Fairlearn Parity Constraints for Mitigating Gender Bias in Binary Classification Models – Comparative Analysis

Inequality is one of the problems of the modern world. Discrimination of various kinds can affect many areas of life. The growing importance of data in the modern world makes it all the more important to ensure that the methods used to analyze it do not return results in which unfairness is present. Unfortunately, there may be situations where there is unfairness in the predictions of machine learning models. In recent years, several IT solutions have been developed to mitigate this phenomenon. One of them is Fairlearn, a Python library dedicated to this type of task. This article presents a comparative analysis of parity constraints used in Fairlearn algorithms. The purpose of this article is to identify which of the constraints is best suited for mitigating gender bias in binary classification models. The following research methods were used: literature review, experiment and comparative analysis. The evaluation of constraints will be based on the value of measures: disparity in recall and disparity in selection rate for the column containing information about the person’s gender. The values of these measures, achieved by binary classification models in which the Threshold Optimizer algorithm with selected parity constraints was implemented, will be compared in order to identify which of the Fairlearn parity constraints is best suited for mitigating gender bias in binary classification models.