Embedded Graph Flows for Categorical Graph Generation
Researchers propose Embedded Graph Flows, a generative model with learned categorical embeddings that beats DiGress and GruM on molecular graph benchmarks.
Embedded Graph Flows (EGF) learns continuous embeddings for node and unordered-edge categories and transports Gaussian noise toward these endpoints using a permutation-equivariant graph transformer. On QM9 it achieves the best result on all four reported metrics, with a Fréchet ChemNet Distance of 0.150 versus 0.717 for DiGress and 0.812 for GruM. On ZINC250k it retains the lowest NSPDK MMD, indicating close agreement with local substructures of reference molecules. Code is released on GitHub.
- EGF replaces one-hot category encodings with learned continuous embeddings
- FCD of 0.150 on QM9 vs 0.717 (DiGress) and 0.812 (GruM)
- Lowest NSPDK MMD on ZINC250k among compared methods
- Uses a permutation-equivariant graph transformer for noise transport
- Products
- Embedded Graph Flows (EGF)
- Organizations
- Trusted-System-Lab
From the original article
1 min read at arxiv.orgGenerating categorical graphs requires choosing node and edge types that form a coherent structure without depending on node order. Many graph generators encode categories as fixed one-hot vectors, which can impose an artificial geometry in which categories are equidistant.
By Ethan Ma · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.