CARE

This repository contains the implementation source code of the following paper:

CARE: Coherent Actionable Recourse based on Sound Counterfactual Explanations

BibTeX:

@article{rasouli2022care,
  title={CARE: Coherent Actionable Recourse based on Sound Counterfactual Explanations},
  author={Rasouli, Peyman and Chieh Yu, Ingrid},
  journal={International Journal of Data Science and Analytics},
  pages={1--26},
  year={2022},
  publisher={Springer}
}

Setup

1- Clone the repository using HTTP/SSH:

git clone https://github.com/peymanrasouli/CARE

2- Install the following package containing GCC/g++ compilers and libraries:

sudo apt-get install build-essential

3- Create a conda virtual environment:

conda create -n CARE python=3.7

4- Activate the conda environment:

conda activate CARE

5- Standing in CARE directory, install the requirements:

pip install -r requirements.txt

Explaining instances

1- To explain a particular instance using CARE, run:

python main.py

2- To explain a particular instance using CARE, CFPrototype, DiCE, and CERTIFAI simultaneously, run:

python care_cfprototype_dice_certifai.py

Explanation Example

An actionable recourse generated by CARE for an instance from Adult data set:

Features	Original Instance	Actionable Recourse
age	25	30
capital-gain	0	_
capital-loss	0	_
hours-per-week	40	45
work-class	Private	_
education	Bachelors	Masters
marital-status	Not-in-family	_
occupation	Sales	Prof-specialty
relationship	Never-married	_
race	White	_
sex	Female	_
native-country	United-States	_
Class	<=50K	>50K

IF 'age was 30' AND 'hours-per-week was 45' AND 'education was Masters' AND 'occupation was Prof-specialty',

THE PERSON WOULD BE CLASSIFIED AS '>50K income' RATHER THAN '<=50K income'.

Reproducing the validation results

1- To reproduce the results of performance validation of CARE's modules, run:

python care_module_performance.py

2- To reproduce the results of soundness validation, run:

python care_soundness.py

3- To reproduce the results of coherency validation, run:

python care_coherency.py

4- To reproduce the validation results of temporal action sequence, run:

python care_temporal_action_sequence.py

Reproducing the benchmark results

1- To reproduce the results of CARE_{VALIDITY} config vs. baselines, run:

python benchmark_validity.py

2- To reproduce the results of CARE_{{VALIDITY, SOUNDNESS}} config vs. baselines, run:

python benchmark_validity_soundness.py

3- To reproduce the results of CARE_{{VALIDITY, SOUNDNESS, COHERENCY}} config vs. baselines, run:

python benchmark_validity_soundness_coherency.py

4- To reproduce the results of CARE_{{VALIDITY, SOUNDNESS, COHERENCY, ACTIONABILITY}} config vs. baselines, run:

python benchmark_validity_soundness_coherency_actionability.py