Reproduce “Edge-Level Explanations for Graph Neural Networks by Extending Explainability Methods for Convolutional Neural Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tetsu Kasanishi, Xueting Wang, Toshihiko Yamasaki
Researchers from The University of Tokyo extended CNN explainability methods, including LIME, Saliency Maps, and Grad-CAM, to Graph Neural Networks for identifying important edges influencing model predictions. Their evaluation found LIME to be the most efficient and accurate approach for real-world graphs, demonstrating up to a 10x speed improvement over the direct removal method while maintaining superior explanation quality.
Key insights
- The architectural resemblance between graph convolution and 2-D convolution allows for successful adaptation of CNN explainability techniques to GNNs.
- LIME's ability to perturb multiple edges concurrently contributes to its superior performance by capturing interactions between graph connections.
- The effectiveness of explainability methods can be highly dependent on dataset properties, particularly the richness and importance of node features.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.