PopFS Selects Robust Features for Diverse Populations

A new research paper presents PopFS, a novel feature selection method designed to address the challenge of deploying models across diverse populations. Unlike standard approaches that optimize for a single large population or robust methods that force a shared model, PopFS learns one universal feature set that works effectively for multiple heterogeneous groups, allowing each group to develop its own specific predictive model. PopFS incorporates a customizable welfare objective, enabling practitioners to weigh the overall predictive accuracy against ensuring adequate performance for populations that might otherwise be underserved. To scale this objective, the method first uses multitask sparse learning to narrow down potential features, then directly searches for optimal feature sets by ranking additions and swaps, only fully refitting a small shortlist. The approach has shown consistent strong performance on both average and worst-case population metrics across eight population splits from six prediction tasks, including a COVID-19 nowcasting study where adjusting the welfare objective improved outcomes for the least-served states with minimal impact on overall performance.