This repository provides the code implementation for FIPSER. This implementation is based on the code of DICE Consequently, the structure of this repository is similar to the structure of DICE, which includes:

• DICE_baseline includes the implementation of baseline NeuronFair
• DICE_data includes functions to load datasets
• DICE_model includes the implementation of DNN model
• DICE_tutorial includes the source code of ADF, EIDIG and DICE
• DICE_utils includes utility functions of DICE
• Raw_data includes raw dataset files(meps15/16)
• Clusters includes pickles from KMeans clustering
• Datasets includes pre-processed datasets
• Models includes trained models checkpoints
• Requirements set of required libraries for running the tool on local machine
• License file
• fairness_improvement includes the implementation of model retraining and evaluation

Python 3.8
numpy==1.22.0
pandas==1.5.1
tensorflow==2.7.0
scipy==1.4.1
argparse==1.1
protobuf==3.9.2
scikit-learn==0.22.2.post1
aif360==0.4.0
IPython==7.13.0
regex==2022.10.31

For the experiment of effectiveness and efficiency, use DICE_tutorial/ADF.py, DICE_tutorial/EIDIG.py, DICE_tutorial/DICE_Search.py, and NF_baseline/NF_main/src/NF.py

For the experiment of retraining, use fairness_improvement/data_process.py to split datasets, fairness_improvement/pretrain.py to pretrain models, fairness_improvement/retrain.py to retrain models, and fairness_improvement/eval.py to evaluate the accuract and fairness of models before or after retraining.

Before running the fairness_improvement/retrain.py, put the search result of DICE into fairness_improvement/generated_idip, rename it to [dataset]_[index of sensitive attribute (starting from 1)]_[sampling method]here's an example:

generated_idip

heart_2_cluster

2_10runs

global_inputs_0.npy

global_inputs_90_0.csv

local_inputs_0.npy

local_inputs_90_0.csv

QID_RQ2_10runs.npy

RQ2_table.csv

total_disc_0.csv

total_inputs_0.npy