Abstract
Traditional algorithmic fairness often relies on rigid mathematical definitions, such as equalizing error rates across demographic groups. However, fairness is frequently a nuanced, subjective concept that varies across contexts and stakeholders. This paper introduces a framework for "Fairness Elicitation," where fairness constraints are derived from pairwise comparisons provided by stakeholders. We propose a provably convergent and oracle-efficient algorithm designed to learn the most accurate model possible while adhering to these elicited constraints. Our approach allows for the integration of subjective human values with traditional statistical metrics. We demonstrate the framework's efficacy through a behavioral study using human-subject constraints on the COMPAS criminal recidivism dataset, exploring the trade-offs between accuracy, individual preferences, and collective fairness.
Introduction
As machine learning models increasingly influence critical life outcomes—from credit scoring to criminal justice—the question of "fairness" has moved to the forefront of research. Most existing approaches impose "top-down" mathematical constraints, such as demographic parity or equalized odds. While computationally convenient, these definitions often fail to capture the complexity of real-world equity and ignore a fundamental question: Who gets to define what is fair?
This research proposes a shift toward a participatory, "bottom-up" approach. By treating fairness as a property to be elicited from stakeholders rather than a formula to be dictated by researchers, the framework bridges the gap between technical implementation and the institutional realities of the people affected by automated decisions.
The Fairness Elicitation Framework
The core of the proposed framework is the translation of subjective values into pairwise constraints. Instead of asking a stakeholder to articulate a universal policy, the system presents them with pairs of individuals $(A, B)$ and asks for a relative judgment:
- Equal Treatment: $A$ and $B$ must be treated the same.
- Ordered Treatment: $A$ should be treated at least as well as $B$ (or vice versa).
- No Constraint: The algorithm is free to treat them differently based on the data.
This method leverages the psychological insight that humans are often better at identifying unfairness in specific instances than articulating abstract moral rules.
Balancing Accuracy and Compliance
To account for the inherent tension between predictive accuracy and human preferences, the framework introduces two relaxation parameters:
- $\gamma$ (Gamma): A "buffer" that allows for minor differences in treatment between two individuals before a violation is recorded.
- $\eta$ (Eta): A "violation budget" that limits the total allowed dissatisfaction across all pairs and stakeholders.
Algorithmic Approach
The learning problem is framed as a zero-sum game between two competing agents:
- The Learner (Primal Player): Seeks to minimize classification error.
- The Auditor (Dual Player): Identifies where fairness constraints are being violated and imposes a "penalty" or cost on the learner for those specific instances.
By reducing this game to a series of Cost-Sensitive Classification problems, the algorithm iteratively converges to an approximate equilibrium. This results in a model that maximizes accuracy within the bounds of the elicited fairness budget. The authors prove that this algorithm is "oracle-efficient," meaning its complexity is comparable to standard unconstrained machine learning.
Behavioral Study: The COMPAS Dataset
The researchers evaluated their framework using the COMPAS criminal recidivism dataset, a standard benchmark in fairness literature. They conducted a behavioral study with 43 human subjects, each providing judgments on 50 random pairs of defendants.
Key Findings
- High Subjectivity: The study revealed significant variability in how individuals define fairness. Some participants' constraints were easily satisfied, while others imposed a sharp trade-off with model accuracy.
- The "Difficult Few": Much of the conflict between accuracy and fairness was driven by a small subset of human-provided constraints, suggesting that a few highly specific cases often dictate the "cost" of fairness.
- Alignment with Statistical Metrics: Interestingly, subjective individual constraints often naturally led to a reduction in the False Positive Rate gap between Black and White defendants, suggesting that human intuition can sometimes—but not always—align with popular statistical fairness definitions.
- Fairness vs. Justice: The authors distinguish between justice (the absolute correctness of an outcome) and fairness (the relative consistency of treatment). Their framework focuses on the latter, ensuring that similar individuals are treated with the consistency required by stakeholders.
Conclusion and Future Directions
This work provides a theoretically sound and empirically grounded method for incorporating human values into machine learning. By proving generalization bounds, the authors show that fairness constraints learned from a small sample of comparisons are likely to hold on unseen data.
Future research will need to address the challenges of stakeholder selection—ensuring that those most affected by a system have a voice in its constraints—and aggregation, developing methods to reconcile conflicting views of fairness from diverse groups. Additionally, "active learning" techniques could be employed to identify the most informative pairs for comparison, making the elicitation process more scalable for real-world applications.
Summary updated