Interpretable Counterfactual Explanations By Minimizing Uncertainty

This repository contains the code for Generating Interpretable Counterfactual Explanations By Implicit Minimisation of Epistemic and Aleatoric Uncertainties; Lisa Schut*, Oscar Key*, Rory McGrath, Luca Costabello, Bogdan Sacaleanu, Medb Corcoran, Yarin Gal.

The paper has been accepted at AISTATS 2021.

It is also available on arXiv: arXiv:2103.08951

Environment set up

Install Miniconda or Anaconda, then use the command: conda env update -f environment.yml

To run the tests: pytest tests

To format the code: black -l 99 **/*.py

Reproducing results

1. Train an ensemble with and without adversarial training. This will generate checkpoints in ~/ces_results/ensembles/[id]:

   python experiments/train_ensemble.py --results_dir=~/ces_results --data_dir=~/pytorch-datasets --epochs=100 --n_ensembles=50 --n_hidden=200 --dataset=mnist --adv_training

   python experiments/train_ensemble.py --results_dir=~/ces_results --data_dir=~/pytorch-datasets --epochs=100 --n_ensembles=50 --n_hidden=200 --dataset=mnist

2. Generate and evaluate the CEs

   a. For quantitative evaluation (generating CEs to examine by eye)

   python experiments/exp_qualitative_evaluation.py --results_dir=~/ces_results --data_dir=~/pytorch-datasets --adv_training_id=[id] --no_adv_training_id=[id]

   b. For qualitative evaluation (computing evaluation metrics over several repeats of the experiment)

   python experiments/exp_quantitative_evaluation.py --results_dir=~/ces_results --data_dir=~/pytorch-datasets --adv_training_id=[id] --no_adv_training_id=[id]

   Note that running the evaluation will generate a set of CEs, and train several evaluation autoencoders. These are cached in results_dir to allow the evaluation to be re-run quickly for debugging purposes. If you alter the code you may need to delete this cache to see the results of your change.

Pointers to useful files

• uces/generators.py contains the algorithm for generating CEs
• uces/im_evaluation.py contains the implementation of the IM1 and IM2 evaluation metrics

Citation

If the paper or code has been useful in your work, please reference it as follows:

@InProceedings{pmlr-v130-schut21a,
  title     = { Generating Interpretable Counterfactual Explanations By Implicit Minimisation of Epistemic and Aleatoric Uncertainties },
  author    = {Schut, Lisa and Key, Oscar and Mc Grath, Rory and Costabello, Luca and Sacaleanu, Bogdan and Corcoran, Medb and Gal, Yarin},
  booktitle = {Proceedings of The 24th International Conference on Artificial Intelligence and Statistics},
  pages     = {1756--1764},
  year      = {2021},
  editor    = {Banerjee, Arindam and Fukumizu, Kenji},
  volume    = {130},
  series    = {Proceedings of Machine Learning Research},
  month     = {13--15 Apr},
  publisher = {PMLR},
  pdf       = {http://proceedings.mlr.press/v130/schut21a/schut21a.pdf},
  url       = {http://proceedings.mlr.press/v130/schut21a.html},
}

License

We release this code under the MIT license, see LICENSE.txt.