FairEdit

Relevent Publication

FairEdit: Preserving Fairness in Graph Neural Networks through Greedy Graph Editing

https://arxiv.org/abs/2201.03681

Requirements

Use the environment.yml to find the associated packages and versions.

Datasets

Example data loaders can be found in data_loader_example.py

Can be run by passing in an argument denoting the data set to load, such as data_loader_example.py -dataset credit

Note for the above command to work the first line in the for loop has to be commented

Models

Folder models holds the various architectures. Examples to load, train, and evaluate can be found in run_models.py. Run_models.py has a lot of possible arguments, here as an example to run:

python adjusted_training.py --dropout 0.5 --hidden 16 --lr 1e-3 --epochs 1000 --model gcn --training_method standard --dataset german --seed 1

where --model can be ['gcn', 'sage'] and --dataset can be ['german', 'credit', 'bail'] and --training_method can be ['standard','brute','fairedit']

model weights are saved to weights folder and evaluation metricss are saved to results.

comparison_training.py

contains models to compare against. Example:

python comparison_training.py --dropout 0.5 --hidden 16 --lr 1e-3 --epochs 1000 --model fairgnn --dataset german --seed 1 --num_layers 32

where --model can be ['fairgnn', 'fairwalk'] and --dataset can be ['german', 'credit', 'bail']

Note: Updated in adjusted_training.py, try this one!

To perform various types of training, code can be found in the training_methods folder

In here, you will find methods such as nifty, brute_force, and fairedit, which will all incorporate some form or fair training.