Reproduce “Score-based Generative Modeling of Graphs via the System of Stochastic Differential Equations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jaehyeong Jo, Seul Lee, Sung Ju Hwang
Researchers at KAIST developed Graph Diffusion via the System of Stochastic Differential Equations (GDSS), a continuous-time score-based model that jointly generates node features and adjacency matrices for graphs. GDSS achieves high chemical validity (95.72% on QM9) and distribution fidelity in molecule generation, along with substantial speed improvements over existing methods.
Key insights
- Continuous-time diffusion models can be effectively adapted for graph generation by formulating the process as a system of SDEs that jointly and dependently model node features and adjacency matrices.
- Explicitly modeling the complex interdependencies between node features and graph structure via this system of SDEs is crucial for generating high-fidelity and chemically valid graphs.
- The proposed system of SDEs allows for more efficient and effective learning of score functions, despite addressing a joint distribution, compared to models with less complete dependency modeling.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.