Distributionally Robust Recourse Action (DiRRAc)

Source-code for paper Distributionally Robust Recourse Action (DiRRAc) (ICLR 2023)

1. Install requirements

pip install -r requirements.txt

2. Experiments with synthetic data:

1. Feasible set of DiRRAc: Figure2

2. Comparison of DiRRAc and ROAR: Figure3

3. Impact of magnitude of distribution shifts to the empirical validity: Figure4

4. Different types of data distribution shifts: Figure5

5. Model parameters in 2D: Figure6

6. Impact of varying parameters of DiRRAc: Figure7

7. Cost of Robustness of DiRRAc: Figure9

Results of each figure are saved in result/

3. Experiments with real-world data:

Generate recourse and evaluate on 3 different real-world datasets:

python train_real_data.py --mode linear --num_samples <number of samples to evaluate>

Example:

python train_real_data.py --mode linear --num_samples 40

Experiments with prior on the covariance matrix:

python train_real_data.py --mode linear --num_samples <number of samples to evaluate> --sigma_identity True

Example:

python train_real_data.py --mode linear --num_samples 40 --sigma_identity True

Result of csv format is saved in result/real_data/

4. Experiments with non-linear model:

Generate recourse for non-linear model and evaluate on 3 different real-world datasets:

python train_real_data.py --mode non --num_samples <number of samples to evaluate>

Example:

python train_real_data.py --mode non --num_samples 20