Neural Graph Generator: Feature-Conditioned Graph Generation using Latent Diffusion Models

In recent years, the field of machine learning on graphs has witnessed an extensive growth, mainly due to the availability of large amounts of data represented as graphs. Indeed, graphs arise naturally in several application domains such as in social networks, in chemo-informatics and in bio-informatics. One of the most challenging tasks of machine learning on graphs is that of graph generation [Zhu et al., 2022]. Graph generation has attracted a lot of attention recently and its main objective is to create novel and realistic graphs. For instance, in chemo-informatics, graph generative models are employed to generate novel, realistic molecular graphs which also exhibit desired properties (e. g., high drug-likeness) [Jin et al., 2018, Zang and Wang, 2020]. Recently, there is a surge of interest in developing new graph generative models, and most of the proposed models typically fall into one of the following five families of models: (1) Auto-Regressive models; (2) Variational Autoencoders; (3) Generative Adversarial Networks; (4) Normalizing Flows; and (5) Diffusion models. These models can capture the complex structural and semantic information of graphs, but focus mainly on specific types of graphs such as molecules [Hoogeboom et al., 2022], proteins [Ingraham et al., 2019], computer programs [Brockschmidt et al., 2019] or patient trajectories [Nikolentzos et al., 2023]. Traditionally, in different application domains, there is a need for generating graphs that exhibit specific properties (e. g., degree distribution, node triangle participation, community structure, etc.).