Faithful and Accurate Self-Attention Attribution for Message Passing Neural Networks via the Computation Tree Viewpoint

Attention weights alone don't tell the full story: The paper reveals that simply using attention weights as explanations for graph neural networks is misleading. Naive approaches to calculating what edges matter most can significantly misrepresent how the model actually makes decisions, which is critical for trustworthy AI systems.

A better method for understanding graph AI decisions: The researchers propose GATT, a more accurate way to trace which connections in a graph are truly important to a model's prediction by mapping the actual computation process. Testing on both synthetic and real-world datasets shows this method substantially improves the quality of explanations compared to existing approaches.

Explainability requires technical rigor: The paper demonstrates that responsible AI isn't just about providing explanations—it's about ensuring those explanations are faithful (accurately reflecting how the model works) and accurate (correctly identifying important features). Without this rigor, explanations can create false confidence in AI systems.

Graph neural networks need specialized explanation methods: Since graph-based AI systems (used in social networks, molecular analysis, recommendation systems, etc.) process information differently than other AI models, they require tailored approaches to explanation rather than borrowed techniques from other domains.