Machine Learning techniques in muliclass problems with application in sensorial analysis

Abstract

Automatic classification methods have been developed in the area of Machine Learning to facilitate the categorization of data. Among the most successful methods are Boosting and Bagging. While Bagging works by combining fit classifiers into the bootstrap samples, Boosting works by sequentially applying a sorting algorithm to reweigh versions of the training dataset, giving more weight to the erroneously classified observations in the previous step. These classifiers are characterized by satisfactory results, low computational cost, and simplicity of implementation. Given these characteristics, there is an interest in verifying the performance of these automatic methods compared to the classical methods of classification in Statistics such as Linear and Quadratic Discriminant Analysis. To compare these techniques, we have used the classification error rates of the models to improve the confidence in the use of Boosting and Bagging methods in more complex classification problem. This study applies these techniques to real and simulated data that have been composed of more than two categories in the response variable. This investigation stimulates the implementation of Boosting and Bagging, by assigning an application in Sensory Analysis. We have concluded that the automatic methods have an optimal classification performance, showing lower error rates compared to the Linear and Quadratic Discriminant Analysis in the tested applications.