This is the official code repository for "LaSCal: Label-Shift Calibration without target labels", published at NeurIPS 2024.
In this work, we propose LaSCal, a label-free, consistent calibration error estimator for label shift, paired with a post-hoc calibration strategy for unsupervised calibration. Label shift assumes changes in the label distribution while keeping the conditional distribution fixed. Our method is validated through extensive empirical analysis, demonstrating its reliability across diverse modalities, model architectures, and shift intensities.