Relevant Walk Search for Explaining Graph Neural Networks
Researchers have developed new polynomial-time algorithms to address the exponential computational complexity of identifying relevant walks in Graph Neural Networks (GNNs). This advancement significantly improves the applicability of GNN-LRP, a method for explaining GNNs by analyzing information flows through walks. The proposed algorithms, based on the max-product method, enable exact and approximate identification of top-K relevant walks, demonstrating effectiveness across various benchmarks including epidemiology, molecular, and natural language tasks. AI
IMPACT Enhances the interpretability and trustworthiness of GNNs, potentially increasing their adoption in safety-critical applications.