Bias Begins with Data: The FairGround Corpus for Robust and Reproducible Research on Algorithmic Fairness

TL;DR

FairGround tackles the reproducibility and diversity bottlenecks in algorithmic fairness by providing a cohesive data-centric framework: a 44-dataset tabular corpus with 136 scenarios, rich metadata, and a reproducible Python pipeline for loading, preprocessing, transforming, and splitting data. It integrates principled dataset annotations with curated collections and a selection algorithm to enable diverse, transferable fairness benchmarking across pre-, in-, and post-processing techniques. Through large-scale experiments evaluating seven debiasing methods, the work reveals substantial variability in method performance across datasets and highlights instability in technique rankings, underscoring the need for diverse benchmarks and data-aware evaluation. The framework promotes transparency, opens new avenues for connecting dataset characteristics to algorithmic fairness outcomes, and supports policy-relevant data governance while remaining extensible to other data modalities and research domains.

Abstract

As machine learning (ML) systems are increasingly adopted in high-stakes decision-making domains, ensuring fairness in their outputs has become a central challenge. At the core of fair ML research are the datasets used to investigate bias and develop mitigation strategies. Yet, much of the existing work relies on a narrow selection of datasets--often arbitrarily chosen, inconsistently processed, and lacking in diversity--undermining the generalizability and reproducibility of results. To address these limitations, we present FairGround: a unified framework, data corpus, and Python package aimed at advancing reproducible research and critical data studies in fair ML classification. FairGround currently comprises 44 tabular datasets, each annotated with rich fairness-relevant metadata. Our accompanying Python package standardizes dataset loading, preprocessing, transformation, and splitting, streamlining experimental workflows. By providing a diverse and well-documented dataset corpus along with robust tooling, FairGround enables the development of fairer, more reliable, and more reproducible ML models. All resources are publicly available to support open and collaborative research.

Figures (10)

• Figure 1: The different components in the FairGround corpus. We provide a comprehensive corpus of datasets and extract diverse collections of datasets via a selection algorithm. Both the corpus, collections and individual datasets are made accessible via a Python package and web-frontend. Collections paired with reproducible dataset loading and preparation allows for novel evaluation suites.
• Figure 2: The pipeline of steps involved when loading and processing a dataset in the package. Datasets can be accessed / exported after each of the steps in the pipeline and most steps allow for configuration.
• Figure 3: Scores from a single dataset (Bank), scenario (sensitive attribute: Age), seed (80539), and metric (Equalized Odds Difference) illustrating how delta scores with respect to baseline logistic regression are calculated. Delta scores correspond to dashed lines.
• Figure 4: Relative performance and efficacy of different fairness interventions is highly variable. Relative ranking of different processing techniques across datasets and seeds (A), as well as prevalence of practical and timeout errors (B).
• Figure 5: The importance of different dataset characteristics can be highly variable between debiasing algorithms. Normalized feature importance of the 10 most important computed metadata features to predict the difference in Equalized Odds Difference across all processing methods, ordered by average importance. Feature importance for Adversarial Debiasing (in) is highlighted in orange.