Generalized Bayes Quantification Learning under Dataset Shift

Quantification learning is the task of prevalence estimation for a test population using predictions from classifier trained on different population. methods assume that sensitivities and specificities are either perfect or transportable training to These assumptions inappropriate in presence dataset shift, when misclassification rates not representative those under shift has been addressed only single-class (categorical) assuming knowledge true labels small subset We propose generalized Bayes quantification (GBQL) uses entire compositional probabilistic classifiers allows uncertainty class limited labeled data. Instead positing full model, we use model-free Bayesian estimating equation approach data Kullback–Leibler loss-functions based first-moment assumption. The idea will be useful analysis general as it robust generating mechanisms 0’s 1’s outputs thereby including categorical special case. show how our method yields existing approaches cases. Extension an ensemble GBQL multiple yielding inference inclusion poor discussed. outline fast efficient Gibbs sampler rounding coarsening approximation loss functions. establish posterior consistency, asymptotic normality valid coverage interval estimates GBQL, which first theoretical results local also finite sample concentration rate. Empirical performance demonstrated through simulations real with evident shift. Supplementary materials this article available online.