A Review on Pre-processing Methods for Fairness in Machine Learning

Abstract

AbstractWith the development of artificial intelligence (AI) technology, the application of machine learning (ML) algorithms has become more extensive, and AI algorithms have begun to make decisions in some important fields (finance, law, and medical health). However, studies have shown that due to social, historical, and other factors, the data for training machine learning algorithms already contain human biases, so machine learning algorithms will learn or even amplify these biases, resulting in unfair decision-making. There have been many studies on fairness in machine learning, including how to define and measure fairness and enhance fairness in ML. The existing means of lightening bias in ML can be classified into three types which are pre-processing, in-processing, and post-processing, according to the life cycle of ML. In this paper, we survey the pre-processing techniques and summarize them according to different categories. At the same time, we also introduce commonly used fairness measures to study fairness.KeywordsMachine learningFairness in machine learningPre-processingAlgorithmic bias