General Quantification of Covariate and Concept Shifts
A paper titled 'General Quantification of Covariate and Concept Shifts' was published on arXiv on 2026-09-10. It proposes a new notion of γ*-concept shifts using entropic optimal transport, derives a general error bound unifying covariate and γ*-concept shifts, and introduces the DataShifts algorithm to quantify distribution shifts and estimate error bounds.
The paper addresses generalization under distribution shift, noting that existing learning bound theory is limited to narrow settings and not estimable from samples. It shows that the existing definition of concept shift breaks when source and target supports mismatch. Using entropic optimal transport, it defines γ*-concept shifts and derives a general error bound applicable to broad loss functions, label spaces, and stochastic labeling. It also develops estimators with concentration guarantees and the DataShifts algorithm for practical quantification.
The work introduces γ*-concept shifts via entropic optimal transport, enabling a unified error bound for covariate and concept shifts. The DataShifts algorithm provides sample-based estimation of these shifts with concentration guarantees, making the theoretical bound practically computable.
This research provides a rigorous tool for analyzing learning error under distribution shift, which is critical for deploying ML models in non-stationary environments. It may influence model monitoring and validation practices in industry.
The method could improve reliability and trustworthiness of ML systems by enabling quantitative assessment of distribution shift, potentially reducing risks in applications like finance, healthcare, and autonomous systems.
Potential next signals include empirical evaluations of DataShifts on real-world datasets, adoption of the shift quantification method in ML tooling, and further theoretical extensions to other shift types.