A Meta-Learning Approach for Training Explainable Graph Neural Networks

Authors: Indro Spinelli, Simone Scardapane, Aurelio Uncini

This repotisory contains the code of MATE (MetA-Train to Explain), a meta-learning framework for improving the level of explainability of a Graph Neural Network at training time. Our approach steers the optimization procedure towards more interpretable minima meanwhile optimizing for the original classification task. Here there is the preprint.

The code is build upon the repository of Re: Parameterized Explainer for Graph Neural Networks and we thanks the authors (Maarten Boon, Stijn Henckens, Lars Holdijk and Lysander de Jong) for making their code accessible to everyone.

IPython Notebooks

• experiment_model_training: Meta-trains models with MATE algorithm.
• experiment_replication:Evaluate model's explainability.

Installation

Install required packages using pip install -r requirements.txt additionally follow the instructions in order to install PyTorch Geometric.