Transparency Tools for Fairness in AI (Luskin)

Mingliang Chen, Aria Shahverdi, Sarah Anderson, Se Yong Park, Justin Zhang, Dana Dachman-Soled, Kristin E. Lauter, Min WuarXiv 2020 · arXiv

Introduction

As machine learning models increasingly govern life-altering decisions—from loan approvals to parole eligibility—the risk of codifying and scaling historical biases grows. The "Luskin" toolkit addresses this challenge by providing policy-makers and data scientists with a suite of algorithms designed to detect and remediate bias. By distinguishing between "controlled fairness" (prioritizing group parity) and "classification parity" (prioritizing error rate equality), the authors offer practical methods to reduce systemic bias in sensitive domains.

The Problem: Biased Data and Opaque Decisions

AI systems often rely on historical data that reflects past human prejudices. The authors identify a critical gap in current AI deployment: a lack of concrete tools to not only measure this bias but also remediate it without rendering the models useless. The paper categorizes the fairness problem into two distinct scenarios:

  1. Absence of "Ground Truth": In domains like police stops, the data consists of past human decisions which may be inherently biased. In these cases, the objective is "parity"—ensuring that similar demographic groups are treated similarly regardless of historical trends.
  2. Objective Ground Truth: In domains like recidivism (re-offending), there is an objective outcome to measure against. Here, the goal is "accuracy-based fairness," ensuring that the model's error rates (false positives and false negatives) are equalized across different demographic groups.

Methodology: Three Pillars of Fairness

The toolkit introduces three primary methodological interventions to achieve these goals:

1. Controlled Fairness Definition

This is a refinement of traditional "statistical parity." It allows for "filtering" or "controlling" for unprotected features. For instance, when comparing loan approval rates between different racial groups, a policy-maker might use this tool to compare only individuals with similar education levels or credit histories, ensuring that the fairness assessment is nuanced rather than a blunt comparison of averages.

2. Retraining for Parity

To address biased training data, the authors propose two algorithms to create "synthetic" fair datasets:

  • Selective Risk Adjustment: This method perturbs the risk scores of a specific group to match the outcome ratio of another.
  • Risk-Based Flipping: This algorithm flips the labels of the minimum number of data points (e.g., changing a "frisk" to "no frisk") based on their risk scores to achieve parity while maintaining as much data integrity as possible.

3. Classification Parity via Optimization

For scenarios with ground truth, the toolkit offers two optimization paths:

  • With Protected Features: Using Particle Swarm Optimization (PSO), the system finds optimal decision thresholds for different groups to equalize True Positive and False Positive rates.
  • Without Protected Features: The authors introduce a novel histogram-based regularization term. By approximating score distributions with Gaussian functions, they penalize the model during training if the score distributions for different groups diverge. This allows for fair outcomes even if the model is legally or practically prohibited from seeing protected features (like race) during the inference phase.

Experimental Evaluation

The toolkit was evaluated on three standard datasets representing high-stakes social domains:

  • NYPD Stop-Question-Frisk (SQF): Analyzing police stop data in New York City.
  • Adult Income (US Census): Predicting whether an individual earns more than $50k/year, used as a proxy for loan eligibility.
  • COMPAS Recidivism: Predicting the likelihood of a criminal re-offending within two years.

Key Insights and Results

The Non-Linearity of Fairness

A significant finding of the research is that when labels are adjusted to make a dataset "fair," the resulting decision boundary often becomes highly non-linear. The authors observed that simple Logistic Regression models frequently failed to learn these new fair boundaries. In contrast, Multi-layer Perceptrons (MLPs) were successful, suggesting that fairness requires more complex model architectures to capture the nuances of corrected data.

Bias Reduction vs. Accuracy Trade-offs

In the SQF dataset, the initial "frisk" ratio gap between White and Non-White individuals (when suspected of assault) was 13%. By applying the retraining algorithms with an MLP, the authors reduced this gap to just 1% to 2%.

In the COMPAS dataset, tuning thresholds via PSO successfully equalized the gaps in True Positive and False Positive rates (which were initially around 0.2) to less than 0.05. This significant improvement in fairness came with only a minor drop in overall accuracy, ranging from approximately 1.7% to 2.3%.

Threshold Invariance

The toolkit’s "equalized distribution" approach is particularly robust because it aims to make the entire score distribution identical across groups. This makes the fairness of the model "threshold-invariant," meaning the model remains fair regardless of where the final decision cutoff is set.

Limitations and Future Work

While the Luskin toolkit provides a powerful framework, it faces several practical hurdles. Many of the most effective methods require access to protected attributes during the training phase, which may be restricted by privacy laws. Additionally, the "controlled fairness" approach relies on policy-makers to correctly identify which features are "relevant" versus "biased"; an incorrect choice could inadvertently mask existing discrimination. Finally, the computational overhead of techniques like PSO and complex regularization requires careful hyperparameter tuning to balance the trade-off between accuracy and social equity.

Summary updated

Abstract

We propose new tools for policy-makers to use when assessing and correcting fairness and bias in AI algorithms. The three tools are: - A new definition of fairness called "controlled fairness" with respect to choices of protected features and filters. The definition provides a simple test of fairness of an algorithm with respect to a dataset. This notion of fairness is suitable in cases where fairness is prioritized over accuracy, such as in cases where there is no "ground truth" data, only data labeled with past decisions (which may have been biased). - Algorithms for retraining a given classifier to achieve "controlled fairness" with respect to a choice of features and filters. Two algorithms are presented, implemented and tested. These algorithms require training two different models in two stages. We experiment with combinations of various types of models for the first and second stage and report on which combinations perform best in terms of fairness and accuracy. - Algorithms for adjusting model parameters to achieve a notion of fairness called "classification parity". This notion of fairness is suitable in cases where accuracy is prioritized. Two algorithms are presented, one which assumes that protected features are accessible to the model during testing, and one which assumes protected features are not accessible during testing. We evaluate our tools on three different publicly available datasets. We find that the tools are useful for understanding various dimensions of bias, and that in practice the algorithms are effective in starkly reducing a given observed bias when tested on new data.
BibTeX
@misc{chen2020transparency,
  author = {Mingliang Chen and Aria Shahverdi and Sarah Anderson and Se Yong Park and Justin Zhang and Dana Dachman-Soled and Kristin E. Lauter and Min Wu},
  title = {{Transparency Tools for Fairness in AI (Luskin)}},
  year = {2020},
  eprint = {2007.04484},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2007.04484},
}
Figure