Binary quantification in non-stationary scenarios

• Ano de defesa: 2019
• Autor(a) principal: Maletzke, André Gustavo
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado de máquina; Armadilhas de mosquito; Concept drift; Contagem; Counting; Machine learning; Mixture models; Modelos de mistura; Mosquito traps; Mudança de conceito
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-19032020-091709/

Quantification is a Machine Learning task similar to classification in the sense that it learns from a training set with labeled data. However, quantification is not interested in predicting the class of each observation, but rather measure the representativeness of each class in the test set. This subtle difference between classification and quantification requires specific algorithms, performance measures, and experimental designs. Moreover, most of the existing quantification algorithms were developed for well-controlled scenarios that rely on the assumption that the only change from training to test data is in the prior probability of the classes. This thesis focuses on providing improvements in quantification algorithms as well as the experimental design, including more realistic assumptions. Specifically, the main contributions of this thesis are the following: (i) the first algorithm to quantify non-stationary data under the concept drift presence; (ii) an unsupervised drift detector that are insensible to class imbalance explicitly; (iii) a mixture model framework for quantification with a deep experimental study, redefining the best parametrization of this sort of method; (iv) we show that the batch size, an ignored question in the literature, changes the ranking of quantification algorithms and thus we proposed a metalearning framework to select the best quantifier dynamically; (v) we describe how existing quantifiers are affected under score quality variability and, as response of it, we proposed a new algorithm that accurately quantifies while allowing changes in the quality of the scores, and; (vi) we show the applicability of our proposals in real-world problem and the efforts to contribute with development of new mechanism for trapping mosquitoes.