Fairness-Aware Post-Processing in Supervised Classification: L1/L2 Norm and Optimal Swapping Methods

Abstract

The growing use of AI in decision-making raises ethical concerns, including fairness in classification tasks. In this paper, we propose two fairness-aware post-processing methods — (1) the least L1/L2 norm with covariance constraints and (2) the optimal swapping — which address group fairness in the outputs of any probabilistic classifier. The first method incorporates fairness constraints and minimizes deviations from the classifier's initial probabilistic predictions through an interpolation between the L1 and L2 norms. The second method is a heuristic approach that directly modifies binary classification decisions through a simple, yet efficient swapping approach. We evaluate our methods on five benchmark datasets and compare them with some well-established baselines. Our results show that optimal swapping achieves the best trade-off between fairness and accuracy on the investigated datasets. While we focus on demographic parity and disparate impact, our swapping post-processing method is adaptable to multi-class settings, some other fairness definitions, and allows for additional linear constraints.