Melhoria dos Processos de Classificação Espectral: Um Estudo sobre Redes Neurais, Técnicas de Normalização e Seleção de Variáveis

Abstract

In recent years, several studies using FTIR spectroscopy and multivariate analysis have been carried out. Their results show the efficiency, feasibility, accuracy, and potential of the method in identifying complex patterns, classifying and discriminating samples, and automating processes. This paper presents a methodology that involves the use of normalization techniques, variable selection, and neural networks for the identification of single nucleotide polymorphisms in bovine DNA (homozygous or heterozygous). The spectral data were pre-processed using the Standard Normal Variate and first and second order derivatives. Principal Component Analysis (PCA) and Boruta were applied as dimensionality reduction techniques and variable selection to feed the neural network. Internal validation techniques using leave-one-out cross validation were used. Optimization of the classifier’s hyperparameters based on improving accuracy and external validation was applied to verify the robustness of the model. The best result demonstrated an internal validation accuracy of 95.33% and external validation of 98%, which is better than the 90% accuracy obtained by the RIOS et. al. methodology for the same classes. This corroborates the use of the method for improving classification processes.