The knowledge behind GNNs is not yet widely accessible because this field has been developing quickly. Currently, the GNN theory is dispersed online through research papers, articles, or blogs. Therefore, a GNN book or a collection of online resources would be helpful for people looking to enter this field.

Graphs are helpful in a wide range of fields, including social networks, physics systems, chemistry, and medicine. This concept piques the interest of deep learning researchers in the graph data structure. Furthermore, large corporations such as Twitter, Google, and Facebook fund GNN research because it outperforms competing machine learning models for graph data.

Let's look at some of the tools and resources available to help you learn more about GNNs.

Resources

Stanford Course Notes — Machine Learning with Graphs

It is a course at Stanford University that focuses on graph-based machine learning. It also includes reading suggestions and public-accessible slides from their lectures. If you want to take a structured course in an organized manner, this is the course for you.

Network Science Book by Albert-László Barabási

This book is the result of extensive collaboration, which shaped everything from the content to the visualizations and interactive tools. Although this book isn't about GNNs, it's a great place to start if you're new to graphs.

Graph Representation Learning Book by William L. Hamilton

This book's pre-publication copies are available online. It provides a concise but comprehensive overview of graph representation learning, including deep generative models of graphs, graph neural networks, and graph data embedding techniques. Almost all of the theories required for graph neural networks are present.

Graph Neural Network Papers With Code

Paper With Code (PwC) is the best place to look for Graph Neural Network models with code implementation that you can use. This website streamlines the availability of technical papers. It has expanded significantly over the last few years. The growth of publicly accessible datasets and modern research has converged toward complete transparency and credibility. As a result, PwC's website has also undergone continuous improvement. By browsing, you can quickly navigate state of the art, either by task or method (e.g. attention, transformers).

Libraries

PyTorch Geometric

PyTorch Geometric (PyG), a Python library, is used for deep learning on atypical structures such as graphs. The project was created and published by Matthias Fey and Jan E. Lenssen, two PhD students at TU Dortmund University. It contains several recently published methods in relational learning and 3D data processing, as well as general graph data structures and processing techniques. PyTorch Geometric achieves high data throughput by utilizing sparse GPU acceleration, providing specialized CUDA kernels, and implementing effective mini-batch handling for input examples of varying sizes.

Deep Graph Library (DGL)

The Deep Graph Library is another easy-to-use, powerful, and scalable Python library for deep learning on graphs (DGL). The Distributed Deep Machine Learning Community, a group of deep learning enthusiasts, created it. Its API is extremely concise and clear. In addition, DGL adds a useful higher-level abstraction that allows for auto-batching.

Graph Nets

Graph Nets is DeepMind's library for building graph networks in Tensorflow and Sonnet. TensorFlow's CPU and GPU versions are both compatible with the library. It allows you to add Temporal Graphs and implement nearly any existing GNN with just six core functions. Unfortunately, graph Nets require TensorFlow 1, so it feels dated despite only being about three years old.

Spektral

Spektral is an open-source Python graph deep learning library built on the Keras API and TensorFlow 2. The primary goal of this library is to provide a simple, adaptable framework for developing GNNs. Spektral can be for any task in which graphs describe data, such as classifying social network users, predicting molecular properties, creating new graphs with GANs, clustering nodes, predicting links, and more. It uses some of the most popular graph deep learning layer implementations. According to Keras principles, this library is to be extremely simple for beginners while still retaining flexibility for experts. Compared to other libraries like DGL and PyG, the slow training speeds for most tasks are an unfortunate trade-off for Spektral's ease of use.