[Paper Review] Empirical observation of negligible fairness-accuracy trade-offs in machine learning for public policy

Kit T. Rodolfa, Hemank Lamba+1 more|arXiv (Cornell University)|Dec 5, 2020

Ethics and Social Impacts of AI37 references83 citations

TL;DR

This paper empirically investigates fairness-accuracy trade-offs in machine learning for public policy, demonstrating that post-hoc disparity mitigation using subgroup-specific score thresholds can significantly improve fairness—particularly in recall—without sacrificing model precision in top-k selections. Across four real-world policy domains—mental health, housing safety, education, and crowdfunding—fairness improvements were achieved with negligible accuracy loss, challenging the assumption that fairness requires complex methods or performance trade-offs.

ABSTRACT

Growing use of machine learning in policy and social impact settings have raised concerns for fairness implications, especially for racial minorities. These concerns have generated considerable interest among machine learning and artificial intelligence researchers, who have developed new methods and established theoretical bounds for improving fairness, focusing on the source data, regularization and model training, or post-hoc adjustments to model scores. However, little work has studied the practical trade-offs between fairness and accuracy in real-world settings to understand how these bounds and methods translate into policy choices and impact on society. Our empirical study fills this gap by investigating the impact of mitigating disparities on accuracy, focusing on the common context of using machine learning to inform benefit allocation in resource-constrained programs across education, mental health, criminal justice, and housing safety. Here we describe applied work in which we find fairness-accuracy trade-offs to be negligible in practice. In each setting studied, explicitly focusing on achieving equity and using our proposed post-hoc disparity mitigation methods, fairness was substantially improved without sacrificing accuracy. This observation was robust across policy contexts studied, scale of resources available for intervention, time, and relative size of the protected groups. These empirical results challenge a commonly held assumption that reducing disparities either requires accepting an appreciable drop in accuracy or the development of novel, complex methods, making reducing disparities in these applications more practical.

Motivation & Objective

• To investigate whether fairness-accuracy trade-offs exist in real-world public policy applications of machine learning.
• To evaluate the practical impact of post-hoc disparity mitigation methods on fairness and accuracy in resource-constrained settings.
• To assess whether fairness improvements can be achieved without significant sacrifice in model precision for top-k interventions.
• To examine the robustness of fairness-accuracy results across diverse policy contexts, group sizes, and resource levels.

Proposed method

• Applied post-hoc disparity mitigation using subgroup-specific score thresholds to balance recall across protected groups.
• Defined fairness using recall disparity (equality of opportunity), with precision in the top-k as the primary accuracy metric.
• Used a group-specific scaling approach to adjust predicted scores, preserving within-group ranking while correcting for imbalance.
• Conducted model training and hyperparameter tuning via grid search across diverse estimators (e.g., Random Forest, Logistic Regression) in each policy context.
• Evaluated performance across temporal validation splits and bootstrap samples to ensure robustness.
• Selected top-k lists based on resource constraints (e.g., 500 individuals, 1,000 projects), measuring precision and recall disparity.

Experimental results

Research questions

• RQ1 Do fairness-accuracy trade-offs exist in real-world public policy applications of machine learning?
• RQ2 Can post-hoc mitigation using subgroup-specific thresholds significantly improve fairness without degrading top-k precision?
• RQ3 How robust are fairness and accuracy improvements across different policy domains and protected group sizes?
• RQ4 Does model selection that accounts for fairness during training yield better outcomes than post-hoc adjustment?
• RQ5 Can simple, scalable methods achieve fairness improvements without complex retraining or architectural changes?

Key findings

• In the Inmate Mental Health setting, recall disparity was reduced by over 50% with only a 1.2-percentage-point drop in precision at the top 500.
• In the Housing Safety context, fairness improvements were achieved with less than 0.5% decrease in top-100 precision, even when disparities favored higher-income neighborhoods.
• For Education Crowdfunding, fairness was enhanced across gender and race subgroups with no measurable drop in precision at the top 1,000 projects.
• In Student Outcomes, over-age students—previously underserved—saw improved recall equity with only a 0.8% decline in top-10,000 precision.
• The method proved robust across temporal validation splits and bootstrap resampling, with consistent fairness gains and minimal accuracy loss.
• Model selection strategies that ignored fairness during training but applied post-hoc thresholds performed nearly as well as those that prioritized fairness in selection, confirming the effectiveness of post-hoc adjustment.