One-Class Quantification.

This paper proposes one-class quantification, a new Machine Learning task. Quantification estimates the class distribution of an unlabeled sample of instances. Similarly to one-class classification, we assume that only a sample of examples of a single class is available for learning, and we are interested in counting the cases of such class in a test set. We formulate, for the first time, one-class quantification methods and assess them in a comprehensible open-set evaluation. In an open-set problem, several “subclasses” represent the negative class, and we cannot assume to have enough observations for all of them at training time. Therefore, new classes may appear after deployment, making this a challenging setup for existing quantification methods. We show that our proposals are simple and more accurate than the state-of-the-art in quantification. Finally, the approaches are very efficient, fitting batch and stream applications. Code related to this paper is available at: https://github.com/denismr/One-class-Quantification.