Domain Adaptation with Extreme Label Shift via Locality-Aware Private Class Identification

A new arXiv preprint (2605.05567) proposes a method for domain adaptation where source and target label spaces have an inclusion relationship, creating private classes present in only one domain. Existing methods assume private class shifts are large enough to be outliers, but the authors argue that variance within a shared class can exceed differences between private and shared classes, increasing classification difficulty. They introduce a locality-aware private class identification approach based on local transportation and metric properties of optimal transport. The work addresses the challenge of extreme label shift in unsupervised domain adaptation.