Long-Term Fairness Inquiries and Pursuits in Machine Learning - A Survey of Notions, Methods, and Challenges

Aug 5, 2026

Upload by Oliver

Widespread deployment of machine learning in everyday, especially high-stakes, settings has intensified scrutiny of fairness. Prior work often treats fairness as static, relying on predefined statistical or causal criteria under fixed environments. This survey emphasizes long-term fairness, where feedback loops, model-environment interactions, and evolving dynamics can misalign short-term objectives with desired outcomes over time. It reviews long-term fairness literature, proposes a taxonomy, highlights key challenges, and outlines future research directions.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages16

Reading Time40 min

AccessFree to download

Introduction

Static vs. long-term fairness

Scope and facets of long-term fairness

Difference from existing surveys

High-level taxonomy

Long-Term Fairness Inquiries and Pursuits in Machine Learning: A Survey of Notions, Methods, and Challenges Usman Gohar 1 , Zeyu Tang2 , Jialu Wang3 , Kun Zhang2 , Peter L. Spirtes2 Yang Liu3 , Lu Cheng4 1Iowa State University 2 Carnegie Mellon University 3 University of California Santa Cruz 4 University of Illinois Chicago [ugohar@iastate.edu](ugohar@iastate.edu), [zeyutang@cmu.edu](zeyutang@cmu.edu), [faldict@ucsc.edu](faldict@ucsc.edu), [kunz1@cmu.edu](kunz1@cmu.edu), [ps7z@andrew.cmu.edu](ps7z@andrew.cmu.edu), [yangliu@ucsc.edu](yangliu@ucsc.edu), [lucheng@uic.edu](lucheng@uic.edu) arXiv :2406 .06736v1 [ cs .LG] 10 Jun 2024 Abstract The widespread integration of Machine Learning systems in daily life, particularly in high-stakes domains, has raised concerns about the fairness implications. While prior works have investigated static fairness measures, recent studies reveal that automated decision-making has long-term implications and that off-the-shelf fairness approaches may not serve the purpose of achieving long-term fairness. Additionally, the existence of feedback loops and the interaction between models and the environment introduces additional complexities that may deviate from the initial fairness goals. In this survey, wereview existing literature on long-term fairness from different perspectives and present a taxonomy for long-term fairness studies. We highlight key challenges and consider future research directions, analyzing both current issues and potential further explorations. 1 Introduction As Machine Learning (ML) algorithms assume increasingly influential roles in high-stake domains traditionally steered by human judgments, an extensive body of research has brought attention to the challenges of bias and discrimination against marginalized groups (Mehrabi et al. 2021; Cheng, Varshney, and Liu 2021) . These issues are pervasive and manifest in different settings, including finance, legal (e.g., pretrial bail decisions), aviation, and healthcare practices, among others (Gohar et al. 2024; Barocas, Hardt, and Narayanan 2023) . The community has led tremendous research efforts toward measuring and mitigating algorithmic unfairness (Mehrabi et al. 2021) . However, the majority of such fairness approaches quantify unfairness based on predefined statistical or causal criteria, often assuming a constant environment and system, defined as static fairness. This myopic conceptualization of fairness usually focuses on short-term outcomes and assesses only the instantaneous impact of interventions at a single snapshot, overlooking system dynamics (dynamic fairness) and/or longterm consequences over a time horizon (Zhang et al. 2020; Liu et al. 2018) . Recent works have brought attention to the misalignment of long-term fairness considerations with the fairness objectives optimized in static settings and shown that imposing Copyright © 2024, Association for the Advancement of Artificial Intelligence ([www.aaai.org](www.aaai.org)). All rights reserved. static fairness criteria often does not guarantee long-term fairness and may even amplify discrimination (Hu, Immorlica, and Vaughan 2019; Liu et al. 2018) . One way static fairness approaches fall short is by failing to account for the future implications of current decisions on individuals or groups, undermining their effectiveness. For example, in predictive policing, Ensign et al. (2018) observed how an initial higher allocation of police in a specific area leads to more reported incidents, perpetuating increased surveillance and exacerbating the marginalization of those communities over time. To this end, a large body of work has proposed various methods to measure, mitigate, and evaluate different aspects of fairness in the long-term setting rather than achieving fairness for a single time step. Simply put, long-term fairness constitutes the settings outside of static fairness framework and short-term outcomes by addressing fairness o

FAQs

What differentiates long-term fairness from static fairness in machine learning?

Static fairness evaluates unfairness using predefined criteria at a single snapshot under an assumed fixed environment. Long-term fairness considers extended time horizons and evolving system dynamics, including feedback effects between models and the environment.

Why might static fairness objectives fail to guarantee long-term fairness?

Imposing static fairness criteria may not account for how current decisions affect future outcomes for individuals or groups. In some settings, the mismatch can even amplify discrimination over time.

What does the survey contribute regarding the study of long-term fairness?

The survey reviews existing work from multiple perspectives and introduces a taxonomy covering dimensions and problem settings of long-term fairness. It also identifies key challenges and proposes future research directions.

Aug 5, 2026

Upload by Oliver