Application of techniques for achieving fair machine learning models

Abstract

Machine learning models are becoming more and more present in everyday life and various industries. People make decisions based on models who learn from available data. The data describe human behavior that has often had a discriminatory character towards certain groups in society. Today, we are witnessing attempts to eliminate discrimination wherever possible, but as machine learning models learn from historical data, it happens that models learn to discriminate against those groups. Therefore, it becomes very important to make a fair model. However, it is certainly not enough for the model to be just fair. The main goal remains to make an accurate model. The challenge lies in finding a compromise between the accuracy and fairness of the model. This paper will focus on answering the question of what is the cost of increasing the fairness of a machine learning model. The paper will provide an overview of the currently most important techniques for achieving fair machine learning models, evaluation methods as well as their costs to explain the trade-off between the accuracy and fairness of models.