Explaining Distribution Shifts Overview

Greeting 👋! This is the repository for the IMCL Publication Towards Explaining Distribution Shifts. In this work we answer the question: ''What is a distribution shift explanation?'' and introduce a novel framework for explaining distribution shifts via transportation maps between a source and target distribution which are either inherently interpretable or interpreted using post-hoc interpretability methods.

Code for the ICML Publication: Towards Explaining Distribution Shifts

To recreate the results for the experiment, first create the corresponding environment (via conda by conda env create -f environment.yml).

Then unzip the data.zip file.

Within the notebooks folder, you can find all the corresponding jupyter notebooks to recreate the experiments. For example, to recreate the shift explanation results for the adult-income dataset, just run the adult-income-experiment.ipynb notebook (and for the baselines, run the adult-income-experiment-baseline.ipynb).

If you have any troubles running these, feel free to reach out to the contacts listed in our ICML paper or via the contact from the first-author's website.

Cheers!

Bibtex Citation

If you reference this work, please consider citing our publication:

@inproceedings{kulinski2023towards,
  title={Towards explaining distribution shifts},
  author={Kulinski, Sean and Inouye, David I},
  booktitle={International Conference on Machine Learning},
  pages={17931--17952},
  year={2023},
  organization={PMLR}
}