Skip to content

Fair ML in credit scoring: Assessment, implementation and profit implications

Notifications You must be signed in to change notification settings

kozodoi/Fair_Credit_Scoring

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

30 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Fair ML in Credit Scoring

pipeline

Code and results accompanying the paper:

Kozodoi, N., Jacob, J., Lessmann, S. (2021).
Fairness in Credit Scoring: Assessment, Implementation and Profit Implications.
European Journal of Operational Research, 297, 1083-1094.

The paper is available at the publisher's website and on arXiv.

Note: the PDF on the publisher's website contains a typo in Equation 5 (the group subscripts in the last term should be swapped). Please refer to the arXiv preprint and/or the code in this repo for the correct version.

Summary

The rise of algorithmic decision-making has spawned much research on fair machine learning (ML). Financial institutions use ML for building risk scorecards that support a range of credit-related decisions. The project makes three contributions:

  1. Revisiting statistical fairness criteria and examine their adequacy for credit scoring.
  2. Cataloging algorithmic options for incorporating fairness goals in the ML model development pipeline.
  3. Empirically comparing multiple fairness processors in a profit-oriented credit scoring context using real-world data.

The empirical results substantiate the evaluation of fairness measures, identify suitable options to implement fair credit scoring, and clarify the profit-fairness trade-off in lending decisions. We find that multiple fairness criteria can be approximately satisfied at once and recommend separation as a proper criterion for measuring scorecard fairness. We also find fair in-processors to deliver a good profit-fairness balance and show that algorithmic discrimination can be reduced to a reasonable level at a relatively low cost.

results

This repo contains codes and files that allow reproducing the results presented in the paper, performing additional analyses and extending the experimental setup. It provides implementations of eight fairness processors used in the paper. Further details on the modeling pipeline are provided in the paper as well as in the codes.

Repo structure

The repo has the following structure:

  • codes/: Python notebooks and R codes implementing data processing and fair ML algorithms
  • data/: raw data sets and processed data exported from the data processing scripts
  • functions/: helper functions accompanying the codes, including a customized aif360 Python module
  • output/: figures and tables with results exported from code_14_results.R and presented in the paper
  • results/: intermediate and final results files exported from the different modeling codes

Further details on the codes are provided in a separate README file in the codes folder.

Data sets

The experiments are performed on seven credit scoring data sets obtained from different sources, including the UCI Machine Learning Repository, Kaggle and PAKDD platforms.

To respect the terms of use of the corresponding data sets and adhere to the file size limit, the raw data is not included in the repo. The full list of data sets with the download links and instructions is provided in a separate README file in the data folder.

Fair ML methods

The paper implements and banchmarks the folloeing fair ML algorithms:

  1. Pre-processors:
  1. In-processors:
  1. Post-processors:

Working with the repo

Dependencies

The codes require Python 3.7+ and R 4.0+. Relevant R packages are installed separately at the beginning of each R script. Packages required to implement each of the considered fairness processors are listed and imported in the corresponding Python scripts.

To work with the Python notebooks, we recommend to create a virtual Conda environment from the supplied environment.yml file:

conda env create --name fairness --file environment.yml
conda activate fairness

Reproducing results

Reproducing results presented in the paper and carrying out additional analysis involves the following steps:

  1. Update working paths in code_00_working_paths.R.
  2. Execute code_14_results.R to perform analysis of the results and save outputs in output/.

Extending experimental setup

Modifying the experimental setup requires the following:

  1. Place the raw data sets in the data/raw/ folder (see instructions).
  2. Update working paths in code_00_working_paths.R.
  3. Execute code_01_data_partitioning.ipynb to process and partition each data set.
  4. Adjust experimental parameters specified in functions/.
  5. Execute the modeling codes according to their name order for each of the data sets.
  6. Execute code_14_results.R to perform analysis of the results and save outputs in output/.

About

Fair ML in credit scoring: Assessment, implementation and profit implications

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published