README.md

Robust Bayesian Recourse

We introduce a model-agnostic recourse that minimizes the posterior probability odds ratio along its min-max robust counterpart with the goal of hedging against future changes in the machine learning model parameters. Robust Bayesian Recourse explicitly takes into account possible perturbations of the data in a Gaussian mixture ambiguity set prescribed using the optimal transport (Wasserstein) distance.

Details of the model architecture and experimental results can be found in our paper

Please CITE our paper whenever this repository is used to help produce published results or incorporated into other software:

@InProceedings{nguyen2022robust,
    title = {Robust {B}ayesian recourse},
    booktitle = {Proceedings of the Thirty-Eighth Conference on Uncertainty in Artificial Intelligence},
    year =   {2022},
    series = {Proceedings of Machine Learning Research},
    publisher = {PMLR},
    author = {Nguyen, Tuan-Duy H. and Bui, Ngoc and Nguyen, Duy and Yue, Man-Chung and Nguyen, Viet Anh}
}

Data

Our numerical experiments employ 1 synthetic and 3 real-world datasets.

The synthetic dataset is automatically generated when you run the experiments.

For real-world datasets, we recommend using pandas for preprocessing, following the steps described in our paper.

• German Credit:
  • Original dataset --> ./data/german_small.csv
  • Shifted dataset --> ./data/corrected_german_small.csv
• Small Business Administration (SBA): Source ('SBA Case' file in Supplementary Material section)
  • 1989 - 2005 --> ./data/sba_8905.csv
  • 2006 - 2014 --> ./data/sba_0614.csv
• Give Me Some Credit (GMC): Source
  • ./data/gmc.csv
  • ./data/gmc_shift.csv

Usage

1. Train a classifier

   python train.py --clf mlp --data synthesis german sba gmc --kfold 5 --num-future 100 --num-proc 16

   • Args:
     • --clf: List of types of classifiers to be trained
     • --data: List of datasets in use
     • --kfold: Number of subsamples for cross-validation
     • --num-future: Number of future (shifted) classifiers
     • --num-proc: Number of parallel processes

2. Run experiments

   python run_expt.py -e 5 --datasets synthesis german sba gmc -clf mlp -uc --methods wachter lime_roar limels_roar rbr --num-proc 32

   • Args:
     • -e: Selected experiment (for customization purpose, check run_expt.py for details)
     • --datasets: List of datasets in use
     • --clf: List of experimenting classifiers
     • -uc: Update configs for multiple trials
     • --methods: Number of experimenting methods
     • --num-proc: Number of parallel processes for each method