Relevant Walk Search for Explaining Graph Neural Networks

Researchers introduced polynomial-time algorithms to identify the top-K most relevant walks in graph neural network explanations.

The approach replaces the exponential search used in GNN-LRP with a scalable max-product-based method.

It showed strong results on epidemiology, molecular, and natural language benchmarks.

The advance makes higher-order GNN interpretability practical for larger real-world graphs and more robust AI systems.