Reproduce “Detecting and Correcting for Label Shift with Black Box Predictors” on OpenResearch

About this paper

Zachary C. Lipton, Yu-Xiang Wang, Alex Smola

Researchers from Carnegie Mellon University and Amazon AI developed a framework to detect and correct for label shift in machine learning models using arbitrary black box predictors. Their method consistently estimates target label distributions and importance weights, enabling robust shift detection and improving classifier accuracy by adapting to changing label prevalences in high-dimensional data.

Key insights

• Under label shift, a change in the marginal label distribution p(y) is equivalent to a change in the marginal distribution of the black box predictor's outputs p(f(x)), simplifying shift detection to a lower-dimensional space.
• The target label distribution can be estimated by solving a linear system where the expected confusion matrix of the predictor (on source data) acts as a linear operator mapping target label prevalences to target output prevalences.
• The accuracy of shift estimation is inversely proportional to the smallest eigenvalue of the predictor's confusion matrix, meaning a more discriminative predictor leads to more precise estimations.

Summary by alphaXiv.