This repository contains the code for the experiments of the paper DispaRisk: A Framework for Assessing and Interpreting Disparity Risks in Datasets.
To run the experiments, please ensure you create a Python environment with python>=3.8.6.
The folder ‘celeba’ contains all the code and experiments on the CelebA dataset. First, you need to download the celeba.hdf5 file from this link and provide the path to this file to the --data-dir.
The folder tabular_datasets contains all the experiments for the three tabular datasets: Census Income (CI), Dutch Census (DC), and Compas Recidivism (CR). You need to provide the name of the dataset for the --datasets parameter. You also need to provide the corresponding model: (1) sgd_lr for Linear family, (2) mlp_one_layer for 1MLP, and (3) two_one_layer for 2MLP.
In each folder, you will find Jupyter notebooks with all the analysis not reported in the paper. Please check Analysis_V_Information.ipynb, sampling_tests.ipynb, and plots.ipynb files in the tabular_datasets folder.
Resulting models trained on the Tabular and CelebA datasets can be found in this Google Drive folder. You will also find the raw data for computing the results reported in the paper: PVIs, fairness metrics, and performances.