On the Problem of Underranking in Group-Fair Ranking.

This repository consists of code and datasets for our paper. We extend the code of FA*IR and implement the DP algorithm in Celis et al. - Ranking with Fairness Constraints, to compare our work with these algorithms as baselines.

Requirements

python, numpy, matplotlib

Re-ranking for fairness

The root directory contains the following bash-script for re-ranking with ALG and also the baselines FA*IR (FAIR) and Celis et al (CELIS).

postprocess.sh

To run the experiments

python3 main.py --postprocess <dataset> <method> --k <top k ranks> --rev_flag <True/False> --multi_group <True/False> re-ranks the true ranking based on the method (ALG/FAIR/CELIS) and writes the re-ranked data in the folder data/<dataset>/..._<experiment>.txt. If rev_flag is set the true ranking is reversed first and then postprocessing is applied. The flag multi_group is set for the experiments on the German Credit dataset with more than 2 groups.

Uncomment the experiment required and use ./postprocess.sh

Result Evaluation

The following bash-script evaluates results on all the datasets.

evaluateResults.sh

Two types of evaluations

1. Top k evaluation, python3 main.py --evaluate <dataset_protectedgroup> --topk <top k ranks> --rev_flag <True/False> evaluates the algorithms for the top k ranks, where k is given as input argument topk. The usage of ref_flag is same as above.
2. Evaluation for k consecutive ranks python3 main.py --evaluate german_age25 --consecutive 40 --start 21 --rev_flag <True/False> evaluates the algorithms for consecutive ranks (use the argument consecutive to choose how many consecutive ranks to evaluate for), starting from rank given as start. The usage of ref_flag is same as above.