Uma estratégia de seleção de variáveis baseada na otimização por colônia de formigas em análise discriminante linear

• Ano de defesa: 2020
• Autor(a) principal: Pontes, Aline Santos de
• 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: Seleção de variáveis; Colônia de formigas; Análise discriminante linear; Variable selection; Ant colony; Linear discriminant analysis; CNPQ::CIENCIAS EXATAS E DA TERRA::QUIMICA
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/23443

In this work the use of a new biophysical optimization strategy in Ants' Colony for selection of variables in classification problems via Linear Discriminant Analysis (LDA) is presented. The proposed algorithm, called ACO-LDA, was implemented in MatLab environment (as well as all other chemometric calculations) and evaluated in two databases already studied and a database developed. In the first study, UV-VIS spectrometry was used to classify four types of edible vegetable oils (corn, soybean, canola and sunflower). In the second case, NIR spectrometry is used to discriminate samples of teas with respect to their varieties (black and green) and geographical origins (Argentinean, Brazilian, and Sri Lankan). In the third case, NIR spectrometry data were used in beans seeds according to the type of cultivar: (perola, pontal and transgenic). The results obtained were compared to other methods of selection of variables good established in the literature as Genetic Algorithm (GA), the Algorithm of Successive Projections (SPA), both associated to LDA modeling and the Discriminante Analysis by Partial Least Squares (PLS-DA). In the first two applications, a correct prediction classification rate was: 100%, 95.7% and 100% of the ACO-LDA, GA-LDA and PLS-DA models respectively for classifying the properties of vegetable plants (UV-Vis). While for the classification of teas (NIR), a correct classification rate was 92%, 84% and 100% of the ACO-LDA, GA-LDA and PLS-DA models respectively. In the third application, a correct classification rate was 100% for the four models (ACO-LDA, GA-LDA, SPA-LDA and PLS-DA) in the classification of debts. Thus, the results indicated that ACO-LDA is promising to classify from NIR and UV-Vis spectrometric data, making it possible to provide results compatible with those chosen by powerful and versatile chemometric tools.