[Paper Review] Achieving Equalized Odds by Resampling Sensitive Attributes

Yaniv Romano, Stephen Bates+1 more|arXiv (Cornell University)|Jan 1, 2020

Adversarial Robustness in Machine Learning2 citations

TL;DR

This paper proposes a differentiable framework that enforces equalized odds fairness in machine learning models by penalizing violations through a resampled sensitive attribute. By constructing a discrepancy functional and using resampling to ensure fairness, the method improves model performance and enables rigorous hypothesis testing for fairness, with applications in regression and multi-class classification.

ABSTRACT

We present a flexible framework for learning predictive models that approximately satisfy the equalized odds notion of fairness. This is achieved by introducing a general discrepancy functional that rigorously quantifies violations of this criterion. This differentiable functional is used as a penalty driving the model parameters towards equalized odds. To rigorously evaluate fitted models, we develop a formal hypothesis test to detect whether a prediction rule violates this property, the first such test in the literature. Both the model fitting and hypothesis testing leverage a resampled version of the sensitive attribute obeying equalized odds, by construction. We demonstrate the applicability and validity of the proposed framework both in regression and multi-class classification problems, reporting improved performance over state-of-the-art methods. Lastly, we show how to incorporate techniques for equitable uncertainty quantification---unbiased for each group under study---to communicate the results of the data analysis in exact terms.

Motivation & Objective

• To address the challenge of ensuring equalized odds fairness in predictive models, especially in scenarios with sensitive attributes like race or gender.
• To develop a differentiable discrepancy functional that quantifies violations of equalized odds, enabling optimization-based fairness improvement.
• To create a formal hypothesis test to rigorously detect whether a model violates equalized odds, a first in the literature.
• To integrate equitable uncertainty quantification techniques that are unbiased across demographic groups, ensuring transparent and fair model interpretation.
• To demonstrate the framework's effectiveness across regression and multi-class classification tasks, outperforming state-of-the-art methods.

Proposed method

• Introduce a differentiable discrepancy functional that measures the degree to which a model's predictions violate equalized odds across sensitive groups.
• Use resampling of the sensitive attribute to construct a proxy that inherently satisfies equalized odds, which is then used in both model training and hypothesis testing.
• Apply the discrepancy functional as a penalty term in the model's objective function to drive parameters toward equalized odds compliance.
• Leverage the resampled sensitive attribute in a formal statistical hypothesis test to detect violations of equalized odds with controlled Type I error rates.
• Integrate equitable uncertainty quantification methods that ensure prediction intervals or credible intervals are unbiased across groups.
• Train models using standard optimization techniques with the penalty term, enabling end-to-end learning while maintaining fairness constraints.

Experimental results

Research questions

• RQ1 Can a differentiable, optimization-based framework be designed to enforce equalized odds fairness in machine learning models?
• RQ2 How can a resampled version of the sensitive attribute be used to both train fair models and test for fairness violations?
• RQ3 Can a formal statistical hypothesis test detect violations of equalized odds with rigorous Type I error control?
• RQ4 Does the proposed method improve model performance compared to state-of-the-art fairness approaches in regression and multi-class classification?
• RQ5 Can equitable uncertainty quantification be integrated to ensure fair interpretation of model predictions across demographic groups?

Key findings

• The proposed framework successfully enforces equalized odds by using a differentiable discrepancy functional that penalizes fairness violations during model training.
• The use of resampled sensitive attributes ensures that both training and hypothesis testing are grounded in a construct that satisfies equalized odds by design.
• The first formal hypothesis test for equalized odds violations is developed, enabling rigorous statistical evaluation of fairness in predictive models.
• Empirical results show improved performance over state-of-the-art methods in both regression and multi-class classification tasks.
• Equitable uncertainty quantification techniques are successfully incorporated, ensuring that prediction intervals are unbiased across demographic groups.
• The framework demonstrates robustness and applicability across diverse machine learning problems while maintaining fairness and interpretability.