Title: Strategies for evaluation of the selected partition in cluster analysis
Authors: Osvaldo Dias Lopes da Silva - Universidade dos Acores (Portugal) [presenting]
Aurea Sandra Toledo de Sousa - Universidade dos Azores (Portugal)
Abstract: In cluster analysis there are several problems, among which, we highlight the absence of knowledge about the real structure of the data, due to the fact that we cannot recognize artificial structures imposed by the algorithms used. The aim is to present a methodological approach in order to evaluate the adequacy of the selected partition, as the most significant, based on stopping criteria rules. The comparison of a partition obtained from real data with several partitions, with the same number of clusters, obtained from randomly generated matrices in the reference hypothesis of absence of structure, allows us to test if the partition under evaluation has a structure which is relevant, that stands out from the inevitable influence of the algorithms used. The application of this methodology is exemplified with a real dataset and the main conclusions are presented and discussed.