FMAPLS: Bayesian label shift estimation based on dynamic dirichlet parameter adaptation
File(s)
Accepted version
Author(s)
Hu, Jiawei
Barria, Javier
Type
Journal Article
Abstract
Label shift, a critical challenge in supervised learning, occurs when the class prior distribution of test data deviates from that of training data, leading to significant degradation in classifier performance. This issue is particularly impactful in real-world applications such as medical diagnosis and satellite imaging, where a reduction in classification accuracy can cause severe consequences. Existing methods, such as Maximum A Posterior Label Shift (MAPLS), rely on strict Dirichlet hyperparameter constraints, thereby imposing a limitation on their adaptability to dynamic learning environments and class-imbalanced scenarios. To address this problem, we propose Full Maximum A Posterior Label Shift (FMAPLS), a Bayesian label shift estimation framework that dynamically co-optimizes Dirichlet hyperparameters and class priors through Expectation-Maximization (EM) iterations. By eliminating rigid hyperparameter constraints and introducing a linear surrogate function, FMAPLS enhances both expressivity and robustness in prior estimation while reducing computational complexity. Extensive experiments on CIFAR100 under shuffled long-tail and Dirichlet imbalanced test priors demonstrate FMAPLS's superiority over state-of-the-art baselines, achieving up to 50% lower KL divergence and maintaining accuracy improvements by up to 0.5% in extreme imbalance settings. The results validate the efficacy of dynamic hyperparameter adaptation in estimating label shift, particularly under high class imbalance and distributional uncertainty.
Date Issued
2025-10-16
Date Acceptance
2025-10-01
Citation
IEEE Signal Processing Letters, 2025
ISSN
1070-9908
Publisher
Institute of Electrical and Electronics Engineers
Journal / Book Title
IEEE Signal Processing Letters
Copyright Statement
Copyright © 2025 IEEE. This is the author’s accepted manuscript made available under a CC-BY licence in accordance with Imperial’s Research Publications Open Access policy (www.imperial.ac.uk/oa-policy)
License URL
Publication Status
Published online
Date Publish Online
2025-10-16