This repository contains the code to conduct classification experiments on fairness datasets using a L2-regularized logistic regression model. This includes the experiments on fairness datasets in the paper "Characterizing the Representation Disparity of Differential Privacy" by Gardner, Morgenstern, and Oh (2021).
Configure environment by running: pip3 install -r requirements.txt.
For our experiments, we use Python 3.6.7 and a laptop CPU. Running each training iteration takes only a few minutes.
The main script used to run experiments is train_l2lr.py. Example commands to train various models are included in the header of that file. One example:
python3 train_l2lr.py --learning_rate 0.01 --epochs 1000 --dataset german --niters 50 \
--sensitive_attribute sex \
--majority_attribute_label male \
--minority_attribute_label female \
--batch_size 64 \
--dp --dp_l2_norm_clip 10.0 --dp_noise_multiplier 0.4
The following datasets are required for the experiments in this repo:
Each dataset should be downloaded and placed in the ./datasets/ directory.
We note that our preprocessing of the COMPAS dataset follows the open-source code of [1], which was accessed on codalab.
For experiments with DP, we use the implementation of DP-SGD from TensorFlow Privacy.
[1] Khani, F., Raghunathan, A., and Liang, P. Maximum weighted loss discrepancy. arXiv preprintarXiv:1906.03518, 2019.