[Paper Review] Augmented Fairness: An Interpretable Model Augmenting Decision-Makers' Fairness

Tong Wang, Maytal Saar‐Tsechansky|arXiv (Cornell University)|Nov 17, 2020

Ethics and Social Impacts of AI19 references4 citations

TL;DR

This paper proposes AuFair, a model-agnostic, interpretable method that enhances fairness in human or black-box decision-makers by replacing biased predictions with a set of short decision rules trained under dual objectives: predictive performance and fairness. Under budget-limited true label acquisition, AuFair achieves superior fairness-accuracy tradeoffs compared to baselines, including state-of-the-art post-processing methods.

ABSTRACT

We propose a model-agnostic approach for mitigating the prediction bias of a black-box decision-maker, and in particular, a human decision-maker. Our method detects in the feature space where the black-box decision-maker is biased and replaces it with a few short decision rules, acting as a "fair surrogate". The rule-based surrogate model is trained under two objectives, predictive performance and fairness. Our model focuses on a setting that is common in practice but distinct from other literature on fairness. We only have black-box access to the model, and only a limited set of true labels can be queried under a budget constraint. We formulate a multi-objective optimization for building a surrogate model, where we simultaneously optimize for both predictive performance and bias. To train the model, we propose a novel training algorithm that combines a nondominated sorting genetic algorithm with active learning. We test our model on public datasets where we simulate various biased "black-box" classifiers (decision-makers) and apply our approach for interpretable augmented fairness.

Motivation & Objective

• To address bias in human decision-makers, who cannot be retrained or accessed on-demand, and whose outputs are discrete, not probabilistic.
• To develop a fairness mitigation method that works with only a small, budget-constrained set of true labels, not full ground truth.
• To ensure interpretability so human experts can vet and approve the augmented model in high-stakes domains.
• To provide a flexible, multi-solution tradeoff between fairness and predictive performance, rather than a single fixed solution.
• To create a hybrid human-machine system where an interpretable surrogate model steps in selectively to correct bias.

Proposed method

• The method uses a multi-objective optimization framework to train an interpretable surrogate model (AuFair) that balances predictive performance and fairness.
• AuFair is trained on the black-box decision-maker’s predictions and a limited set of true labels acquired via active learning.
• A novel training algorithm combines nondominated sorting genetic algorithm (NSGA-II) with active learning to iteratively select the most informative instances for label acquisition.
• The surrogate model consists of interpretable association rules that act as a 'fair surrogate' in regions of the feature space where the original decision-maker is biased.
• The model is evaluated on the Pareto frontier of fairness (measured as absolute difference in true positive rates) and error rate.
• Label acquisition is guided by uncertainty and diversity criteria to maximize information gain under budget constraints.

Experimental results

Research questions

• RQ1 Can a post-processing method achieve better fairness-accuracy tradeoffs than existing baselines when only a limited number of true labels are available?
• RQ2 Can an interpretable surrogate model effectively replace biased decisions by a human or black-box decision-maker while remaining understandable to experts?
• RQ3 How does the performance of the proposed method scale with increasing label acquisition budgets?
• RQ4 Does the active learning strategy for label acquisition significantly improve model performance compared to random or non-adaptive sampling?
• RQ5 Can the method provide a range of solutions on the fairness-accuracy frontier, enabling task-specific preference selection?

Key findings

• AuFair achieves significantly better fairness-accuracy tradeoffs than the EOP baseline, particularly in reducing bias while maintaining comparable error rates.
• Even with only 1% of true labels, AuFair outperforms baselines that use 100% of the true labels in terms of both fairness and error.
• The Pareto frontier of AuFair improves substantially between 1% and 10% label acquisition, with diminishing returns beyond 10%.
• AuFair dominates the EOP baseline across all budget levels, offering solutions with lower bias and/or lower error.
• The active learning strategy effectively identifies informative instances, reducing variance and improving model performance with minimal label queries.
• The method successfully provides a flexible set of solutions on the fairness-accuracy frontier, enabling selection of task-specific optimal tradeoffs.