Home

Awesome

Conformal time-series forecasting

Implementation for Stankevičiūtė et al. "Conformal time-series forecasting", NeurIPS 2021.

This codebase builds on the implementation for "Frequentist Uncertainty in Recurrent Neural Networks via Blockwise Influence Functions" (ICML 2020), available at https://github.com/ahmedmalaa/rnn-blockwise-jackknife under the BSD 3-clause license.

Installation

Python 3.6+ is recommended. Install the dependencies from requirements.txt.

Replicating Results

To replicate experiment results, run the notebooks:

You can download the publicly available data for this work at the UKHSA data dashboard and UCI Machine Learning Repository (note the data format may have changed by maintainers of the datasets). As the MIMIC-III dataset requires PhysioNet credentialing to access, you must become a credentialed user on PhysioNet before accessing the data. To get access to the dataset as used in this work, please contact @DrShushen or @ahmedmalaa and provide proof of your PhysioNet credentialing.

Citing

If you use our code in your research, please cite:

@inproceedings{stankeviciute2021conformal,
  author = {Stankevičiūtė, Kamilė and Alaa, Ahmed M. and {van der Schaar}, Mihaela},
  title = {Conformal time-series forecasting},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2021}
}