50 Years of Test (Un)fairness: Lessons for Machine Learning (1811.10104v2)

Abstract: Quantitative definitions of what is unfair and what is fair have been introduced in multiple disciplines for well over 50 years, including in education, hiring, and machine learning. We trace how the notion of fairness has been defined within the testing communities of education and hiring over the past half century, exploring the cultural and social context in which different fairness definitions have emerged. In some cases, earlier definitions of fairness are similar or identical to definitions of fairness in current machine learning research, and foreshadow current formal work. In other cases, insights into what fairness means and how to measure it have largely gone overlooked. We compare past and current notions of fairness along several dimensions, including the fairness criteria, the focus of the criteria (e.g., a test, a model, or its use), the relationship of fairness to individuals, groups, and subgroups, and the mathematical method for measuring fairness (e.g., classification, regression). This work points the way towards future research and measurement of (un)fairness that builds from our modern understanding of fairness while incorporating insights from the past.

Analysis of "50 Years of Test (Un)fairness: Lessons for Machine Learning"

The paper by Ben Hutchinson and Margaret Mitchell offers an extensive retrospective analysis of fairness concepts in testing over the past fifty years and presents notable connections and lessons for current ML applications, especially concerning fairness in algorithms. This document traces the evolution of fairness definitions from the domains of educational and employment testing to their parallels and implications in modern machine learning, emphasizing the persistent challenges and opportunities inherent in defining and implementing fairness criteria.

Historical Overview and Evolution

The authors dissect the timeline beginning in the 1960s, when societal transformations, spurred largely by the Civil Rights Movement in the United States, demanded quantitative measures of fairness primarily focused on racial discrimination in testing. The early work, such as Cleary's definition of test bias and Guion's employment fairness, emphasized prediction errors and equal probabilities, setting the foundation for subsequent fairness frameworks.

During the 1970s, a shift toward defining what constitutes fairness, rather than mere absence of unfairness, emerged. Researchers like Thorndike refined fairness notions, introducing contextual fairness with an emphasis on test application rather than test attributes alone. Concepts from this era, such as equalizing opportunity and minimizing prediction disparities among subgroups, predate but align closely with modern ML fairness metrics such as equality of opportunity and equalized odds.

The analysis indicates that despite rigorous statistical exploration, by the end of the 1970s, a concrete and universal method to determine test fairness remained elusive. As detailed by Nancy Cole, the lack of consensus on fairness criteria led to stagnation in research without solving the core issues of bias — a cautionary tale for the burgeoning field of ML fairness research today.

Implications for Machine Learning Fairness

Hutchinson and Mitchell uncover philosophical and technical overlaps between historical fairness in testing and contemporary considerations in machine learning. The parallels, including conflicts between individual and group fairness and the inherent incompatibility of certain fairness criteria, underscore significant lessons for AI researchers:

1. Historical Definitions and ML Metrics: Many legacy testing criteria translate directly into ML terminologies. For instance, Cleary's fairness based on regression equality presages the sufficiency in ML, and early concepts around demographic parity mirror today's independence criteria.
2. Real-World Context in Fairness Evaluation: Emphasizing Thornton and Peterson and Novick's arguments, fairness should not be assessed in a vacuum. Understanding a model's context and intended use is integral and requires both domain-specific knowledge and sensitivity to societal values, an aspect partially overlooked in the ML domain.
3. Inclusivity of Historical Insights in Fairness Constructs: Reinforcing earlier lessons on fairness within defined societal contexts and historical value systems can guide present-day ML ethics. This historical insight urges ML practitioners to account for complex societal variations beyond simplistic statistical balance.

Forward-Looking Perspectives

Moving beyond the existing paradigms, the authors highlight untapped areas such as identifying sources of unfairness, analogous to Differential Item Functioning (DIF) in test design, which can potentiate advanced fairness assessments in machine learning. Additionally, redefining fairness along compromise spectrums and incorporating quantifiable value trade-offs can balance accuracy and equity interests in algorithmic systems.

By learning from the "disappointing" outcomes of prior fairness endeavors in educational and hiring exams, the ML community can strive toward developing fairness definitions that resonate with public perceptions while balancing ethical complexities. This approach requires interdisciplinary efforts, cooling the inclination towards overly rigid metric adherence, and must be grounded in values reflecting societal equity and inclusiveness.

In summary, Hutchinson and Mitchell's paper offers a robust synthesis of fairness in testing history, proposing valuable insights for emerging machine learning contexts. By bridging past and present, fairness can be critically evaluated in ML systems, paving the way for equitable algorithmic decision-making processes responsive to both technical rigor and societal justice.