FairDisCo

implementation of paper Fair Representation Learning: An Alternative to Mutual Information (KDD 2022)

Requirements

• python >= 3.6
• pytorch >= 1.4

Usage

For dataset Adult-sex with default parameters $lg\beta \in [0,1,2,3,4,5,6,7,8,9]$, use following command

python FairDisCo.py --cuda 0 --data Adult-sex

For other dataset, you can change the data parameter

python FairDisCo.py --cuda 0 --data Adult-sex/Adult-race/Health/Bank/German/Creadit/Student/Compas-sex/Compas-race

For dataset MNIST or MNIST-color with default parameters $lg\beta=7$, use following command

python FairDisCoImage.py --cuda 0 --data color/number

Repository contents

file	description
data/	The datasets dir
res/	The results dir
FairDisCo.py	The training code of fairdisco
FairDisCoImage.py	The training code of fairdisco with image data
module.py	The definition of fairdisco's model structure
utils.py	The data loading function and evaluation function

Experimental Result