Research map / Learning theory and algorithms / Fairness

A Taxonomy of Challenges to Curating Fair Datasets

NeurIPS 2024 · Learning theory and algorithms › Fairness · approach: Fairness auditing and regularization · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Interviews with 30 ML dataset curators yield a taxonomy of challenges and trade-offs in fair dataset curation across the dataset lifecycle and a five-level fairness landscape, plus recommendations.

Problem

There is limited understanding of the practical challenges and trade-offs dataset curators face when trying to build fairer ML datasets, beyond guideline-focused scholarship.

Approach

Grounded qualitative study: semi-structured interviews with 30 fair-dataset curators are thematically analyzed to build a taxonomy of challenges mapped onto the dataset lifecycle (requirements, design, implementation, evaluation, maintenance) and a social-ecological model of the broader fairness landscape (individual, discipline, organization, regulatory, socio-political).

Key results

• Thirty semi-structured interviews (45-60 min each) with ML dataset curators, 23 of 30 from university settings, were thematically analyzed into a lifecycle taxonomy of challenges.
• Challenges are organized into five landscape levels: individual, discipline, organization, regulatory, and socio-political.
• Recommendations target composition, process, and release dimensions of fairness, emphasizing systemic change rather than individual effort.

Surprising findings

• Participants were skeptical of synthetic data as a fix for bias, arguing it keeps the same underlying distribution.
• Standard practices such as majority voting and annotator agreement metrics were seen as unsuitable for fair datasets with subjective labels.

Limitations

• Sample skewed toward North America and Europe, limiting generalizability and Global South perspectives.
• Qualitative findings reflect participants' experiences and may not be universally applicable.
• Only 30 interviews, with most participants in academia.

Experimental setup

Datasets: interview transcripts from 30 ML dataset curators

Builds on: datasheets for datasets social ecological model

Closest papers by meaning

• The State of Data Curation at NeurIPS: An Assessment of Dataset Development Practices in the Datasets and Benchmarks Track (NeurIPS 2024 · similarity 0.90)
• A Systematic Review of NeurIPS Dataset Management Practices (NeurIPS 2024 · similarity 0.87)
• Position: Measure Dataset Diversity, Don't Just Claim It (ICML 2024 · similarity 0.85)
• The Fragility of Fairness: Causal Sensitivity Analysis for Fair Machine Learning (NeurIPS 2024 · similarity 0.84)
• FairDD: Fair Dataset Distillation (NeurIPS 2025 · similarity 0.84)
• Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.84)