[Paper Review] Auditing and Achieving Intersectional Fairness in Classification Problems

Giulio Morina, Viktoriia Oliinyk+3 more|arXiv (Cornell University)|Nov 4, 2019

Ethics and Social Impacts of AI32 references17 citations

TL;DR

This paper presents a comprehensive framework for auditing and achieving intersectional fairness in binary classification by extending fairness metrics to multiple sensitive attributes, developing robust estimation techniques for underrepresented subgroups, and introducing post-processing methods that preserve predictive performance while guaranteeing fairness. The approach ensures intersectional fairness across groups like 'Black women' or 'older Hispanic men' without relying on model assumptions.

ABSTRACT

Machine learning algorithms are extensively used to make increasingly more consequential decisions about people, so achieving optimal predictive performance can no longer be the only focus. A particularly important consideration is fairness with respect to race, gender, or any other sensitive attribute. This paper studies intersectional fairness, where intersections of multiple sensitive attributes are considered. Prior research has mainly focused on fairness with respect to a single sensitive attribute, with intersectional fairness being comparatively less studied despite its critical importance for the safety of modern machine learning systems. We present a comprehensive framework for auditing and achieving intersectional fairness in classification problems: we define a suite of metrics to assess intersectional fairness in the data or model outputs by extending known single-attribute fairness metrics, and propose methods for robustly estimating them even when some intersectional subgroups are underrepresented. Furthermore, we develop post-processing techniques to mitigate any detected intersectional bias in a classification model. Our techniques do not rely on any assumptions regarding the underlying model and preserve predictive performance at a guaranteed level of fairness. Finally, we give guidance on a practical implementation, showing how the proposed methods perform on a real-world dataset.

Motivation & Objective

• To address the critical gap in algorithmic fairness research by focusing on intersectional fairness, where multiple sensitive attributes (e.g., race and gender) intersect.
• To develop fairness metrics that assess bias not just per individual attribute, but across combinations of attributes such as 'Black women' or 'older Hispanic men'.
• To provide robust estimation techniques for fairness metrics in data with underrepresented intersectional subgroups due to data collection biases.
• To design post-processing methods that mitigate detected intersectional bias while maintaining high predictive performance.
• To demonstrate the framework’s practical utility through experiments on the Adult Income dataset and a synthetic dataset.

Proposed method

• Extends single-attribute fairness metrics—such as statistical parity, equal opportunity, and equalized odds—to the intersectional setting by defining metrics for data (lift and impact ratio) and model outputs (equal opportunity and equalized odds).
• Proposes three robust estimation techniques: smoothed empirical, bootstrap, and Bayesian estimation, to improve confidence in fairness estimates, especially for rare subgroups.
• Frames post-processing as an optimization problem that adjusts decision thresholds and prediction probabilities per intersectional subgroup to satisfy fairness constraints.
• Introduces three post-processing strategies: deterministic (threshold-only), sequential (randomization on top of optimal scores), and overall (joint optimization of thresholds and probabilities).
• Uses a fairness constraint based on ε-differential fairness, ensuring that the odds ratio of true positive rates across subgroups remains within a predefined bound (ε ≤ 2.15).
• Employs a baseline 'optimal score model' to compare predictive performance before and after post-processing, ensuring minimal performance degradation.

Experimental results

Research questions

• RQ1 How can fairness metrics be meaningfully extended to capture intersectional disparities across multiple sensitive attributes in classification models?
• RQ2 What estimation techniques can reliably measure intersectional fairness when some subgroups are severely underrepresented in the data?
• RQ3 Can post-processing techniques achieve guaranteed intersectional fairness without sacrificing predictive performance?
• RQ4 How do different post-processing strategies (deterministic, sequential, overall) compare in terms of fairness and performance trade-offs?
• RQ5 To what extent can robust estimation methods detect and quantify intersectional bias in real-world datasets like Adult?

Key findings

• All post-processed models achieved the target ε-differential fairness constraint of approximately 2.15 for equalized odds, as confirmed by smoothed empirical, bootstrap, and Bayesian estimators.
• The 'sequential' and 'overall' post-processing methods performed nearly as well as the optimal score model in terms of predictive performance, with minimal loss in accuracy.
• The 'deterministic' post-processed model incurred a significantly higher expected loss compared to the other post-processing strategies, indicating suboptimal trade-offs.
• Randomization applied after the optimal score model improved performance on a small, underrepresented subgroup, demonstrating the method’s ability to correct for model bias.
• The Bayesian and bootstrap estimators provided reliable 95% confidence intervals that contained the target fairness threshold, validating the robustness of the estimation framework.
• The framework successfully identified and mitigated intersectional bias in the Adult dataset, particularly for subgroups like 'Black women' and 'older Hispanic men', which are often marginalized in standard fairness evaluations.