This repository contains source code necessary to reproduce some of the main results in the paper:
If you use this software, please consider citing:
@article{agarwal2021unified,
title={Towards a Unified Framework for Fair and Stable Graph Representation Learning},
author={Chirag Agarwal and Himabindu Lakkaraju* and Marinka Zitnik*},
year={2021},
booktitle={arXiv},
}
Our framework NIFTY can learn node representations that are both fair and stable (i.e., invariant to the sensitive attribute value and perturbations to the graph structure and non-sensitive attributes) by maximizing the similarity between representations from diverse augmented graphs.
This repository is built using PyTorch. You can install the necessary libraries by pip installing the requirements text file pip install -r ./requirements.txt
After installing the packages from the requirements.txt, install the PyTorch Geometric packages following the instructions from here.
Note: We ran our codes using python=3.7.9
We ran our experiments on three high-stake read-world datasets. All the data are present in the './datasets' folder. Due to space constraints the edge file of the credit dataset is zipped.
The main scripts running the experiments on the state-of-the-art GNNs and their NIFTY-augmented counterparts is in nifty_sota_gnn.py
Script 1: Evaluate fairness and stability performance of GCN (for German Graph dataset)
python nifty_sota_gnn.py --dropout 0.5 --hidden 16 --lr 1e-3 --epochs 1000 --model gcn --dataset german --seed 1
The AUCROC of estimator: 0.7605
Parity: 0.3952 | Equality: 0.2731
F1-score: 0.8078
CounterFactual Fairness: 0.2960
Robustness Score: 0.1160
Script 2: Evaluate fairness and stability performance of NIFTY-GCN (for German Graph dataset)
python nifty_sota_gnn.py --drop_edge_rate_1 0.001 --drop_edge_rate_2 0.001 --drop_feature_rate_1 0.1 --drop_feature_rate_2 0.1 --dropout 0.5 --hidden 16 --lr 1e-3 --epochs 1000 --model ssf --encoder gcn --dataset german --sim_coeff 0.6 --seed 1
The AUCROC of estimator: 0.7205
Parity: 0.0104 | Equality: 0.0199
F1-score: 0.8235
CounterFactual Fairness: 0.0
Robustness Score: 0.0
Script 3: Evaluate fairness and stability performance of FairGCN baseline (for German Graph dataset)
python baseline_fairGNN.py --dropout 0.5 --hidden 16 --lr 1e-3 --epochs 1000 --dataset german --seed 1 --model gcn
The AUCROC of estimator: 0.7549
Parity: 0.2763 | Equality: 0.1723
F1-score: 0.8251
CounterFactual Fairness: N/A
Robustness Score: 0.0440
Script 4: Evaluate fairness and stability performance of RobustGCN (for German Graph dataset)
python nifty_sota_gnn.py --dropout 0.5 --hidden 16 --lr 1e-3 --epochs 1000 --model rogcn --dataset german --seed 5
The AUCROC of estimator: 0.6230
Parity: 0.2449 | Equality: 0.2048
F1-score: 0.6143
CounterFactual Fairness: 0.0880
Robustness Score: 0.1320
Note that the code in this repository is licensed under MIT License. Please carefully check them before use.
If you have questions/suggestions, please feel free to email or create github issues.