Adversarially Robust Recourse

This repository provides the code necessary to reproduce the experiments of the paper "On the Adversarial Robustness of Causal Algorithmic Recourse".

Prequisites

Install the packages in requirements.txt, for instance using

python -m venv myenv/
source myenv/bin/activate
pip install -r requirements.txt

Reference implementations

Our proposed method to generate adversarially robust recourse

Please refer to evaluate_recourse.py for an overview. The implementations are found in recourse.py.

In the linear case, the method used to generate standard recourse is LinearRecourse. To generate robust recourse against some uncertainty epsilon, we implement Equation 5 in the paper as

w, b = model.get_weights()
Jw = w if scmm is None else scmm.get_Jacobian().T @ w
dual_norm = np.sqrt(Jw.T @ Jw)
b = b + dual_norm * epsilon

In the nonlinear case, please refer to DifferentiableRecourse.

Our proposed ALLR regularizer

Please refer to train_classifiers.py for an overview. The implementations are found in train_classifiers.py.

In the linear setting, refer to LogisticRegression. In the non-linear setting, refer to LLR_Trainer.

Plotting the figures of the paper

Since running the full set of experiments can be relatively time-consuming, we already provide the numerical results of the experiments in the folder results/. To plot the results, simply run

python plot_figure1.py     # plots Figure 3 in the paper
python plot_figure2.py     # plots Figure 4 in the paper
python plot_figures3_4.py  # plots Figure 5 and Figure 6 in the paper

Figures will be created in the folder figures/

Running the experiments

Simply run

python run_benchmarks.py --seed 0
python run_benchmarks.py --seed 1
python run_benchmarks.py --seed 2
python run_benchmarks.py --seed 3
python run_benchmarks.py --seed 4

If you wish to retrain the decision-making classifiers and the structural equations, simply delete the folders models/ and scms/ before running run_benchmarks.py. Otherwise, the pretrained classifiers and SCMs will be used.