Fairness vs Performance: Characterizing the Pareto Frontier of Algorithmic Decision Systems

A new academic paper submitted to arXiv by researchers Mieke Wilms and Christoph Heitz investigates the critical trade-off between model performance and fairness in algorithmic decision systems. The study conceptualizes binary prediction-based decision-making as a multi-objective optimization problem, simultaneously considering decision-maker utility and group fairness. The authors characterize the Pareto frontier, revealing that it consists of deterministic, group-specific threshold rules applied to individuals' success probabilities. Contrary to existing literature that primarily posits lower-bound threshold rules, this research demonstrates that the Pareto frontier may also include upper-bound thresholds, potentially favoring individuals with lower success probabilities depending on the fairness metric used. Crucially, the findings indicate that the location of this frontier depends solely on population characteristics, utility functions, and fairness scores, remaining independent of the algorithm's technical design, whether pre-, in-, or post-processing. These results generalize previous optimality theorems and provide a principled foundation for evaluating algorithmic systems, bridging formal fairness research with legal and ethical requirements for identifying less discriminatory alternatives in automated decision-making processes.