fast-graph-gp 0.1.3

Graph Random Features for Scalable Gaussian Processes (GRF-GP)

Fast-Graph-GP

Fast-Graph-GP is the package for performing fast Gaussian Process (GP) inference on graphs. Internally, it uses Graph Random Features (GRFs) to compute a unbiased & sparse estimate of a family of well-known graph node kernels. It further uses path-wise conditioning to leverage the sparsity of the kernel approximation, enabling you to perform GP model train / inference in $\mathcal{O}(N^{3/2})$ time and $\mathcal{O}(N)$ space complexity.

This package is in beta. Package development TODOs:

• support graph GP inference / training with GRFs
• supoort low-rank GRF with Johnson-Lindenstrauss Transform (JLT)
• support solving low-rank linear system solver using Woodbury formula
• support GRF++ samplers

Examples

For a detailed example of training and using a Graph GP model, refer to the example notebook.

Installation

Install Fast-Graph-GP via pip:

pip install fast-graph-gp

Documentation

Hosted documentation is available at:

https://matthewzhang473.github.io/Fast-Graph-GP/index.html

Citing Us

If you use Fast-Graph-GP, please cite the following papers:

@article{zhang2025graph,
title={Graph random features for scalable Gaussian processes},
author={Zhang, Matthew and Lin, Jihao Andreas and Choromanski, Krzysztof and Weller, Adrian and Turner, Richard E and Reid, Isaac},
journal={arXiv preprint arXiv:2509.03691},
year={2025}
}