Modelos alternativos para classificação em dados desbalanceados

• Ano de defesa: 2023
• Autor(a) principal: de la Cruz Huayanay, Alex
• Orientador(a): Bazán Guzmán, Jorge Luis
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa Interinstitucional de Pós-Graduação em Estatística - PIPGEs
• Palavras-chave em Português: Métricas para classificação binária; Potência Gumbel; Distribuição potência; Ligação assimétrica; Dados desbalanceados; Modelo de efeitos mistos; Asymmetric link
• Palavras-chave em Inglês: Metrics for binary classification; Power Gumbel; Power distribution; Imbalanced data; Mixed-effects model
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::PROBABILIDADE E ESTATISTICA
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/18630

In binary classification, the most used method is logistic regression model. However, several authors indicate that this model is not suitable when the data are imbalanced; for this, different asymmetric link functions as alternatives for binary response models have been proposed, for example, in recent years the power (P) and reverse power (RP) distributions have been presented. In this work we develop new properties of the P and RP distributions in the context of models for classification on imbalanced data. Also, some metrics for classification are studied through a simulation study, and an application of the studied methodology is presented. In addition, we extend the binary regression models to the case of mixed models for binary classification in the context of a longitudinal studies. To evaluate the performance of the models, a simulation study is performed. Additionally, an application is considered concerning the studied methodology in a dataset in which the response is longitudinal and imbalanced. For parameter estimation the Bayesian approach is considered using a MCMC procedure through the No-U-Turn Sampler (NUTS) algorithm. Further predictive checks, randomized Bayesian quantile residuals and a measure of Bayesian influence are considered for model diagnosis. Different models are compared using model selection criteria.