Towards Few-shot Self-explaining Graph Neural Networks

TL;DR

This work tackles the challenge of explainable graph neural networks in few-shot settings. It introduces MSE-GNN, a two-stage self-explaining architecture comprising an explainer and a predictor, guided by task information prototypes to enable rapid adaptation across tasks via MAML-inspired meta-training. The model employs graph augmentation and a combined loss with supervised, contrastive, and regularization terms to improve both predictive accuracy and explanation quality. Empirical results on four datasets show that MSE-GNN outperforms existing meta-learning and self-explaining baselines in few-shot graph classification while producing high-quality explanations, highlighting its potential for domains with limited labeled graph data. The approach offers practical impact for domains like chemistry and biomedicine where interpretability and data scarcity are critical.

Abstract

Recent advancements in Graph Neural Networks (GNNs) have spurred an upsurge of research dedicated to enhancing the explainability of GNNs, particularly in critical domains such as medicine. A promising approach is the self-explaining method, which outputs explanations along with predictions. However, existing self-explaining models require a large amount of training data, rendering them unavailable in few-shot scenarios. To address this challenge, in this paper, we propose a Meta-learned Self-Explaining GNN (MSE-GNN), a novel framework that generates explanations to support predictions in few-shot settings. MSE-GNN adopts a two-stage self-explaining structure, consisting of an explainer and a predictor. Specifically, the explainer first imitates the attention mechanism of humans to select the explanation subgraph, whereby attention is naturally paid to regions containing important characteristics. Subsequently, the predictor mimics the decision-making process, which makes predictions based on the generated explanation. Moreover, with a novel meta-training process and a designed mechanism that exploits task information, MSE-GNN can achieve remarkable performance on new few-shot tasks. Extensive experimental results on four datasets demonstrate that MSE-GNN can achieve superior performance on prediction tasks while generating high-quality explanations compared with existing methods. The code is publicly available at https://github.com/jypeng28/MSE-GNN.

Figures (6)

• Figure 1: Paradigm of "explainer-predictor" two-stage self-explaining models. The first part is composed of a explainer which selects an explanation subgraph for each input graph. The second part is a predictor which makes predictions based on the explanation subgraph. Given an input example from Synthetic datasetying2019gnnexplainer, explainer select as explanation, then predictor predicts $\hat{y}=house$ based on .
• Figure 2: Overall architecture of MSE-GNN. The model employs a "explainer-predictor" 2-stage self-explaining structure. The explainer selects explanation subgraphs for each input graph. The predictor mimics the decision-making process, which makes predictions solely based on the generated explanation.
• Figure 3: Raw figure of MNIST-sp and visualization of explanations generated by CAL(a), GREA(b) and MSE-GNN(c). Darker nodes indicate higher importance scores.
• Figure 4: Classification Performance and quality of explanation selected on Synthetic and OGBG-Molsider with different $\gamma$.
• Figure 5: Classification Performance and quality of explanation selected on Synthetic and OGBG-Molsider with different size of support sets.