Characterizing Fairness Over the Set of Good Models Under Selective Labels

Amanda Coston, Ashesh Rambachan, Alexandra Chouldechova

TL;DR

This paper introduces a framework called FaiRS (Fairness in the Rashomon Set) to characterize and optimize predictive fairness across a set of "good" models that achieve similar overall performance. It specifically addresses the challenge of "selective labels" (where outcomes are only observed for a subset of the data), enabling more robust fairness audits and model improvements in real-world scenarios.

What Problem Does This Paper Address?

In high-stakes decision-making (like lending or criminal justice), algorithmic models are often used. A common issue is the "Rashomon Effect," where many different models can achieve similar overall accuracy but make very different predictions for individual cases. This diversity means these models can have vastly different fairness properties across different demographic groups. Furthermore, in many real-world applications, we don't have outcome labels for everyone (e.g., we only know if a loan applicant defaulted if they were approved and funded). This "selective labels" problem makes it difficult to accurately assess and improve fairness. The paper aims to bridge this gap by providing a method to explore the fairness landscape within the set of "good" models, even when dealing with incomplete outcome data, and to find models that minimize predictive disparities.

Methods

The core of the paper is the FaiRS (Fairness in the Rashomon Set) framework. It works by:

  1. Defining the "Set of Good Models": This set includes all prediction functions that achieve performance within a specified tolerance (e.g., a small percentage of the loss of a benchmark model).
  2. Characterizing Predictive Disparities: FaiRS defines a flexible measure of predictive disparity that can capture various fairness notions (like statistical parity or balance for the positive class).
  3. Optimization via Reductions: The paper adapts a "reductions approach" from prior work to efficiently search within the set of good models. This involves transforming the problem into a constrained optimization problem (specifically, a min-max problem) that can be solved using an exponentiated gradient algorithm.
  4. Handling Selective Labels: A key innovation is the extension of FaiRS to handle selective labels. This is achieved by using "outcome modeling" (specifically, extrapolation techniques like RIE and IE) to estimate the unobserved outcomes, allowing for fairness evaluation on the entire target population.

Datasets

The paper uses the following datasets for evaluation:

  • Commonwealth Bank of Australia (CommBank) credit-scoring dataset: A real-world dataset with selectively labeled outcomes (default status only observed for funded loans).
  • ProPublica's COMPAS recidivism data: A publicly available dataset used for auditing fairness in criminal justice risk assessment.
  • Communities & Crime dataset: Used for regression experiments to demonstrate FaiRS's performance on continuous outcomes.

Key Insights and Ideas

  • The "Rashomon Effect" is a double-edged sword for fairness: While it offers flexibility to find fairer models, it also means that simply having a "good" model doesn't guarantee fairness. We need to actively explore the space of good models.
  • Selective labels are a major hurdle for fairness: Evaluating fairness on only the observed data can be highly misleading. FaiRS's ability to handle this is crucial for real-world applicability.
  • Auditing for "business necessity" is enabled: By showing that fairer models exist within the "good" set, FaiRS can help challenge the justification of existing models that exhibit disparate impact.
  • FaiRS provides a spectrum of fairness-performance trade-offs: It doesn't just find one "fair" model, but allows users to explore the range of possible disparities for a given performance constraint.

Key Results

  • Credit Lending: FaiRS successfully reduced predictive disparities across geographic regions while maintaining comparable AUCs to benchmark models, outperforming other methods like TFCS and Fairlearn.
  • Recidivism Prediction: FaiRS demonstrated that the COMPAS risk assessment exhibits larger disparities than many models within the "good" set, suggesting that fairer alternatives exist without sacrificing performance.
  • Regression Tasks: FaiRS also showed improved statistical parity on the Communities & Crime dataset without compromising performance compared to a benchmark linear regression model.
  • Selective Labels Handling: The RIE and IE approaches for handling selective labels were shown to be effective, producing better fairness-performance trade-offs than methods that ignore the selective nature of the data.

Limitations and Future Work

  • Assumption of Unconfoundedness: The framework assumes that the selection decision and outcome are unconfounded given observed data features. Violations of this assumption could impact results.
  • Computational Complexity: While FaiRS is designed to be efficient, the optimization process can still be computationally intensive, especially with large datasets or complex function classes.
  • Choice of Fairness Metric: The paper focuses on specific disparity measures, but the framework is flexible. Future work could explore a wider range of fairness definitions and their interactions.
  • Practical Implementation: The paper acknowledges that the stochastic nature of the returned solution can pose practical challenges, though they propose methods to shrink its support. Further research could explore more streamlined implementations.

Summary updated

Abstract

Algorithmic risk assessments are used to inform decisions in a wide variety of high-stakes settings. Often multiple predictive models deliver similar overall performance but differ markedly in their predictions for individual cases, an empirical phenomenon known as the "Rashomon Effect." These models may have different properties over various groups, and therefore have different predictive fairness properties. We develop a framework for characterizing predictive fairness properties over the set of models that deliver similar overall performance, or "the set of good models." Our framework addresses the empirically relevant challenge of selectively labelled data in the setting where the selection decision and outcome are unconfounded given the observed data features. Our framework can be used to 1) replace an existing model with one that has better fairness properties; or 2) audit for predictive bias. We illustrate these uses cases on a real-world credit-scoring task and a recidivism prediction task.
BibTeX
@inproceedings{coston2021characterizing,
  author = {Amanda Coston and Ashesh Rambachan and Alexandra Chouldechova},
  title = {{Characterizing Fairness Over the Set of Good Models Under Selective Labels}},
  booktitle = {International Conference on Machine Learning},
  year = {2021},
}
Figure