Algoritmo de classificação por particionamento hierárquico

Abstract

This dissertation proposes a new method of hierarchical partitioning classification that aims not only to return a response regarding the class of an element, but also to provide more information about the classification process and the spatial arrangement of the classes along the attribute space. Through iterative partitioning and the use of concepts such as divergence between distributions, the method seeks to find regions where there is a predominant class and regions where overlap between classes makes classification more complex. Experiments with artificial and real databases were performed to demonstrate the competitiveness of the method and its advantage in separating regions of easy classification of more complex regions, both for own classification and to obtain more information on the performance of well-known methods.