fair learning without sensitive attributes

Official code for WSDM 2022 paper: Towards Fair Classifiers Without Sensitive Attributes: Exploring Biases in Related Features

Dataset

Two datasets, Law_school and Compas, are provided in this project.

Algorithm

We provide several algorithms for fair learning:

corre: constrain correlation with sensitive attributes
groupTPR: regularize group-wise true positive rate for fairness
remove: remove related attributes
learnCorre: learn to constrain correlation with related attributes

Example

An example on adult dataset is provided here:

python main.py --method="learnCorre" --dataset=adult --related age --r_weight 0.1 --weightSum=0.1 --beta=0.4 --seed=42