Towards Automated Evaluation of Explanations in Graph Neural Networks

Explanation quality is hard to measure objectively: The paper identifies that current methods for evaluating whether GNN explanations are actually good lack standardization, making it difficult to ensure explanations are trustworthy before deploying AI systems to real users.

User-centered evaluation matters more than technical metrics: The authors emphasize that explanations should be evaluated based on how actual end users understand and use them, rather than only through computational measures—a shift that could improve real-world AI accountability.

Graph Neural Networks need better explanation tools: GNNs are increasingly used in applications affecting people's lives, but explaining their decisions remains "an unsolved problem," highlighting a gap between AI capability and AI transparency that needs urgent attention.

Automated evaluation could scale responsible AI practices: By developing automatic methods to evaluate explanations, organizations could more systematically check whether their AI systems provide understandable justifications before deployment, supporting broader responsible AI adoption.

Real-world experience reveals current limitations: The authors draw on practical experience with actual applications to show that existing explanation approaches fall short of what users need, suggesting that responsible AI requires grounding in genuine use cases rather than theoretical assumptions.