Feature Responsiveness Scores: Model-Agnostic Explanations for Recourse

Requirements

• Python 3.9+

To install requirements:

pip install -r requirements.txt

CPLEX

To run experiment code, you will need to download CPLEX (v. 22.1). CPLEX is cross-platform optimization tool solver a Python API. Running the pip install above will download the community edition (CE) from pip. The problem size allowed on the CE version will not be enough to run the entire experiment. But the full version of CPLEX is free for students and faculty members at accredited institutions. To download CPLEX:

1. Register for IBM OnTheHub
2. Download the IBM ILOG CPLEX Optimization Studio from the software catalog
3. Install CPLEX Optimization Studio.
4. Setup the CPLEX Python API as described here.

If you have problems with CPLEX, please check the CPLEX user manual or the CPLEX forums.

Directory Structure

├── data         # datasets
├── reach        # modified code from reachml
├── src          # our source code for generating explanations
└── scripts      # scripts that call the source code for testing and running experiments

reachml refers to the source code to generate reachable sets from Kothari et al. (2024).

Running Experiment

To run the experiments in the paper, run from the repository:

python3 scripts/experiments/run_experiment_pipeline.py --data_name DATASET_NAME

where DATASET_NAME is one of german, fico, givemecredit.

To generate the metrics shown in Table 3:

python3 scripts/metrics/combine_metrics.py

will result in a csv file named all_metrics.csv in a results directory.