Position: Measure Dataset Diversity, Don't Just Claim It

TL;DR

This paper critically examines how 'diversity' in ML datasets is often vaguely defined and proposes a measurement-based approach to better conceptualize and evaluate dataset diversity.

Contribution

It introduces a measurement theory framework to analyze and improve the conceptualization and operationalization of diversity in ML datasets.

Findings

Analysis of 135 datasets reveals inconsistent diversity definitions

Applying social science principles improves diversity measurement

Recommendations for clearer, more precise dataset diversity evaluation

Abstract

Machine learning (ML) datasets, often perceived as neutral, inherently encapsulate abstract and disputed social constructs. Dataset curators frequently employ value-laden terms such as diversity, bias, and quality to characterize datasets. Despite their prevalence, these terms lack clear definitions and validation. Our research explores the implications of this issue by analyzing "diversity" across 135 image and text datasets. Drawing from social sciences, we apply principles from measurement theory to identify considerations and offer recommendations for conceptualizing, operationalizing, and evaluating diversity in datasets. Our findings have broader implications for ML research, advocating for a more nuanced and precise approach to handling value-laden properties in dataset construction.