Introduction to graph neural networks

This is GitHub repository that is aimed to help newcomers into Graph NN to understand basics of graph convolution method, how classification on graphs work and what are the common practices in this sphere.

This tutorial contains information how to implement graph convolution on raw numpy, but it is worth mentioning that this tutorial can be extended (it already has all code, we made an open-source pytorch-like library that is written on numpy) to even be used as introduction to linear models and how forward/backward passes works.

How to run

The only file you need to run is Tutorial.ipynb. But before running please make sure you have networkx, matplotlib and numpy installed. If no, you can run next command bellow:

pip install -r requirement.txt

Worth mentioning that you may require inno-graph library if you will use tutorial independently (outside this repository).

To run notebook in jupyter-notebook environment you can use any IDE, service, or even from terminal with jupyter-notebook command.

Some insides into tutorial

We used Karate-club dataset, however it was not our first iteration. From the beginning we decided to use IMDB-BINARY dataset. The main problem we faced is that it's not that simple as we thought it was, it was hard to understand (in deprecated version we dedicated notebook to fully-explain it). We also were sceptic that it was simple to train and our Pytorch-implementation proved that. For such reasons we decided to switch to Karate-club.

In addition, we wanted to provide in-depth tutorial with examples, so we decided to implement full GCN-network by only using numpy. All interested people would have a chance to implement all layers by hand, or import some from our library. Flexibility - our strong part!

Repository structure

GCN-tutorial
│   README.md       # Introduction to project (this file)
│   Tutorial.ipynb  # Main tutorial file we worked on
│
└───igraph          # Pytorch-styled library we developed
│   │   layers
│   │   losses
│   │   models
│   │   preprocessing
│   
└───examples        # Examples how to use our library
    │   train_test_example.py
    

Other files - legacy / important for other reasons

Credits

This tutorial was created by:

• Polina Zelenskaya - responsible for library design, layers and tutorial
• Lev Rekhlov - responsible for library testing, models and tutorial
• Said Kamalov - responsible for library losses, preprocessing and tutorial