Um classificador baseado na profundidade de Tukey para geração de uma aproximação do ponto de Bayes

Abstract

Ensembles for creating classifiers, in the area of machine learning, have been widely used since they are a more efficient and more accurate way when compared to a single classifier. This increase in the accuracy of ensembles is directly related to the capacity to combine the acquired knowledge from the classifier components to generate a solution that surpasses the solutions of each individual. An ensemble can be understood as a group of members and, from that point of view, it can be related to the field of statistics and, therefore, apply concepts of depth data to these members. A well-known concept is Tukey’s depth, also known as the half-space depth, which is defined as the smallest area of the probability of a closed half-space that contains an analyzed point, thus resulting in the depths of data in a multivariate group. The deepest point regarding Tukey’s depth is commonly seen as Tukey’s median. This dissertation has as its main objective to develop tools and analyses to be applied to an ensemble of classifiers, being the version space created from balanced Perceptrons, aiming to find the representative member that wins the most in the voting or the deepest member, then coining Tukey’s Perceptron, besides presenting the algorithms based on the concepts of the depth-first search and breadth-first search for identifying the winner with most votes. An experimental study has been performed to evaluate the proposed methods, and the results show that the introduced methods are able to overcome, in most cases, other algorithms such as Support Vector Machine, Bayes Point Machine, Version Space Reduction Machine, and Evolutionary Analytic Center Classifier.