DiGress: Discrete Denoising Diffusion for Graph Generation

Explore a comprehensive lecture on DiGress, a discrete denoising diffusion model for graph generation with categorical node and edge attributes. Delve into the intricacies of this innovative approach, which defines a diffusion process for progressive graph editing and employs a graph transformer network to revert the process. Learn about the Markovian noise model that preserves marginal distribution of node and edge types during diffusion, and the incorporation of auxiliary graph-theoretic features. Discover how DiGress achieves state-of-the-art performance on molecular and non-molecular datasets, including its scalability to the large GuacaMol dataset. Gain insights into the challenges, training methods, and properties of DiGress, as well as its potential for future developments in graph generation.

- Intro
- Denoising Diffusion Models
- The Price of Efficiency
- Discrete Diffusion
- DiGress: Discrete Graph Denoting Diffusion
- DiGress: Challenges and Training Methods
- DiGress: Denoising Network
- DiGress: Properties
- Results
- Improving DiGress with Marginal Transitions and Structural Features
- Final Results
- Summary and What’s Next
- Q+A