General Quantification of Covariate and Concept Shifts

Addressing the breakdown of classical learning bounds when source and target data supports mismatch, this paper introduces $\gamma^*$-concept shifts grounded in entropic optimal transport. The authors derive a unified generalization bound that accommodates broad loss functions, continuous or discrete label spaces, and stochastic labeling. To bridge theory and practice, they design finite-sample estimators alongside DataShifts, an algorithm capable of numerically quantifying distribution shifts and evaluating target error bounds directly from empirical datasets.