Group-aware Shift Explanations

This is the official repository for the paper Rectifying Group Irregularities in Explanations for Distribution Shift.

Before running any scripts, download the necessary datasets into the data/ directory by following the directions under "Dataset Setup." The scripts can then be run using Docker as described in Section "Running with Docker."

Dataset Setup

Adult Dataset

Download the Adult tabular dataset from https://archive.ics.uci.edu/ml/machine-learning-databases/adult/adult.data and create the directory data/adult/ where adult.data should be placed.

Breast Dataset

Download the Breast tabular dataset from https://www.kaggle.com/datasets/uciml/breast-cancer-wisconsin-data and create the directory data/breast where data.csv should be placed.

ImageNet Datset

Download the ImageNet dataset from https://www.kaggle.com/c/imagenet-object-localization-challenge/overview/description. Then, create a soft link to the directory ILSVRC/Data/CLS-LOC/val/ in the ImageNet data to the directory data/imagenet/ in this repository.

Running with Docker

Navigate to the main directory explainable_domain_shift/ and build the docker image with the following:

docker build -t explainshift:latest .

To run the experiments in the container, run the following:

docker run -it explainshift:latest \
	-v $(pwd)/base:/workspace/explainable_domain_shift/data/base \
	python base_exp.py --method kmeans --dataset adult --train

Worst-case Robustness experiments:

docker run -it explainshift:latest \
	-v $(pwd)/base:/workspace/explainable_domain_shift/data/base \
	python base_exp.py --method kmeans --dataset adult --adv --train

The allowed values for the --method flag are kmeans for K-cluster, dice, and ot. The allowed values for the --dataset flag are adult, breast, nlp, nlp-amazon, fmow, iwildcam, imagenet.

Citation

@article{stein2023rectifying,
  title={Rectifying Group Irregularities in Explanations for Distribution Shift},
  author={Stein, Adam and Wu, Yinjun and Wong, Eric and Naik, Mayur},
  journal={arXiv preprint arXiv:2305.16308},
  year={2023}
}