Score-based Generative Modeling of Graphs via the System of Stochastic Differential Equations

Explore a comprehensive lecture on score-based generative modeling of graphs using stochastic differential equations. Delve into the challenges of graph generation and learn about a novel approach that overcomes limitations of previous methods. Examine the proposed graph diffusion process, which models joint distribution of nodes and edges through a system of SDEs. Discover new score matching objectives and an efficient solver for sampling from the reverse diffusion process. Analyze the model's performance on diverse datasets, including molecule generation. Investigate the forward and reverse diffusion processes, various GDSS variants, and the design of score-based models. Conclude with a discussion on model limitations and potential future directions in this cutting-edge field of graph generation.

- Intro & Overview
- Challenges of Graph Generation
- Our Approach
- Graph Diffusion via the System of SDEs GDSS
- Forward Diffusion Process of GDSS
- GDSS Variants
- Designing Score-based Models
- Reverse Diffusion Process
- Model Results: Generic Graph and Molecule Generation
- Model Limitations
- Future Directions + Conclusions