Fairness with Overlapping Groups; a Probabilistic Perspective

Forest Yang, Mouhamadou Cisse, Oluwasanmi KoyejoNeurIPS 2020 · OpenReview

Achieving algorithmic fairness is a complex challenge, particularly when individuals belong to multiple, overlapping demographic groups. This paper introduces a unified probabilistic framework designed to ensure fairness across these intersecting identities—such as race, gender, and age—simultaneously. By identifying the Bayes-optimal classifier for these complex constraints, the authors present a statistically consistent algorithm that outperforms traditional regularization methods and handles "fairness gerrymandering" more efficiently than previous state-of-the-art approaches.

The Challenge of Fairness Gerrymandering

Most algorithmic fairness research focuses on a simplified scenario: ensuring parity between two mutually exclusive groups (e.g., Group A vs. Group B). However, real-world identity is intersectional. A model might appear fair when evaluating "Race" and "Gender" as independent attributes, yet still discriminate against specific subgroups, such as Black women. This phenomenon is known as fairness gerrymandering.

Diagram illustrating overlapping demographic groups and their intersections

The primary hurdle in addressing this is the "curse of dimensionality." Enforcing fairness across every possible intersection of attributes quickly leads to data scarcity—some intersections may have zero samples—and computational explosions. This research bridges the gap between independent group fairness (which is easier to compute but prone to gerrymandering) and intersectional fairness (which is ideal but statistically difficult) by providing a rigorous mathematical foundation for overlapping group constraints.

A Probabilistic Approach to Overlapping Constraints

The authors tackle the problem through a probabilistic population analysis, characterizing the "Bayes-optimal" classifier—the theoretical best-case predictor—under fairness constraints. They demonstrate that even when groups overlap, the optimal fair classifier can be decomposed into a combination of weighted classifiers based on intersectional groups.

Key components of their methodology include:

  • Lagrangian Formulation: Fairness is treated as a constrained optimization problem. Lagrange multipliers balance the trade-off between prediction accuracy and fairness violations.
  • GroupFair Algorithm: A general-purpose procedure that finds a "saddle point" in the Lagrangian, iterating between updating a classifier and adjusting fairness penalties.
  • Two Oracle Flavors:
    1. Plugin Oracle: Estimates conditional class probabilities first, then applies a threshold. It is computationally highly efficient.
  • Weighted ERM (Empirical Risk Minimization): Solves a cost-sensitive classification problem in each round, offering more robustness when underlying probability estimates are difficult to ascertain.

The authors provide theoretical guarantees that their empirical estimators will converge to the population-optimal fair classifier as more data becomes available.

Key Results and Performance Trade-offs

The methods were evaluated on four standard fairness benchmarks: Communities and Crime, Adult Census, German Credit, and Law School datasets. These datasets contain multiple sensitive attributes, allowing for rigorous testing of overlapping constraints.

Test frontiers showing classification error versus fairness violation across multiple datasets

The results highlight several significant advantages of the proposed framework:

  • Superior Trade-offs: In "fairness frontier" evaluations (mapping error rate against fairness violation), the proposed methods (WERM and Plugin) consistently achieved better accuracy for the same level of fairness compared to baseline regularizers. Simple regularization baselines often failed to reach the optimal fairness frontier entirely.
  • Computational Efficiency: Previous methods for handling gerrymandering often relied on brute-force enumeration or "fictitious play." The GroupFair approach, particularly the Plugin version, proved to be 1–2 orders of magnitude faster. For instance, on the Adult dataset under gerrymandering constraints, the GroupFair-Plugin completed in roughly 700 seconds, compared to over 2200 seconds for prior baseline methods.
  • Metric Flexibility: The framework successfully handles any fairness metric that can be expressed as a linear function of the confusion matrix, such as Demographic Parity or Equal Opportunity.

Limitations and Future Directions

While the framework offers a robust solution to overlapping group fairness, it is not without limitations. The statistical bounds depend on the size of the smallest group; if a specific intersectional subgroup is extremely small, the fairness guarantees weaken. Additionally, the current work focuses on fairness metrics that are linear functions of the confusion matrix and is built specifically for multiclass classification. Extending these insights to more complex fractional metrics, as well as to regression or ranking problems, remains an exciting open challenge for future research.

Summary updated

Abstract

In algorithmically fair prediction problems, a standard goal is to ensure the equality of fairness metrics across multiple overlapping groups simultaneously. We reconsider this standard fair classification problem using a probabilistic population analysis, which, in turn, reveals the Bayes-optimal classifier. Our approach unifies a variety of existing group-fair classification methods and enables extensions to a wide range of non-decomposable multiclass performance metrics and fairness measures. The Bayes-optimal classifier further inspires consistent procedures for algorithmically fair classification with overlapping groups. On a variety of real datasets, the proposed approach outperforms baselines in terms of its fairness-performance tradeoff.
BibTeX
@inproceedings{yang2020fairness,
  author = {Forest Yang and Mouhamadou Cisse and Oluwasanmi Koyejo},
  title = {{Fairness with Overlapping Groups; a Probabilistic Perspective}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2020},
  url = {https://openreview.net/forum?id=VWy-7AvS5s9},
}
Figure