Diffusion scores learn the graph and the filter at once: TopoScore-FAF turns residual scores into topology gradients without transition densities
Synopsis
Inferring a network's interaction graph together with its hidden states is hard when observations are intermittent and transition noise is unknown and non-Gaussian, because an unconstrained generative model can absorb graph structure into its own mean. TopoScore-FAF restricts the graph to the conditional mean of a leaky diffusion and lets a frozen diffusion emulator represent only the centered residual, then projects the residual score onto edge directions through a score-topology duality, so topology gradients need neither transition densities nor backpropagation through the sampler. Experiments on synthetic and four real datasets show the combination improves topology recovery and probabilistic calibration under non-Gaussian forcing.
Interpretation
Topology inference and state filtering are placed in one closed loop: the graph sets how states propagate, and the estimated states refine the graph in turn. Earlier graph-learning methods relied on specified transition models or second-order statistics, while generative filters assumed the transition mechanism was known; joining them directly lets an unconstrained generative model obscure the graph-dependent dynamics. Under partial observation with switching-forcing residuals, TopoScore-FAF attains lower relative Laplacian error on ER graphs than the strongest second-order baseline, GKNet, and a higher edge F1 score, and the improvement holds across ER, BA and random geometric graph families.
The construction separates structural dynamics from stochastic forcing: the graph carries only the conditional transition mean, and a frozen diffusion emulator represents only the centered residual law. If the residual is not centered, the emulator absorbs topology error into its learned mean, and the graph becomes unidentifiable even from fully observed states. In the ablation, an emulator trained without detrending leaves state-estimation metrics nearly unchanged but raises topology error markedly, so good state estimation does not certify a correct graph.
Projecting the residual score onto edge-excitation directions yields the topology likelihood gradient, so updating the graph needs only forward denoiser calls. A diffusion emulator supplies scores or denoisers rather than tractable transition densities, which rules out standard likelihood-based estimation. The duality holds at finite noise and in the zero-noise limit, and the objective reduces to generalized least squares for Gaussian residuals; on real data, replacing the score with a Gaussian score drops topology accuracy back to the GKNet level.
Perspective
The demonstrated scope is static undirected graphs with known leakage coefficient, diffusion step and observation model, under intermittent masked observations; synthetic experiments cover ER, BA and random geometric graphs with five residual laws, and real data cover temperature, traffic, hydrology and resting-state fMRI networks.
The graph is assumed static, and time-varying graphs would need tracking analysis; the closed-loop theorem relies on compact dynamics and a filter-forgetting assumption that is not verified from primitives; emulator transfer across graphs is assumed, and its error enters the error floor and may violate the small-gain condition.
