Fairness in Machine Learning: A Review for Statisticians

With the widespread application of machine learning algorithms in daily life, it is crucial to mitigate the risk of these algorithms producing socially undesirable outcomes that may disproportionately disadvantage certain groups or individuals based on demographic characteristics such as gender, race, or disabilities. In recent years, machine learning fairness has gained increasing attention from both researchers and the public. This article provides a comprehensive overview of fairness-enhancing mechanisms designed to mitigate such risks, along with the fairness criteria they aim to achieve. We organize these fairness-enhancing mechanisms into three categories—pre-processing, in-processing, and post-processing—corresponding to different stages of the machine learning lifecycle and varying levels of access to the underlying algorithm. The discussion focuses on fairness in binary classification models using numerical tabular data, which serve as a foundation for addressing fairness in more complex algorithms. Additionally, we present experimental results that offer a comparative evaluation of representative fairness-enhancing approaches.