This repository implements different fairness algorithms initally shown by the aif360 project[1]. It uses different data sets and different approaches to display potential biases in dataset. It is largly inspired by IBM. Generally speaking there are three possibilities to correct a bias in a dataset pre, inter, and post. The majority of the approaches used in this repository focus on the preprocessing step as it looked most promising.

Further in order to achieve a better understanding I highly recommend the aif360 repository. Especially the guides and tutorials offered: https://github.com/Trusted-AI/AIF360/tree/master/examples. Most of the following code is inspired by these.

• notebooks
  • adult Contains two different notebooks varying in their protected attribute. Both use the aif360 classifier metric.
  • compas Contains 3 different notebooks relevant for fairness:
    • lr_classifier: Classifier metric with linear regression classifier
    • lr_pipeline: Simple pipeline for a linear regression model
    • sklearn_classifier: Instead of using custom aif360 functions this approach uses standarized methods (pandas) to improve the handling and processing of the fairness data.
  • credit score Contains three notebooks of which two differ in their protected attribute. Whereas, sexclassifier.ipynb and sexbinary.ipynb vary in their classifier approach.
• aif360-Fork. I would advise to check the commit messages for further insides. I added some simple helper functions to improve procedures.

• The documentation of aif360 provides great detail about what is possible.
• At the beginning using the binary approach might be easier, but it is limited in functionality. Thus, I recommend using a proper classifier.
• I was not able to produce the exact same results as IBM, but the results of the compas notebooks came very close.
• The original paper from IBM

[1] Bellamy, R. K., Dey, K., Hind, M., Hoffman, S. C., Houde, S., Kannan, K., ... & Zhang, Y. (2018). AI Fairness 360: An extensible toolkit for detecting, understanding, and mitigating unwanted algorithmic bias. arXiv preprint arXiv:1810.01943.