Variational Bayesian Flow Network for Graph Generation

Graph generation aims to sample discrete node and edge attributes while satisfying coupled structural constraints. Diffusion models for graphs often adopt largely factorized forward-noising, and many flow-matching methods start from factorized reference noise and coordinate-wise interpolation, so node-edge coupling is not encoded by the generative geometry and must be recovered implicitly by the core network, which can be brittle after discrete decoding. Bayesian Flow Networks (BFNs) evolve distribution parameters and naturally support discrete generation. But classical BFNs typically rely on factorized beliefs and independent channels, which limit geometric evidence fusion. We propose Variational Bayesian Flow Network (VBFN), which performs a variational lifting to a tractable joint Gaussian variational belief family governed by structured precisions. Each Bayesian update reduces to solving a symmetric positive definite linear system, enabling coupled node and edge updates within a single fusion step. We construct sample-agnostic sparse precisions from a representation-induced dependency graph, thereby avoiding label leakage while enforcing node-edge consistency. On synthetic and molecular graph datasets, VBFN improves fidelity and diversity, and surpasses baseline methods.

💡 Research Summary

The paper introduces the Variational Bayesian Flow Network (VBFN), a novel framework for graph generation that overcomes key limitations of existing Bayesian Flow Networks (BFNs) and diffusion‑based models. Traditional BFNs assume a factorized Gaussian belief over the latent graph signal, which forces Bayesian updates to be performed element‑wise. This independence geometry fails to capture the intrinsic coupling between node attributes and edge attributes, leading to brittle discrete decoding and violations of structural constraints in generated graphs.

VBFN lifts the factorized belief family to a tractable joint Gaussian variational family whose uncertainty is governed by a structured precision matrix. Concretely, the latent graph signal (z\in\mathbb{R}^D) (concatenating node features (X) and edge features (A)) is endowed with a belief (q_t(z)=\mathcal{N}(z;\theta_t,P_t^{-1})) where the precision (P_t) evolves over continuous time. The precision is the sum of two components: a prior precision (\Omega_{\text{prior}}) that encodes graph‑structural relationships, and a time‑dependent observation precision (\beta(t)\Omega_{\text{obs}}) that shapes the noise injected by the sender channel.

Both (\Omega_{\text{prior}}) and (\Omega_{\text{obs}}) are constructed from a sample‑agnostic dependency graph (H). Vertices of (H) correspond to individual entries of node and edge tensors; edges connect node‑edge pairs, edge‑edge pairs, and symmetric edge pairs (if the adjacency is undirected). Edge weights are set by hyper‑parameters (\lambda_X) and (\lambda_A). From this weighted adjacency a masked combinatorial Laplacian (L) is built, and the prior precision is defined as (\Omega_{\text{prior}} = M L M + \epsilon I), where (M) masks out entries that are not present in a particular graph and (\epsilon>0) guarantees positive definiteness. Crucially, (H) never queries the actual adjacency values, thus avoiding label leakage while still enforcing basic node‑edge consistency.

The sender transmits a noisy message (y) according to (p_S(y|z;t)=\mathcal{N}(y;z,