Classificador por arestas de suporte (CLAS): métodos de aprendizado baseados em Grafos de Gabriel

Abstract

This work presents a methodology directed to pattern classification problems. The goal is to design large margin classifiers where the information necessary is obtained from the geometric structure of data. Through the Gabriel graph, the data set is turned into a planar graph, where the edges with vertices of distinct labels corresponds to the samples which are on the margin of separation between the classes. These edge set is named as support edges and forms the basis for the development of a family of methods, such as a decision-maker for multi-objective learning of neural networks; a strategy for selecting parameters in RBF neural networks. Finally, the design of new large margin classifiers. Results with benchmarks known in the literature show that our approaches maximize the margin and increase the classifier generalization ability .