Generative Diffusion Models of Stochastic Graph Signals

Abstract: Sampling stochastic signals supported on a graph underlies many graph machine learning tasks, including recommender systems, forecasting in financial markets, and wireless network optimization. In these settings, the target signals are realizations of unknown conditional distributions. However, prevailing approaches rely mostly on intricate, application-tailored designs that often regress to a conditional mean instead of sampling from the conditional law. This paper unifies such problems as conditional graph signal generative modeling and tackles them with a single denoising diffusion framework. We learn a reverse diffusion process, parametrized by graph neural networks (GNNs), that draws graph signals conditioned directly on the graph topology and on node-feature side information. The reverse process is realized by a novel architecture, the U-Graph Neural Network (U-GNN), which generalizes the image-convolutional U-Net to graph-structured signals. The U-GNN performs multi-resolution encoder--decoder processing in which pooling and unpooling reduce to a learned node selection, expressed by nested selection matrices, and a zero-padded lifting of coarse signals back to the full node set. The graph convolutions are carried out on the original graph, with a stride that sets their hop reach, so the U-GNN bypasses explicit graph coarsening at every resolution. We demonstrate our method on two generative tasks: stock price forecasting and optimal wireless resource allocation, with extensive numerical results in both domains.

• The paper introduces a unified conditional denoising diffusion framework with a novel U-GNN that samples true conditional distributions of graph signals.
• The methodology leverages a multi-resolution encoder-decoder with data-driven node selection to improve scalability and performance in applications like financial forecasting and wireless resource allocation.
• The model achieves notable improvements, reducing prediction errors by 21–28% and capturing realistic uncertainty and rare events critical for decision-making.

Generative Diffusion Models of Stochastic Graph Signals

Problem Formulation and Motivation

The paper "Generative Diffusion Models of Stochastic Graph Signals" (2607.06833) addresses a key gap in graph signal modeling: prevailing methods regress deterministic conditional means rather than sampling from true conditional distributions of graph signals, limiting their efficacy in domains where uncertainty and diversity are essential, such as recommender systems, financial forecasting, and wireless resource allocation. The authors pose the conditional graph signal generation problem as learning a parameterized generator that matches unknown conditional distributions of signals defined over fixed graphs, conditioned on topology and node features.

The fundamental requirement is an architecture that leverages inductive biases appropriate for graph-structured domains. While denoising diffusion models have excelled in regular domains (e.g., images), adapting them to graphs, which lack canonical pooling lattices, remains challenging. The paper proposes a unified conditional denoising diffusion framework for graph signal generation, parametrized by a novel U-shaped graph neural network architecture.

Architectural Design: U-Graph Neural Network (U-GNN)

Central to the proposed framework is the U-GNN, which generalizes the U-Net convolutional backbone to irregular graph-structured signals. Key innovations include:

• Multi-resolution encoder-decoder cascade: Node (un)pooling is realized via learned, data-driven selection matrices, retaining the most informative nodes at each hierarchy level. This avoids explicit graph coarsening typically present in graph pooling schemes.
• Lift-filter-reduce paradigm: Convolutions are always performed on the original graph topology. Pooling achieves dimensionality reduction through node selection and zero-padding, while a stride parameter sets the receptive field, analogous to dilated convolutions.
• Permutation equivariance and transferability: The architecture inherits these from its GNN layers, crucial for stability and adaptability across diverse graph topologies.
• Block-level conditioning and skip connections: Node states and diffusion step embeddings are globally fused into each block, augmenting the signal path and supporting conditional inference.

These mechanisms together achieve expressive generative modeling of signals conditioned on graph structure and node features, with improved interpretability and scalability over prior deterministic or application-specific architectures.

Diffusion Process

The generative process employs a forward Markov chain to incrementally corrupt clean graph signals with noise, driving them toward a tractable Gaussian prior, and a learned reverse process (parameterized by the U-GNN) to denoise and reconstruct samples from the data distribution. The reverse process is trained via a noise-prediction objective aligned with evidence lower bound maximization, and sampling uses DDIM updates for accelerated inference.

Conditioning is directly integrated by embedding node-feature side information and diffusion steps, enabling principled conditional generation. Model evaluations are accelerated by subsampling denoising steps without retraining.

Numerical Evaluation

Stock Price Forecasting

• Dataset: Daily S&P 500 log-returns modeled as signals over a correlation graph of 468 stocks. Conditioning incorporates both graph topology from fundamental data and 12 market node features.
• Results: The U-GNN consistently outperforms the geometric random walk baseline across accuracy metrics (CRPS, RMSE, MIS90, MAE), reducing price prediction errors by 21–28%. Directional accuracy improves above chance, capturing nontrivial market structure.
• Distributional Fidelity: The U-GNN reproduces stylized statistical properties (autocorrelation, volatility clustering, dominant eigenvalues) and places non-negligible mass on rare, extreme moves, reflecting realistic uncertainty. Under both GRW and U-GNN likelihoods, generated samples track the real distribution more closely than baseline models.
• Implications: The capability to sample diverse, realistic trajectories enables modeling uncertainty and rare tail events in financial forecasting, directly informing risk assessment and scenario generation.

Wireless Resource Allocation

• Setup: Optimal power control in multi-user interference networks with 400 tx-rx pairs, modeled as signals over channel-interference graphs. Conditioning is on channel topology and node features describing link and interference statistics.
• Results: The U-GNN achieves near-expert performance on ergodic rates and feasibility, matching the reference primal-dual policy over the full distribution (e.g., only 5% and 0.9% gap to expert at p5 and p10 percentiles, respectively), outperforming full power and average power deterministic baselines.
• Stochastic Time-Sharing: The U-GNN learns a stochastic policy, time-sharing between distinct allocations to attain ergodic feasibility for all users, which deterministic allocations fail to do. Size-transferability experiments show robustness and scalability to larger graphs without retraining.
• Implications: Bypassing expensive online optimization, the stochastic diffusion policy yields optimal allocations in a single sampling pass, supporting scalable, feasible resource management in complex wireless environments.

Theoretical and Practical Implications

The paper makes several strong claims:

• Generalization: The diffusion framework and U-GNN architecture unify generative modeling of stochastic graph signals, subsuming application-specific approaches and supporting diverse domains.
• Distributional modeling: Sampling from conditional laws (rather than regressing means) captures uncertainty, diversity, and rare events, critical for decision-making under uncertainty.
• Architectural scalability: The learned node selection and stride design ensures scalability and transferability, with empirical evidence of size-invariant performance on large graphs.
• Stochastic policy learning: In resource allocation, generative stochastic policies outperform deterministic baselines by leveraging time-sharing and sampling diversity.

The implications for future developments include integrating latent graph-aware diffusion processes, graph transformers for enhanced spatio-temporal modeling, and extending these architectures to more general generative settings in graph-structured domains.

Conclusion

The paper introduces a unified methodology for generative modeling of stochastic signals on graphs, leveraging a conditioned denoising diffusion process parametrized by the U-GNN, a graph-domain adaptation of U-Net architectures. Empirical validation in financial forecasting and wireless resource allocation demonstrates superior performance in both accuracy and distributional fidelity, substantiating the paper's claims about generalization and scalability. The methodological and architectural advancements have significant theoretical and practical implications for uncertainty modeling and resource optimization in complex graph-structured systems. Future work should explore hybrid architectures and latent diffusion for further flexibility and expressiveness in generative graph machine learning.