Embedded Graph Flows for Categorical Graph Generation
EGF replaces one-hot graph categories with learned continuous endpoints, then maps them back to discrete nodes and edges.
The paper says this avoids treating every category as equally distant from every other category. Its permutation-equivariant graph transformer moves Gaussian noise toward learned node and unordered-edge embeddings. On QM9, EGF leads the three reported methods on all four metrics, including FCD 0.150 versus 0.717 for DiGress and 0.812 for GruM. On ZINC250k, it reports the lowest NSPDK MMD, pointing to closer local molecular substructures. ArXiv · AI/CL/LG's note
The paper says this avoids treating every category as equally distant from every other category. Its permutation-equivariant graph transformer moves Gaussian noise toward learned node and unordered-edge embeddings. On QM9, EGF leads the three reported methods on all four metrics, including FCD 0.150 versus 0.717 for DiGress and 0.812 for GruM. On ZINC250k, it reports the lowest NSPDK MMD, pointing to closer local molecular substructures. ArXiv · AI/CL/LG's note
score 4