Uncertainty-Aware Finite-Sample Quantification
MCML Authors
Abstract
Abstract
A quantification learning (QL) method estimates the label distribution of unlabeled data. Existing methods typically target the population label distribution, even though many applications and common evaluation protocols ask for the label distribution realized in a particular finite sample. This distinction matters for small target samples, where the finite-sample label distribution can differ substantially from the population label distribution. We formalize this estimand as finite-sample quantification and propose FS-UP-ACC, a Bayesian hard-classifier-based quantifier that assigns posterior mass directly to feasible finite target counts while propagating calibration uncertainty. The method retains the correction logic of Adjusted Classify & Count (ACC), admits a closedform binary update, and avoids the continuous posterior over population label distributions defined by existing Bayesian black-box-estimatorbased quantifiers. Synthetic and real-data experiments show that FS-UPACC improves point estimates and uncertainty intervals, especially for small target samples.
inproceedings DH26
ECML-PKDD 2026
1st International Workshop on Quantification and Classification under Dataset Shift at the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases. Naples, Italy, Sep 07-11, 2026.Authors
C. Damke • E. HüllermeierLinks
URLResearch Area
BibTeXKey: DH26