Exploring Fairness Interventions in Open Source Projects
Sadia Afrin Mim 1 , Fatema Tuz Zohra 2 Justin Smith 3 , Brittany Johnson 4 · 2025
Read Exploring Fairness Interventions in Open Source Projects by Sadia Afrin Mim 1 , Fatema Tuz Zohra 2 Justin Smith 3 , Brittany Johnson 4 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-07026.txt.
Introduction
Sadia Afrin Mim, Fatema Tuz Zohra, Justin Smith, and Brittany Johnson surveyed open source tools that fight bias in machine learning, in a 2025 paper. Biased models have harmed people in justice, health, and other high-stakes fields, and many fairness interventions exist to detect or reduce bias. Yet practitioners rarely use them, often not knowing they exist. The team hunted down sixty-two such projects and studied their features in depth.
Only about a third had been actively maintained within the past year. Half could both detect and mitigate bias, most of them acting during model training. The catalog exposes a gap between what researchers publish and what survives as usable software. The authors argue that knowing which tools are alive, and what they offer, helps practitioners choose. It also shows funders where the ecosystem is thinning out.
Catalog record
| CREATOR | Sadia Afrin Mim 1 , Fatema Tuz Zohra 2 Justin Smith 3 , Brittany Johnson 4 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-07026
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Exploring Fairness Interventions in Open Source Projects Sadia Afrin Mim 1 , Fatema Tuz Zohra 2 Justin Smith 3 , Brittany Johnson 4 Abstract The deployment of biased machine learning (ML) models has resulted in adverse effects in crucial sectors such as criminal justice, and healthcare. To address these challenges, a diverse range of machine learning fairness interventions have been developed, aiming to mitigate bias and promote the creation of more equitable models. Despite the growing availability of these interventions, their adoption in real-world applications remains limited, with many practitioners unaware of their existence. To address this gap, we systematically identified and compiled a dataset of 62 open source fairness interventions and identified active ones. We conducted an in-depth analysis of their specifications and features to uncover considerations that may drive practitioner preference and to identify the software interventions actively maintained in the open source ecosystem. Our findings indicate that 32% of these interventions have been actively maintained within the past year, and 50% of them offer both bias detection and mitigation capabilities mostly during inprocessing. I Introduction Machine learning (ML) applications and frameworks are having a major impact on modern society. In safety-critical domains such as healthcare, automotive, and manufacturing, ML systems have surpassed human performance by processing large amounts of data and making predictions or classifications with greater reliability [ 1 ] . Although these advances are promising, numerous studies have also revealed that many ML applications are prone to undesirable bias. Examples include, but are not limited to, biases in the evaluation of physical appearance [ 2 ] , demographic biases in loan application approvals [ 3 ] , gender bias in employee recruitment [ 4 ] , natural language processing [ 5 ] and racial bias in determining criminal sentencing [ 4 ] . In response to growing concerns regarding bias in machine learning models, researchers and practitioners have proposed a variety of potential solutions. The most common form of support provided is automated software interventions designed to help identify, mitigate and evaluate bias within machine learning models and their training datasets [ 6 , 7 , 8 ] . The availability of fairness interventions provides practitioners with an option for addressing fairness concerns in practice, however, research indicates that practitioners may lack the necessary expertise to use them effectively [ 7 ] . Furthermore, as the range of interventions continues to expand, the challenge of identifying and evaluating the most suitable options has also become complex. The goal of our research is to provide a comprehensive exploration of open source fairness interventions to form foundations that can simplify thieir selection and evaluation process. To this end, we conducted an in-depth analysis of currently available interventions, focusing on their features and specifications for use. We compiled a collection of available, fully functional fairness interventions and analyzed the information provided in their documentation and tutorials. Additionally, we determined which interventions are active and inactive based on their recent commit activity. We further explored the underlying ecosystems, algorithmic fairness approaches, domain-specific applications, and mitigation techniques for each intervention to better understand the ways in which these interventions might meet the needs of practitioners. The contributions of this work are: • An extensive compilation of 62 open source fairness interventions available for practitioners to incorporate into their workflows. • A qualitative assessment and categorization of the features and specifications offered by existing interventions that can support their adoption and engagement in the open source community. The remainder of this paper is organized as follows: Section II reviews the related literature. Section III outlines our methodology, beginning with the formulation of research questions that guided our analysis of fairness interventions, including their features, specifications, documentation, and tutorials. Section IV presents the results of our analysis. Insights derived from these findings are discussed in Section V and some possible threats to the validity of our work are in Section VI . We conclude the paper in Section VII . II Related Work Numerous efforts have focused on exploring various aspects of software interventions and machine learning fairness. Relevant to our work, prior research has examined the taxonomy of software interventions, provided insights into machine learning fairness, and proposed approaches for mitigating bias in machine learning models. II-A Software Interventions Taxonomies Creating software intervention taxonomies is a common approach in the software engineering research community. Researchers have introduced various taxonomies over time to assist developers in selecting the interventions best suited to their tasks. Delgado and colleagues developed a structured taxonomy for runtime software-fault monitoring interventions [ 9 ] . They categorized the interventions based on key components such as specification languages, monitoring mechanisms, and event handlers. This taxonomy helps developers in understanding, classifying, and distinguishing advancements in monitoring approaches. Shaukat and colleagues developed a taxonomy to classify and compare automated software testing interventions [ 10 ] . This taxonomy assists professionals in selecting the most appropriate interventions by analyzing 32 automated testing interventions based on attributes like operating system compatibility, programming language support, licensing, browser support, and cost. The work by Lee and colleagues represents a significant contribution to the field of machine fairness intervention analysis [ 7 ] . Although the paper does not explicitly offer a taxonomy, it provides a comparative analysis of six prominent open source fairness toolkits, such as IBM Fairness 360 (AIF360) and Google What-If intervention. The authors assessed the interventions by examining their features and limitations, including aspects such as fairness metrics, models, and bias mitigation algorithms. This analysis emphasizes functionality, usability, and applicability, highlighting the differences in the design, features, and target audiences of these toolkits. Pagano and colleagues reviewed current research on bias and unfairness in machine learning (ML) models. They examined 45 studies published between 2017 and 2022 and organized information into categories such as types of bias (data, algorithm, user interaction), fairness metrics, datasets, and mitigation techniques (preprocessing, inprocessing, postprocessing) [ 11 ] . It helps researchers select a intervention by providing detailed evaluations of interventions like AIF360, FairLearn, and Aequitas, outlining their features, limitations, and suitability for different bias identification and mitigation scenarios. II-B Preliminary Concepts of Machine Learning Fairness Fairness in machine learning has gained significant importance as an essential topic to address as it important to develop systems that are unbiased and promote equity. Mehrabi and colleagues explored bias and fairness in machine learning, identifying key sources of bias and discussing their real-world implications [ 2 ] . They also reviewed mitigation strategies and fairness toolkits, providing valuable insights for addressing unfairness in ML systems. Zhang and colleagues investigated structural biases and limitations in fairness-aware algorithms, emphasizing the need for interventions that address both technical and sociopolitical biases in ML workflows [ 12 ] . Makhlouf and colleagues examined how fairness in machine learning models interacts with the broader sociotechnical context, highlighting the impact of abstractions on fairness outcomes and pointing the importance of a contextual understanding of fairness beyond technical solutions [ 13 ] . Caton and colleagues analyzed datasets, fairness metrics, and bias mitigation techniques, discussing open source libraries and how feature alteration and training data size influence fairness [ 14 ] . Lee and colleagues categorized fairness-aware datasets by domain, such as healthcare, finance, and