations make AI more accessible: GraphXAIN converts technical GNN outputs (subgraphs and feature scores) into story-like narratives that non-experts can understand, addressing a key barrier to responsible AI—the ability for stakeholders to comprehend how automated decisions are made.

Significant improvements in user understanding and trust: A survey of ML researchers and practitioners found that GraphXAIN enhanced four critical explainability dimensions (understandability, satisfaction, convincingness, and suitability), with 95% of participants rating it as valuable, suggesting narrative explanations are more effective than traditional technical outputs.

Works across different AI systems: GraphXAIN is "model- and explainer-agnostic," meaning it can be applied to various graph neural networks and explanation methods without requiring redesign, making responsible AI practices more scalable and practical to implement.

Combines multiple AI techniques for better explanations: By using Large Language Models to translate technical explanations into narratives, the method demonstrates how different AI tools can work together to improve transparency—an important consideration for responsible AI development.

Builds trustworthiness through clarity: When combined with other explanation methods, GraphXAIN further improved trustworthiness, insightfulness, and user confidence, indicating that clearer communication about how AI systems work directly supports more responsible deployment and user confidence.