Shedding light on graph neural networks

This repository contains the code and slides for the "Shedding Light on Obscure Graph Deep Learning" workshop presented by Visium at AMLD 2021.

Abstract

Deep learning techniques on graphs have achieved impressive results in recent years. Graph neural networks (GNN) combine node feature information with the graph structure in order to make their predictions. However, this strategy results in complex models, whose predictions can be hard to interpret.

In this workshop, you will get familiar with graph data and you will learn the fundamental concepts behind graph neural networks. After building and training your own GNN, you will be introduced to GNNExplainer, a model-agnostic framework for interpreting GNN results. Thus, you will be able to visualize explanations for your GNN's predictions, giving a sense to the model outputs.

Workshop setup

Running the code of this repository is, for the most part, computationally expensive and it is best to have GPU access. For the workshop users, we have prepared the notebook so that you can easily load it into Google Colab and enjoy the free GPU services offered by Google.

1. Open Google Colab and sign in with your Google account or create a new one.

2. Click on File and then Open notebook.... Select the Github tab, look for VisiumCH to find this repository and select the workshop_notebook.ipynb.

3. Once you are in the notebook, click on Runtime, Change runtime type and then select GPU as hardware accelerator.

4. Finally, click on Connect and you should be ready to go!

Requirements

For users outside the workshop who would like to experiment more with the repo outside of colab, here are the installation steps:

Create a virtual environnement and install the dependencies:

python3 -m venv env
source env/bin/activate
xargs -a requirements.txt -L 1 pip install

Basic usage

The workshop_notebook.ipynb contains the theory and the code to learn about Graph Neural Networks and explainability with an hands on approach. It contains some exercises which you shall complete in order to advance in the workshop.

Solutions are provided in workshop_notebook_sol.ipynb. This notebook can be run from top to bottom.

We provide some additional code resources in the src module, which is installed by the requirements files. It implements a few more advanced methods which are used through the notebooks.

Citations

The GNN Explainer algorithm we present has been developed by Ying et al. If you intend to use it for anything, please consider citing the original authors in your work:

@misc{ying2019gnnexplainer,
  title={GNNExplainer: Generating Explanations for Graph Neural Networks},
  author={Rex Ying and Dylan Bourgeois and Jiaxuan You and Marinka Zitnik and Jure Leskovec},
  year={2019},
  eprint={1903.03894},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}

The code in this repo is adapted and corrected for the AMLD workshop. Our implementation of the explainer is based on PyTorch Geometric.

We also borrowed the notebook structure from PyTorch Geometric tutorials and inspired our slides on the very good seminar from Petar Veličković, from which we got some figures.