[2N6-GS-2x] Machine learning

5:30 PM - 5:45 PM JST(8:30 AM - 8:45 AM UTC)

[2N6-GS-2x-01] Generative Modeling of Graph-Structured Data via Latent Flow Matching

〇Sho Kuno¹, Kengo Nakajima^2,3 (1. Graduate School of Information Science and Technology, The University of Tokyo, 2. Information Technology Center, The University of Tokyo, 3. RIKEN R-CCS)

[[online]]

Keywords:

Non-Euclidean Generative Modeling,Flow Matching

Graph foundation models require generators that are permutation-consistent, scalable, and useful beyond unconditional sampling. Latent Graph Diffusion (LGD) (Zhou et al., NeurIPS’24) maps graphs to a Euclidean latent space and unifies generation and supervised prediction via conditional completion, but its iterative diffusion sampler is slow. We propose \emph{Latent Graph Flow Matching} (LGFM), replacing LGD's latent diffusion sampler with continuous-time flow matching and enabling sampling via a single ODE solve in latent graph space. To our knowledge, LGFM is the first approach to bring LGD's unified conditional-generation interface to a flow-matching formulation. On QM9 unconditional molecule generation, LGFM achieves $\sim$94\% validity with competitive distributional metrics while improving throughput by $7.8\times$ (35 graphs/s vs.\ 4.5 for LGD-small). We further demonstrate the same conditional-generation interface on molecular property regression by generating masked graph-level labels.