The Two Faces of Outliers in Distributionally Robust Learning: Exclusion and Emphasis via One-Sided Partial Optimal Transport

Introduction:

Rare or unusual samples are often the ones a model learns least from, even when they represent important real-world conditions. This paper introduces an upper-tail learning framework that deliberately gives more attention to these difficult but informative samples. By combining one-sided partial optimal transport with Wasserstein distributionally robust optimization, the method improves learning in underrepresented regimes while remaining robust to uncertainty.