Representation Matters: Assessing the Importance of Subgroup Allocations in Training Data
This work addresses the challenge of disparate model performance for different demographic groups, providing a framework for objective-aware data collection, though it is incremental in advancing existing discussions on fairness and representation.
The paper tackles the problem of understanding how diverse representation in training data affects machine learning performance across subpopulations, demonstrating that intentional dataset design based on subgroup allocations is key to improving both subgroup and population-level outcomes.
Collecting more diverse and representative training data is often touted as a remedy for the disparate performance of machine learning predictors across subpopulations. However, a precise framework for understanding how dataset properties like diversity affect learning outcomes is largely lacking. By casting data collection as part of the learning process, we demonstrate that diverse representation in training data is key not only to increasing subgroup performances, but also to achieving population level objectives. Our analysis and experiments describe how dataset compositions influence performance and provide constructive results for using trends in existing data, alongside domain knowledge, to help guide intentional, objective-aware dataset design.