Seleção de variáveis robustas para transferência de modelos de classificação empregando o algoritmo das projeções sucessivas

• Ano de defesa: 2017
• Autor(a) principal: Milanez, Karla Danielle Tavares de Melo
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Química; Programa de Pós-Graduação em Química; UFPB
• Palavras-chave em Português: Transferência de classificação multivariada; Modelagem robusta; Algoritmo das projeções sucessivas; Métodos de padronização; Etanol hidratado combustível; Azeite de oliva extra virgem; Multivariate classification transfer; Robust modeling; Successive projections algorithm; Standardization methods; Hydrated ethyl fuel; Extra virgin olive oil; CNPQ::CIENCIAS EXATAS E DA TERRA::QUIMICA
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/11879

This work proposes two new criteria for selection of robust variables for classification transfer employing the successive projections algorithm (SPA). These variables are used to construct models based on linear discriminant analysis (LDA) that are robust to the differences between the responses of the instruments involved or to the experimental conditions. For this purpose, transfer samples are included in the calculation of the cost for each subset of variables under consideration. The proposed methods are evaluated for four datasets involving identification of adulteration of hydrated ethanol fuel (HEF) and extra virgin olive oil (EVOO). To investigate HEF, near infrared (NIR) spectroscopy was used. In the EVOO study, were used UV-Vis spectrometry, molecular fluorescence spectrometry and digital images. In all cases, better classification transfer results using the two criteria, obtained for a test set measured in the secondary instrument, were compared with direct standardization (DS) and piecewise direct standardization (PDS). When one of the criteria was applied to the test set measured in the secondary instrument, the accuracy of the model increased by about 50%, 33%, 3% and 12% for NIR, UV-Vis, fluorescence emission and digital imaging data, respectively. These results are compatible, sometimes superior to those obtained by the standardization methods, demonstrating that, when the differences between the instrumental responses did not present a drastically high magnitude (NIR and UV-Vis), either of the criteria proposed can be used for building robust models as an alternative to the standardization of spectral responses for transfer of classification. For the data with a large difference between the responses of each instrument (fluorescence and digital images), the spectra needed to be corrected with DS and PDS standardization to perform classification transfer effectively. The results suggest that the proposed approach is a promising alternative to full recalibration of the model or standardization methods, especially if the primary and secondary instruments are not located in the same laboratory, if the samples are deteriorated, when it is difficult to transport the samples or if the primary instrument is no longer available.