Manipulating SHAP via Adversarial Data Perturbations

This repository contains data and code for the article:

H. Baniecki, P. Biecek. Manipulating SHAP via Adversarial Data Perturbations (Student Abstract). In: AAAI Conference on Artificial Intelligence (AAAI), 36(11):12907-12908, 2022. https://doi.org/10.1609/aaai.v36i11.21590

2022-06-12 Update. Values of Kendall tau distance reported in the article come from the scipy.stats.kendalltau() function, which in fact computes the Kendall tau coefficient (see the SciPy GitHub issue on "Kendall tau distance" scipy/scipy#7089). Knowing that the (normalized) distance equals (1 - coefficient) / 2, the actual distance values equal 0.20 and 0.07, respectively. We updated export_and_table.ipynb to account for this error, which doesn't change the conclusion.

Python version: 3.9.2

• alg directory with the algorithm's code implementation
• data directory with the datasets
• results directory with the pickled metadata and logs
• scenario_heart.ipynb recreates the heart analysis
• scenario_apartment.ipynb recreates the apartment analysis
• export_and_table.ipynb converts the .p result files and computes Kendall tau
• figures_and_table.R creates Figures and computes the remaining distances