Core Concepts
Explaining node predictions in GNNs using simple surrogates can improve interpretability and performance.
Abstract
Distill n’ Explain (DnX) proposes a method to explain graph neural networks by distilling knowledge into a simpler surrogate model. The surrogate model is then used to extract node-level explanations efficiently. DnX and its faster version, FastDnX, outperform existing GNN explainers while being significantly faster. The method leverages the linear nature of the surrogate model to speed up the explanation process. Experimental results show superior performance and speed compared to state-of-the-art methods across various benchmarks.
Stats
DnX learns a surrogate GNN via knowledge distillation.
DnX and FastDnX often outperform state-of-the-art GNN explainers.
FastDnX presents a speedup of up to 65K× over GNNExplainer.
Quotes
"Explanations for GNNs have recently gained interest due to their lack of interpretability."
"Distill n’ Explain proposes a new framework for explaining GNNs using simple surrogates."