Graph-Aware Generative Diffusion (GAD)

• Graph-Aware Generative Diffusion (GAD) is a model that integrates the graph Laplacian into a noise-diffusion process to generate graph signals that respect the underlying topology.
• It employs a heat equation-based noise schedule that progressively smooths signals, resulting in a stationary Gaussian Markov random field distribution.
• The backward process is framed as graph-signal denoising using GCNN-based estimators, yielding superior performance compared to standard diffusion baselines.

A graph-aware generative diffusion model (GAD) is a class of generative models that produce graph signals by embedding important structural information of the graph directly within the generative diffusion process. Unlike graph-agnostic diffusion approaches, GAD incorporates the graph Laplacian and solves the forward process as a heat equation perturbed by a learnable noise schedule. This results in a forward dynamics with stationary distribution given by a Gaussian Markov random field whose covariance is parameterized by the topology of the underlying graph, and a backward process that operates as a series of graph-signal denoising problems. The GAD framework is specifically designed to generate realistic signals defined on arbitrary graphs, such as sensor networks or traffic systems, and outperforms standard variance-preserving or variance-exploding models in these settings.

1. Graph-Aware Diffusion Process: Heat Equation Formulation

At the core of GAD is the explicit use of the graph Laplacian in the stochastic differential equation driving the forward process. Rather than introducing independent Gaussian noise, the model evolves the signal according to the heat equation on the graph: $$dx_t = -c_t (L + \gamma I) x_t\, dt + \sqrt{2c_t}\, \sigma\, dB_t$$ where $L$ is the graph Laplacian, $\gamma > 0$ ensures positive definiteness, $c_t$ is a time-dependent drift coefficient, $\sigma$ is the noise level, and $B_t$ is standard Brownian motion. The innovative feature is the introduction of a "floor-constrained polynomial scheduler," with $c_t = c_{\text{min}} + k u^\alpha$ and $u = t/T$, controlling the rate of spectral decay.

This heat equation formulation ensures that the forward process acts as a graph low-pass filter, progressively smoothing the signal according to the spectral decomposition of $L + \gamma I$: $$\Phi_t = \exp(-\bar{c}_t (L + \gamma I)),\quad \bar{c}_t = \int_0^t c_s\, ds$$ Hence, the noising phase is not isotropic but respects the correlations imposed by graph connectivity.

2. Stationary Distribution and Theoretical Analysis

The forward SDE admits closed-form solutions for the mean and covariance of the signal at any $L$0: $L$1

$L$2

As $L$3 and $L$4, the process converges to a stationary measure that is a Gaussian Markov random field (GMRF) with covariance $L$5. This analytically demonstrates that the generative process has a topology-aware stationary distribution, unlike standard diffusion models that ignore the graph structure.

3. Backward Process as Graph-Signal Denoising

The backward SDE is formulated as: $L$6 The score function $L$7 is not directly available; it is estimated using Tweedie’s formula in the context of the forward heat equation: $L$8 This means the reverse process can be interpreted as a sequence of graph denoising problems, where at each timestep, the model must "deconvolve" the noise added by graph-based smoothing.

To implement $L$9, a graph convolutional neural network (GCNN) with spectral filtering layers is employed, explicitly matching the graph structure to the learning process: $\gamma > 0$0

4. Practical Implementation and Closed-Form Estimators

The MMSE estimator for the denoising at a fixed $\gamma > 0$1 is given in closed form: $\gamma > 0$2 where

$\gamma > 0$3

This ARMA-type (autoregressive moving average) graph filter provides explicit control over the degree of spectral noise removal and high-frequency signal retention.

5. Empirical Results and Benchmarking

The GAD model was validated on multiple datasets:

• Synthetic SBM graph: Signals constructed with community-specific means, low-pass filtered per graph spectrum, showed that GAD yields lower sample discrepancies (average MMD) than variance-preserving and variance-exploding baselines.
• METR-LA traffic sensor network: GAD more effectively recovers spatial correlations with computationally efficient generation (few time steps), outperforming graph-agnostic baselines.
• Molene temperature sensor graph: GAD generates spatially realistic signals matching network-induced dependencies. In all cases, leveraging the graph Laplacian within the forward noise process and interpreting the reverse dynamics as denoising yields superior signal fidelity compared to conventional methods.

6. Context, Significance, and Future Directions

Injecting graph structure into diffusion modeling via the heat equation and time-warped scheduler directly ties the stationary generative distribution to the graph’s topology, producing GMRF-correlated outputs. This connection to established graph signal processing (GSP) methods enhances interpretability and algorithmic control.

Potential future research directions include:

• Design of more refined time schedules, possibly varying per Laplacian eigenvalue, to control noise injection across spectral components.
• Use of more advanced SDE integration schemes beyond Euler–Maruyama for higher-quality signal generation.
• Extension to dynamic and heterogeneous graphs, or domain-specialized applications (e.g., recommender systems, sensor arrays) that demand topology-aware generative signals.

This approach demonstrates a principled path for graph-aware generative modeling that tightly integrates graph theory, spectral analysis, and diffusion-based deep learning.