Partição Espacial Utilizando Triangulação de Delaunay e Hiperplanos de SVM para Classificação Multiclasse

Luciana Babberg AbiuziCarlos Henrique Quartucci Forster

Support Vector Machines are considered robust tools with great ability for generalization. They were originally designed to perform classification between two pattern classes; however, many real applications require a method for discrimination among multiple classes, such as biometric identification. We propose in this article a mechanism to construct the spatial subdivision of the classification feature space by combining the hyperplanes obtained through the training of SVMs. The geometrical structure of the partition follows the Delaunay Triangulation having the sample means of each class as vertices. By visualizing the results obtained in the plane, we are able to assess the feasibility of generating regions for multiclass classification.

Caso o link acima esteja inválido, faça uma busca pelo texto completo na Web: Buscar na Web

Biblioteca Digital Brasileira de Computação - Contato:
     Mantida por: