Generating graphs with graph neural networks
Deep graph generative models are GNN-based architectures that are more expressive than traditional techniques. However, it comes at a cost: they are often too complex to be analyzed and understood, like classical methods. We list three main families of architecture for deep graph generation: VAEs, GANs, and autoregressive models. Other techniques exist, such as normalizing flows or diffusion models, but they are less popular and mature than these three.
This section will describe how to use VAEs, GANs, and autoregressive models to generate graphs.
Graph variational autoencoders
As seen in the last chapter, VAEs can be used to approximate an adjacency matrix. The Graph Variational Autoencoder (GVAE) model we saw has two components: an encoder and a decoder. The encoder uses two GCNs that share their first layer to learn the mean and the variance of each latent normal distribution. The decoder then samples the learned distributions...
