Understanding the Impossibility of Machine Learning Fairness with Data Examples

Abstract

This review paper examines how the three common criteria of machine learning fairness, despite their common sense and moral appeal, are often mutually exclusive. This frequently presents a challenge and the need to prioritize one criterion above the others. Specifically, the paper highlights the general overview of a machine learning model and the historical aspect that can play a role in the existing biases in models today. The paper then dives into the three criteria of machine learning fairness: independence, separation, and sufficiency. It explains the conflict between them and how it is impossible to satisfy all three conditions simultaneously, creating an imperfect machine-learning model. This paper illustrates how this comes up and plays out with real problems with real data examples and code.