Título: Segmentação do SOM baseada em particionamento de grafos
Autores: Costa, José Alfredo Ferreira; Netto, Márcio Luiz A.
Resumo: Clustering methods have been studied and applied in a diversity of problems involving multidimensional data. The objective is to classify N unlabeled objects in a P-dimensional space into groups based on their similarities. Difficulties include determining the real number of categories and a metric which optimally adapt to data. Conventional methods, such as k-means, may impose a structure on data rather than finding it. This paper focuses the usage of self-organizing feature map (SOM) as a clustering tool. Although SOM had been applied as visualization tool of high-dimensional data some additional procedures are required to enable a meaningful cluster’s interpretation. It is shown that the map can be partitioned by analyzing inconsistent neighboring relations between neurons. The results are sets of connected neurons that represent data clusters. The number of clusters and its membership neurons are determined by the algorithm.
Código DOI: 10.21528/CBRN2003-096
Artigo em PDF: 6CBRN_096.PDF
Arquivo BibTex: 6CBRN_096.bib