Types of Graph Neural Networks

We have different types of graph neural network architectures. Some of the popular ones include the following:

• Graph convolutional neural networks

• Graph attention networks

• Recurrent graph neural networks

Let's go through them one by one.

Graph convolutional neural networks

Graph convolutional neural networks (GCNs) are a variety of convolutional neural networks that work on graphs. We can broadly divide GCN models into spectral-based and spatial-based graph convolutional networks.

Spectral-based networks

The foundation of spectral-based methods comes from graph signal processing. The core idea behind the spectral networks is to use the spectrum of the graph, i.e., the eigenvalues and eigenvectors, to perform convolution operations on the graph.

We define a set of basis functions that captures the graph structure well. These basis functions are used as filters (similar to what we see in a traditional convolutional network).

The neural network model has multiple layers which perform the graph convolution operations. The convolution operation is as follows: