Source-linked AI summary
Bias Mitigation Post-processing for Individual and Group Fairness
Pranay K. Lohia, Karthikeyan Natesan Ramamurthy, Manish Bhide, Diptikalyan Saha, Kush R. Varshney, Ruchir Puri
TL;DR
Existing post-processing methods mainly address group fairness, leaving a gap in jointly improving individual and group fairness. The paper proposes IGD, which detects likely individual bias and prioritizes those samples for disparate-impact mitigation. Across real-world tasks, it reports superior combined accuracy and fairness with little accuracy loss.
Problem
Existing post-processing algorithms primarily address group fairness, while methods jointly addressing individual and group fairness are lacking.
Method
IGD trains an individual bias detector from validation samples and prioritizes likely biased samples for post-processing toward group fairness.
Results
Across six real-world classification tasks, IGD showed superior combined classification accuracy, individual fairness, and group fairness, with little accuracy reduction and improved fairness measures.
Takeaways & Limitations
Post-processing can target individual-bias cases while improving both individual and group fairness without much loss in classification accuracy.
Takeaways & Limitations
The exposition assumes scalar binary protected attributes, whereas realistic settings may require several model evaluations for individual-bias detection.
Abstract
from arXiv · showhide
Whereas previous post-processing approaches for increasing the fairness of predictions of biased classifiers address only group fairness, we propose a method for increasing both individual and group fairness. Our novel framework includes an individual bias detector used to prioritize data samples in a bias mitigation algorithm aiming to improve the group fairness measure of disparate impact. We show superior performance to previous work in the combination of classification accuracy, individual fairness and group fairness on several real-world datasets in applications such as credit, employment, and criminal justice.
1. INTRODUCTION
The paper addresses the gap that existing post-processing methods largely target group fairness, proposing a black-box approach that also targets individual fairness. It prioritizes likely individually biased samples and reports improved fairness with little accuracy loss across real-world tasks.
- Post-processing is suitable for runtime, black-box deployment because it does not require access to training procedures or model internals.
- Existing post-processing algorithms primarily address group fairness, while the proposed method considers both group and individual fairness.
- The method prioritizes samples that have or are likely to have individual fairness issues rather than selecting samples randomly or solely by prediction uncertainty.
- The individual bias detector tests whether predictions change when protected attributes change while other features remain constant, then generalizes from a small set of checked points.
- Across six real-world classification tasks, the method achieved superior combined classification accuracy, individual fairness, and group fairness, with little accuracy reduction and substantial fairness improvement.
2. INDIVIDUAL AND GROUP FAIRNESS
The paper distinguishes individual fairness, which seeks similar treatment for similar individuals, from group fairness, which compares statistical measures across protected groups. It defines individual bias and disparate impact and reviews post-processing alternatives.
- Individual bias occurs when a sample’s prediction changes after switching its protected attribute while keeping other features constant.
- The individual bias score bS,i is the privileged-group score minus the unprivileged-group score, while the summary statistic averages binary bias across test samples.
- Disparate impact is a group-fairness notion based on the ratio of favorable prediction rates between unprivileged and privileged groups, with violations outside specified bounds.
- Individual bias detection must identify biased cases and determine bias status across samples, but realistic settings may involve nonscalar or nonbinary protected attributes.
- Reject option classification changes predictions for uncertain samples near the classification threshold, while equalized odds post-processing targets equal false-positive and false-negative rates across groups.
3. PROPOSED ALGORITHM
The proposed IGD algorithm uses a validation-set detector to identify likely individual bias and selectively alter unprivileged-group predictions. It chooses samples under a disparate-impact constraint and applies the corresponding privileged-group outcome at runtime.
- PROPOSED ALGORITHM: IGD selects samples likely to have individual bias, exploiting the exchangeability of samples under expectation-based group-fairness metrics.
- Individual Bias Detector: The method evaluates individual bias on an unlabeled validation partition and trains a detector under an assumption that individual bias is coherent or smooth in feature space.
- Individual Bias Detector: Validation samples receive binary detector labels according to a threshold on soft individual-bias scores chosen to satisfy the disparate-impact constraint.
- Individual Bias Detector: The detector avoids repeated costly evaluations for complex protected attributes and limits remediation to the highest-bias examples selected by the group-fairness constraint.
- Overall Algorithm: At runtime, likely biased unprivileged-group samples are assigned the outcome they would receive in the privileged group, while other samples remain unchanged.
4. EMPIRICAL RESULTS
Experiments on three real-world datasets compare the proposed IGD post-processing method with EOP and ROC using individual bias, disparate impact, and balanced classification accuracy. IGD consistently improves both fairness measures while preserving accuracy close to the original classifier.
- Experimental setup: Experiments use UCI Adult, UCI Statlog German Credit, and ProPublica COMPAS datasets, with logistic regression and random forests as black-box classifiers.Datasets cover income, credit risk, and recidivism prediction, with protected attributes including sex, race, and age.
- Bias detector: All bias-detector accuracy values exceed 0.85 for logistic regression, while random forests achieve a minimum accuracy of approximately 0.80.The detector is evaluated on unprivileged-group samples and generalizes to unseen test data.
- Evaluation measures: The evaluation compares individual bias, disparate impact, and balanced classification accuracy for EOP, ROC, IGD, and the original predictions.Balanced classification accuracy is the mean of true positive and true negative rates.
- Fairness comparisons: IGD is the only compared method that consistently improves both individual fairness and group fairness while keeping accuracy close to the original classifier.The reported comparisons use individual bias, disparate impact, and balanced classification accuracy across the evaluated tasks.
- Fairness comparisons: IGD performs best for individual bias on the German and COMPAS datasets and consistently improves disparate impact over the original predictions.ROC outperforms IGD on disparate impact in five of six cases, while EOP performs poorly on that measure.
- Accuracy comparisons: IGD best preserves balanced classification accuracy relative to the original predictions despite using no ground-truth labels in the validation partition.Similar results are also observed with random forests, although those results are omitted from the figures.
5. CONCLUSION
The paper presents a black-box, run-time post-processing algorithm that targets individually biased samples to improve both individual and group fairness. Empirical results show these improvements across real-world datasets without much loss in classification accuracy, while avoiding ground-truth labels for validation.
- Conclusion: The proposed post-processing algorithm targets samples with individual bias to improve both individual and group fairness metrics.The method is evaluated empirically on several real-world datasets.
- Conclusion: The algorithm treats the classifier as a complete black box and does not require ground-truth class labels for the validation set.This makes it a pure run-time approach within the paper’s stated deployment setting.
- Conclusion: The method improves fairness metrics on several real-world datasets without much loss in classification accuracy.The conclusion positions post-processing as useful when deployers have limited access to pre-trained model internals.