New Features of Ordered Predictors Selection for Multivariate Regression and Classification

• Ano de defesa: 2019
• Autor(a) principal: Roque, Jussara Valente
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Viçosa
• Palavras-chave em Português: Variáveis (Matemática); Sistema de reconhecimento de padrões; Calibração; Mínimos quadrados; Análise discriminatória; Química Analítica
• Link de acesso: https://locus.ufv.br//handle/123456789/29234

New variable selection methods for multivariate regression and classification based on ordered predictors selection (OPS) were developed in this work. Initially, the new OPS strategies for regression were developed and applied to the six datasets used in the original OPS paper to compare their prediction performances. After that, twelve new datasets were used to test and compare the new OPS approaches for regression with other variable selection methods, genetic algorithm (GA), the interval successive projections algorithm for partial least squares (iSPA), and recursive weighted partial least squares (rPLS). Simulated datasets were used to evaluate the computational performance of variable selection methods, being then the new OPS approaches for regression, GA, iSPA, and rPLS. All methods were evaluated by using a central composite design varying the matrix dimensions of simulated datasets and the number of latent variables. For classification, OPS methods for feature selection in the discriminant analysis (OPSDA) were developed. OPSDA methods were applied to three datasets with different numbers of classes, and classification models were built using different classification methods. The new OPS approaches for regression outperformed the first OPS version and the other variable selection methods. Results showed that in addition to higher predictive capacity, the accuracy in the selection of expected variables is highly superior with the new OPS approaches for regression. The computational performance of OPS approaches was mainly influenced by the number columns of the data matrix, as well as the GA. On the other hand, iSPA and rPLS were mainly influenced by the number of rows. In classification, the OPSDA methods provided the best set of selected variables to build more predictive models using different classification methods. Besides, they could be applied to classification problems, independent of the number of classes. Overall, the new OPS methods provided the best set of selected variables to build more predictive and interpretative regression and classification models. The new OPS methods proved to be efficient for variable selection in different types of datasets. Keywords: Variable Selection, Multivariate Regression, Supervised Pattern Recognition.