Generating Graphs Using Graph Neural Networks
Graph generation consists of finding methods to create new graphs. As a field of study, it provides insights into understanding how graphs work and evolve. It also has direct applications in data augmentation, anomaly detection, drug discovery, and so on. We can distinguish two types of generation: realistic graph generation, which imitates a given graph (for example, in data augmentation), and goal-directed graph generation, which creates graphs that optimize a specific metric (for instance, in molecule generation).
In this chapter, we will explore traditional techniques to understand how graph generation works. We will focus on two popular algorithms: the Erdős–Rényi and the small-world models. They present interesting properties but also issues that motivate the need for GNN-based graph generation. In the second section, we will describe three families of solutions: variational autoencoder (VAE)-based, autoregressive...
