Graph Neural Network explanations lack canonicality due to input symmetries

A new research paper explores a fundamental issue in explaining Graph Neural Networks (GNNs), specifically how their explanations can be non-canonical due to input symmetries. The study highlights that gradient-based GNN explainers assign identical attribution scores to chemically equivalent atoms, but the final reported explanation, such as the top-k edges, can arbitrarily favor one over the other based on array ordering. This arbitrariness is a structural obstruction, particularly problematic for datasets like Mutagenicity where such symmetries are common. AI

IMPACT Highlights a fundamental limitation in the interpretability of graph neural networks, potentially impacting trust and reliability in their explanations for molecular analysis.

RANK_REASON Academic paper on graph neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• graph neural networks
• Lean 4 Programming Language
• Mutag Dataset

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →