GPLFR (Gaussian Process Latent Factor Regression)

Lightweight reference implementation of Gaussian Process Latent Factor Regression (GPLFR), accompanying the paper

Install

pip install -e ".[dev]"

This installs GPLFR plus the dependencies needed for demos and tests.

Quickstart

Start with demos/quickstart.ipynb, which generates a small synthetic problem, fits GPLFR, and visualizes held-out predictions.

predict returns the posterior predictive mean; pass return_std=True for the predictive standard deviation (add include_noise=True to include observation noise), or call sample(X_new, n_samples) to draw from the predictive distribution.

For a script-style reproduction of the synthetic learning-curve experiment, see demos/synthetic_learning_curve/; the plot is written to demos/synthetic_learning_curve/learning_curve.png.

Source files

Path	Role
model.py	GPLFR model
synthetic.py	synthetic data generator
kernels.py	covariance kernels
demos/	quickstart and synthetic data learning-curve example
tests/	smoke tests plus numerical-correctness tests (collapsed likelihood, predictive mean)

Relation to exoplanet climate prediction / ThousandWorlds

GPLFR was motivated by the exoplanet climate prediction problem. This repository is a lightweight reference implementation of GPLFR. For an expanded version of the exoclimate dataset, more thorough benchmarking, and the problem-specific adaptation of GPLFR, see ThousandWorlds and the associated paper.

Citation

@article{gplfr2026,
  title   = {Gaussian Process Latent Factor Regression for Low-Data, High-Dimensional Output Problems},
  author  = {Stevenson, Edward T. and Wolf, Eric T. and Mak, Mei Ting and Mayne, N. J. and Cranmer, Miles},
  journal = {arXiv preprint arXiv:2606.06576},
  year    = {2026},
  url     = {https://doi.org/10.48550/arXiv.2606.06576}
}