Non-stationary and unpredictable data distributions in classification and quantification

• Ano de defesa: 2020
• Autor(a) principal: Reis, Denis Moreira dos
• 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 com dados positivos e não- rotulados; Classificação; Classification; Concept drift; Dados não estacionários; Mudança de conceito; Nonstationary data; Positive and unlabeled learning; Quantificação; Quantification
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-27072020-174834/

In the last years, we observed a crescent academic interest on nonstationary data. On the one hand, differences between the data that was used to induce a model and the data that is found after the model is deployed cause a decrease of performance for several tasks, such as classification. On the other hand, in several tasks, such as quantification, we are explicitly interested in measuring how a distribution changes over time. For any of these problems, however, we generally run into solutions that rely on strong assumptions, which are impractical or even impracticable in real world applications. In this thesis, we provide solutions that rely on less restrictive and/or more realistic assumptions in order to allow such methods to be employed in real applications. In the concept drift detection area, we introduce unsupervised drift detection methods that allow for performing classification and quantification without ever requesting true labels after deployment. In the quantification area, we bootstrap a new research topic called one-class quantification. Similarly to one-class classification, in one-class quantification we are able to avoid strong assumptions regarding the negative class, which is deemed unpredictable. Our experimental results are promising and provide favorable evidences about the viability of solutions that are focused on solving real-world problems.