Reproduce “Graph Signal Generative Diffusion Models” on OpenResearch

About this paper

Yigit Berkay Uslu, Samar Hadou, Sergio Rozada, Shirin Saeedi Bidokhti, Alejandro Ribeiro

Researchers from the University of Pennsylvania developed a U-Graph Neural Network (U-GNN) architecture for generative diffusion models, enabling the generation of stochastic signals on fixed graph structures. This framework, demonstrated through stock price forecasting, generated realistic future stock return trajectories and achieved lower Continuous Ranked Probability Scores (CRPS) compared to a Geometric Random Walk baseline, indicating improved probabilistic prediction calibration.

Key insights

• The U-Net's effective encoder-decoder structure and skip connections can be successfully translated to graph domains by designing appropriate graph convolutional and hierarchical pooling layers.
• Employing zero-padding with nested sampling matrices allows for multi-resolution feature extraction on graphs while ensuring all graph convolutions respect the original graph's connectivity.
• Integrating diffusion models with specialized graph architectures offers a powerful approach for generating stochastic graph signals, capable of capturing inherent uncertainties and generating diverse future trajectories.

Summary by alphaXiv.