Regularização de classificadores geométricos de margem larga baseados no grafo de Gabriel

Abstract

The present work is aimed at the study of new ways to build regularization based only on information extracted from the Gabriel graph. There are two main contributions: first, a preliminary study evaluates how the Gabriel graph can be used in the regularization of RBF neural networks and how this structure can be informative. Characteristics extracted from the graph were used to remove radial functions estimated by the CG-RBF algorithm, which also uses the graph in its construction. Second, a novel filtering approach is proposed for a classifier designed with information extracted from the Gabriel graph, the CHIP-CLASS. This classifier does not use neither optimization algorithms nor parameter definition by the user. Previous work has shown that efficient classifiers can be designed as such. However, there is still much to progress in regularization of these classifiers. The results show that a special set of Gabriel graph vertices is very informative of the classes separation region and that the filtering of samples based on characteristics of the graph can be used to control the capacity of the proposed model.