Fairness-Utility Trade-off via Wasserstein Projection

2025-05-16

AI Summary

This paper addresses binary intervention decision-making under causal fairness, aiming to jointly optimize demographic-group fairness and overall utility. We propose a Wasserstein-projection-based propensity score calibration method that maximizes total utility subject to an ε-tolerant demographic parity constraint, enabling tunable fairness–utility trade-offs. To our knowledge, this is the first work to incorporate the Wasserstein distance into a utility-constrained fairness evaluation framework, providing rigorous theoretical guarantees and a statistically testable hypothesis-testing mechanism for transparent trade-off analysis. Empirical evaluation across multiple benchmark datasets demonstrates that our method achieves strict fairness compliance while preserving at least 95% of the original utility; moreover, it enables policymakers to quantify societal costs and performance degradation in a principled manner.

Technology Category

• Philosophy and Ethics of AI: Bias, Fairness & Equity
• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Decision/Utility Theory

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: Fairness, privacy, and diversity in economic environments
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

Ensuring fairness in data-driven decision-making is a critical concern, but existing fairness constraints often involve trade-offs with overall utility. We propose a fairness framework that enforces strong demographic parity-related fairness criteria (with $epsilon$-tolerance) in propensity score allocation while guaranteeing a minimum total utility. This approach balances equity and utility by calibrating propensity scores to satisfy fairness criteria and optimizing outcomes without incurring unacceptable losses in performance. Grounded in a binary treatment and sensitive attribute setting under causal fairness setup, our method provides a principled mechanism to address fairness while transparently managing associated economic and social costs, offering a practical approach for designing equitable policies in diverse decision-making contexts. Building on this, we provide theoretical guarantee for our proposed utility-constrained fairness evaluation framework, and we formalize a hypothesis testing framework to help practitioners assess whether the desired fairness-utility trade-off is achieved.

Problem

• Balancing fairness and utility in decision-making
• Enforcing demographic parity with minimal utility loss
• Providing theoretical guarantees for fairness-utility trade-offs

Innovation

• Wasserstein projection enforces demographic parity fairness
• Calibrates propensity scores to balance equity and utility
• Provides theoretical guarantees for fairness-utility trade-off