[Paper Review] Equalizing Recourse across Groups

Vivek Gupta, Pegah Nokhiz+2 more|arXiv (Cornell University)|Sep 7, 2019

Explainable Artificial Intelligence (XAI)Computer Science22 references17 citations

TL;DR

This paper introduces a method to equalize recourse across demographic groups by minimizing the disparity in distance to a classifier's decision boundary, using a regularized objective in both linear and non-linear models. It achieves significant reductions in recourse unfairness—up to 100% improvement—while maintaining high prediction accuracy across diverse datasets and model types.

ABSTRACT

The rise in machine learning-assisted decision-making has led to concerns about the fairness of the decisions and techniques to mitigate problems of discrimination. If a negative decision is made about an individual (denying a loan, rejecting an application for housing, and so on) justice dictates that we be able to ask how we might change circumstances to get a favorable decision the next time. Moreover, the ability to change circumstances (a better education, improved credentials) should not be limited to only those with access to expensive resources. In other words, \emph{recourse} for negative decisions should be considered a desirable value that can be equalized across (demographically defined) groups. This paper describes how to build models that make accurate predictions while still ensuring that the penalties for a negative outcome do not disadvantage different groups disproportionately. We measure recourse as the distance of an individual from the decision boundary of a classifier. We then introduce a regularized objective to minimize the difference in recourse across groups. We explore linear settings and further extend recourse to non-linear settings as well as model-agnostic settings where the exact distance from boundary cannot be calculated. Our results show that we can successfully decrease the unfairness in recourse while maintaining classifier performance.

Motivation & Objective

• To address ethical concerns in ML decision-making by ensuring individuals from all demographic groups have comparable opportunities to achieve favorable outcomes through feasible actions.
• To formalize recourse as the distance from an individual’s feature vector to the decision boundary of a classifier, enabling quantifiable fairness metrics.
• To develop a regularized learning framework that equalizes this recourse distance across groups without compromising model accuracy.
• To extend the approach to non-linear and model-agnostic settings where exact boundary distances are unavailable.
• To empirically validate that equalizing recourse reduces unfairness while preserving predictive performance across diverse datasets and classifiers.

Proposed method

• Define recourse as the absolute distance from an individual’s feature vector to the decision boundary of a classifier, using this as a fairness metric.
• Formulate a regularized objective function that minimizes the difference in average recourse distance across demographic groups during model training.
• Apply the method to linear models and extend it to kernelized models (e.g., SVMs) using polynomial and radial basis function kernels.
• For model-agnostic settings, use a re-weighting strategy based on estimated recourse distances to balance group disparities.
• Leverage sample weights in models like Random Forest and Logistic Regression to approximate the regularization effect without requiring explicit boundary distance computation.
• Use percentile-based distributions and median analysis to evaluate robustness of recourse equality across multiple runs and datasets.

Experimental results

Research questions

• RQ1 Can we reduce disparities in the feasibility of recourse across demographic groups while maintaining high model accuracy?
• RQ2 How effective is a regularized objective that minimizes inter-group differences in recourse distance in both linear and non-linear models?
• RQ3 To what extent can recourse fairness be improved in model-agnostic settings where the decision boundary is not explicitly known?
• RQ4 Does equalizing recourse lead to measurable reductions in unfairness, as quantified by mean, median, and percentile distributions of recourse differences?
• RQ5 How does the proposed method affect model accuracy across different datasets and classifier types?

Key findings

• The proposed method reduced recourse disparity by up to 100% on the givemecredit dataset using polynomial kernels, with 95th percentile improvements exceeding 90% in several cases.
• On the credit dataset, recourse difference was reduced by 90% in training and 100% in testing using logistic regression in the model-agnostic setting.
• Median recourse differences decreased across all datasets and models, indicating that at least half of the runs achieved lower disparity after equalization.
• Model accuracy was preserved within 1% of baseline in most cases, with only minor drops (up to 7%) observed in specific cases like propublica on AdaBoost.
• The linear kernel showed less improvement than the polynomial kernel, likely due to fewer model parameters and reduced flexibility in adjusting the decision boundary.
• The re-weighting strategy in model-agnostic settings successfully reduced recourse disparity across Random Forest, Logistic Regression, and AdaBoost, demonstrating generalizability beyond explicit boundary computation.