Quantificação em problemas com mudança de domínio

Abstract

Several machine learning applications use classifiers as a way of quantifying the prevalence of positive class labels in a target dataset, a task named quantification. For instance, a naive way of determining what proportion of positive reviews about given product in the Facebook with no labeled reviews is to (i) train a classifier based on Google Shopping reviews to predict whether a user likes a product given its review, and then (ii) apply this classifier to Facebook posts about that product. Unfortunately, it is well known that such a two-step approach, named Classify and Count, fails because of data set shift, and thus several improvements have been recently proposed under an assumption named prior shift. However, these methods only explore the relationship between the covariates and the response via classifiers and none of them take advantage of the fact that one often has access to a few labeled samples in the target set. Moreover, the literature lacks in approaches that can handle a target population that varies with another covariate; for instance: How to accurately estimate how the proportion of new posts or new webpages in favor of a political candidate varies in time? We propose novel methods that fill these important gaps and compare them using both real and artificial datasets. Finally, we provide a theoretical analysis of the methods.