Reduced, Reused and Recycled: The Life of a Dataset in Machine Learning Research

TL;DR

This study empirically examines how benchmark datasets are used in machine learning by analyzing the Papers With Code corpus from 2015 to 2020. It shows increasing concentration on a small set of datasets within many task communities, widespread adoption of datasets originally created for other tasks, and a heavy concentration of dataset origins in a small number of elite institutions. Using the Gini coefficient and related metrics, the authors quantify within-task, cross-task, and institutional concentration, and reveal broader implications for scientific rigor, ethics, and equity in AI research. The work calls for reforms in dataset development and evaluation to mitigate inequities and improve ecological validity in benchmarking practices.

Abstract

Benchmark datasets play a central role in the organization of machine learning research. They coordinate researchers around shared research problems and serve as a measure of progress towards shared goals. Despite the foundational role of benchmarking practices in this field, relatively little attention has been paid to the dynamics of benchmark dataset use and reuse, within or across machine learning subcommunities. In this paper, we dig into these dynamics. We study how dataset usage patterns differ across machine learning subcommunities and across time from 2015-2020. We find increasing concentration on fewer and fewer datasets within task communities, significant adoption of datasets from other tasks, and concentration across the field on datasets that have been introduced by researchers situated within a small number of elite institutions. Our results have implications for scientific evaluation, AI ethics, and equity/access within the field.

Figures (8)

• Figure 1: Top: Predicted concentration on datasets across task communities over time. Gini predicted by Model 1 holding task size/age at means. Green plots show the estimated effects of the full dataset, other colors are fixed effects for categories. 95% confidence intervals shown. Bottom: Distributions of concentrations. Higher Gini indicates greater concentration on fewer datasets. We observe significant spread of Gini across tasks, with the median increasing over time.
• Figure 2: Adoption (Top) and Creation (Bottom) Proportions for PWC Parent Tasks. Full dataset in green, tasks in the Computer Vision category in purple, Natural Language Processing tasks in orange, and Methods tasks in red,. Red dot and line in boxplot indicate median. Width of violins indicates distribution of tasks.
• Figure 3: Increases in concentration of dataset usages on institutions and datasets (non-task specific) over time.Left: Map of dataset usages per institution as of June 2021. Dot size indicates number of usages. Blue dots indicates for-profit institutions and orange dots indicate not-for-profit. Institutions accounting for 50%+ of usages labeled. Right: Gini coefficient for concentration of dataset usages across the whole PWC dataset over time for both institutions and datasets. Ribbons indicate 95% CI; dot size indicates number of usages that year.
• Figure 4: Top datasets used across Image Generation and Face Recognition task communities: (a) Origin task communities of top Image Generation datasets. Only 7.49% of Image Generation papers in PWC evaluate on datasets developed for Image Generation. (b) Names of top Image Generation datasets. Only one of the top datasets, FFHQ ffhq, was developed for the task. (c) The small number of datasets in usage within the high stakes domain of Face Recognition. Two of the datasets, MegaFace kemelmacher2016megaface and MS-Celeb-1M msceleb (in white), have been recently retracted, the latter due to serious ethical violations exposing_msceleb.
• Figure A1: Increases in concentration on datasets within task communities over time. Higher Gini coefficient indicates greater concentration on fewer datasets. We observe significant spread of Gini across different task communities, with the median trending upwards over time for all modalities. Green is the full dataset, other colors indicate subsets of the data by PWC task category.