Processos t-student em classificação

• Ano de defesa: 2021
• Autor(a) principal: Assunção, Alan da Silva
• Orientador(a): Andrade, José Aílton Alencar
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Classificador de processo gaussiano; Robustez; Classificador de processo t-student; Modelagem não-paramétrica; Gaussian process classifier; Robustness; T-student process classifier; Non-parametric modeling
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/60533

Gaussian Process regression models (GPR) are excellent non-parametric alternatives for modeling complex problems, among the advantages, we can mention: good predictive performance, non-parametric flexibility, interpretability and easy computational implementation. Thus, the proposal for GP classification models is useful to deal with most diverse classification problems. However, Gaussian Process models are not robust to outliers, due to the light-tailed nature of the Gaussian distribution. In this work, we propose a new t-Student Process classifier (TPC), as an alternative to Gaussian Processes. The TPC aproach is able to deal most adequately with classification problems which input data x are contaminated by outliers. The proposed classifier had its performance evaluated with the traditional Gaussian Process classifier (GPC) in real data sets from the biomedical area, where the outliers were generated artificially. For applications in the case of binary classification, spinal diagnostic data and breast cancer diagnosis were used. For applications in the multiclass case, the set of vertebral column observations in its multiclass version was considered. The inferences about the models covered in this work were made using the NUTS method, an MCMC technique variant of Hamiltonian Monte Carlo. Due to the results of the applications carried out in this work, the TPC classifier achieved very promising results, mainly in the task of multiclass classification, in which the proposal of robustness in data contaminated by textit outliers was well attended.