How many labelers do you have? A closer look at gold-standard labels
This addresses a foundational problem in machine learning dataset construction, offering insights for researchers and practitioners on when to use aggregated vs. non-aggregated labels, though it is incremental as it builds on existing labeling theory.
The paper questions the standard practice of aggregating multiple labels into a gold-standard for supervised learning, showing theoretically that using non-aggregated label information can enable more feasible training of well-calibrated models, with specific conditions where it leads to faster convergence rates.
The construction of most supervised learning datasets revolves around collecting multiple labels for each instance, then aggregating the labels to form a type of "gold-standard". We question the wisdom of this pipeline by developing a (stylized) theoretical model of this process and analyzing its statistical consequences, showing how access to non-aggregated label information can make training well-calibrated models more feasible than it is with gold-standard labels. The entire story, however, is subtle, and the contrasts between aggregated and fuller label information depend on the particulars of the problem, where estimators that use aggregated information exhibit robust but slower rates of convergence, while estimators that can effectively leverage all labels converge more quickly if they have fidelity to (or can learn) the true labeling process. The theory makes several predictions for real-world datasets, including when non-aggregate labels should improve learning performance, which we test to corroborate the validity of our predictions.