RS3GP: Recurrent Sparse Spectrum Signature Gaussian Processes

RS3GP is a Python package that implements Random Fourier Signature Features for Gaussian Processes, enabling scalable kernel methods for sequential data. It provides efficient approximations of signature kernels, facilitating their application in large-scale machine learning tasks.

Features

• Random Fourier Signature Features: Efficient approximation of signature kernels using random features.
• Scalable Gaussian Processes: Apply Gaussian Process models to sequential data with reduced computational complexity.
• Probabilistic Time Series Forecasting: Outputs full predictive distribution for time series forecasting by a single pass through the time series using efficient an recurrent formulation, enabling uncertainty quantification and robust decision-making.

Installation

To install RS3GP using pip:

pip install git+https://github.com/tgcsaba/rs3gp.git

Citation

If you use this code, please cite:

Csaba Tóth, Masaki Adachi, Michael A. Osborne, Harald Oberhauser.
Learning to Forget: Bayesian Time Series Forecasting using Recurrent Sparse Spectrum Signature Gaussian Processes.
Proceedings of AISTATS 2025, PMLR 258:4654–4662
Link to paper

@InProceedings{pmlr-v258-toth25b,
  title = {Learning to Forget: Bayesian Time Series Forecasting using Recurrent Sparse Spectrum Signature Gaussian Processes},
  author = {T{\'o}th, Csaba and Adachi, Masaki and Osborne, Michael A. and Oberhauser, Harald},
  booktitle = {Proceedings of The 28th International Conference on Artificial Intelligence and Statistics (AISTATS)},
  pages = {4654--4662},
  year = {2025},
  publisher = {PMLR},
  url = {https://proceedings.mlr.press/v258/toth25b.html}
}

Contact

For questions or issues, please open a GitHub issue or contact the authors via the email provided in the paper.