Abstract
While the accuracy-fairness trade-off has been frequently observed in the literature of fair machine learning, rigorous theoretical analyses have been scarce. To demystify this long-standing challenge, this work seeks to develop a theoretical framework by characterizing the shape of the accuracy-fairness trade-off Pareto frontier (FairFrontier), determined by a set of all optimal Pareto classifiers that no other classifiers can dominate. Specifically, we first demonstrate the existence of the trade-off in real-world scenarios and then propose four potential categories to characterize the important properties of the accuracy-fairness Pareto frontier. For each category, we identify the necessary conditions that lead to corresponding trade-offs. Experimental results on synthetic data suggest insightful findings of the proposed framework: (1) When sensitive attributes can be fully interpreted by non-sensitive attributes, FairFrontier is mostly continuous. (2) Accuracy can suffer a sharp decline when over-pursuing fairness. (3) Eliminate the trade-off via a two-step streamlined approach. The proposed research enables an in-depth understanding of the accuracy-fairness trade-off, pushing current fair machine-learning research to a new frontier.
Introduction
In the field of fair machine learning, a persistent tension exists between model accuracy and fairness. Practitioners often find that increasing the fairness of a model necessitates a drop in its predictive accuracy. However, much of this understanding has remained empirical, derived from trial-and-error on specific datasets. This research addresses a fundamental theoretical gap: what is the actual shape of this trade-off curve, known as the FairFrontier? By characterizing this frontier, the authors provide a roadmap for determining whether a model is truly optimal or if a superior balance of accuracy and fairness is achievable.
The FairFrontier Taxonomy
The researchers utilize Bayesian optimal classifiers to analyze the FairFrontier, focusing on the positioning of decision boundaries. They categorize the FairFrontier into four distinct types based on their geometric properties:
- Continuous: A smooth trade-off where small gains in fairness result in small losses in accuracy.
- Sharp Decline in Accuracy: A catastrophic drop in accuracy occurs when attempting to reach high fairness levels.
- Sharp Decline in Fairness: Small accuracy improvements lead to large fairness penalties.
- Sharp Decline in Both: Abrupt changes in both metrics simultaneously.
The existence of these sharp declines suggests that the FairFrontier is often non-convex, challenging the common assumption in fair ML that these trade-offs are well-behaved and easily optimized.
Unfairness Decomposition and the "Two-Step Cure"
A key methodological contribution of this work is the decomposition of total unfairness ($F_U$) into two components:
- Data Unfairness ($F_{DU}$): Disparities inherent in the underlying data distribution.
- Model Unfairness ($F_{MU}$): Disparities introduced by how the model aligns decision boundaries across different groups.
This decomposition allows for a "two-step cure" to the trade-off. First, practitioners can tackle data unfairness through augmentation or re-sampling. Second, they can address model unfairness by transforming data to better align decision boundaries across sensitive groups.
Key Insights: The "Sharp Decline" Warning
The theoretical framework identifies the "necessary conditions" that trigger specific frontier shapes. A critical finding is the warning against "over-pursuing" fairness. The paper proves that accuracy can suffer a sharp decline when fairness constraints force the model into a decision boundary that is highly unfavorable for a specific group, effectively "trivializing" the classifier for that segment of the population. Conversely, in "ideal" cases where non-sensitive features fully capture the information in sensitive attributes, the FairFrontier remains mostly continuous, allowing for a more predictable trade-off.
Conclusion
By establishing the mathematical existence of the accuracy-fairness trade-off (Theorem 1) and defining an accuracy upper bound under fairness constraints (Theorem 2), this work provides a rigorous foundation for future fair ML research. While currently validated on synthetic distributions, the framework offers a powerful lens for understanding why models fail to be both fair and accurate and provides a path toward eliminating these trade-offs through targeted data and model interventions.
Summary updated