AIMing for Standardised Explainability Evaluation in GNNs: A Framework and Case Study on Graph Kernel Networks

Explainability evaluation for AI systems lacks standardization: The paper identifies that existing frameworks for evaluating how well AI models explain their decisions are fragmented and inconsistent, making it difficult to compare different approaches—a critical gap for responsible AI deployment.

A new comprehensive evaluation framework (AIM) measures three key dimensions: The framework assesses Accuracy (whether predictions are correct), Instance-level explanations (why specific decisions were made), and Model-level explanations (how the overall system works), providing a more complete picture of explainability than previous single-focus approaches.

Inherently interpretable models can be systematically improved: By applying AIM to graph kernel networks, the researchers developed an updated version (xGKN) that maintained prediction accuracy while significantly improving explainability—demonstrating that transparency and performance need not be trade-offs.

Standardized evaluation enables better model comparison: The flexible design of AIM allows fair comparison of explanations across different AI models, addressing a practical challenge for organizations trying to select which AI systems to deploy responsibly.

Practical tools for XAI are advancing: This work contributes concrete methodology that can help practitioners evaluate whether their AI systems are genuinely explainable to stakeholders, moving beyond theoretical discussions toward actionable standards for responsible AI development.