Graph Diffusion Model

Graph diffusion models are a class of generative models designed to create new graphs—networks of interconnected nodes—by iteratively adding or removing noise from an initial graph structure. Current research focuses on improving these models' ability to incorporate domain-specific constraints (e.g., planarity, acyclicity), handle large graphs efficiently, and ensure permutation invariance, leading to the development of architectures like autoregressive diffusion models and those leveraging graph neural networks. These advancements are impacting diverse fields, enabling applications such as predicting brain connectivity, generating realistic molecular structures, and solving complex reassembly problems in 2D and 3D.