Artificial Intelligence systems add significant value to decision-making. However, the systems must be fair because bias creeps into the system from sources like data and preprocessing algorithms. In this work, we explore fairness metrics discussing the shortfalls and benefits of each metric. The fairness metrics are demographic, statistical, and game theoretic. We find that the demographic fairness metrics are independent of the actual target value and hence have limited use. In contrast, the statistical fairness metrics can provide the thresholds to maximize fairness. The Minimax criterion was used to guide the search and help recommend the best model where the error among protected groups was minimum.
CITATION STYLE
Irfan, Z., McCaffery, F., & Loughran, R. (2023). Evaluating Fairness Metrics. In Communications in Computer and Information Science (Vol. 1840 CCIS, pp. 31–41). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-031-37249-0_3
Mendeley helps you to discover research relevant for your work.