Unification of Algorithms for Quantification and Unfolding

Abstract

Quantification is the supervised learning task of predicting the prevalences of classes in a data sample. Physics literature knows the same task under a different name: unfolding. However, the literature on quantification and the literature on unfolding are largely disconnected from each other. We bridge this interdisciplinary gap by proposing a common framework that integrates algorithms from both fields in a unified form. Instantiations of our framework differ from each other in terms of the loss functions, the regularizers, and the feature transformations they employ.