Desenvolvimento de um método de clustering baseado em grafos de similaridade

Abstract

The clustering process consists in, starting from a dataset, extracting clusters which contains similar data. On the other hand, not all the clustering methods aim to reduce or remove the overlap between identified clusters. Therefore, the present work proposes the development of a clustering method based on similarity graphs extracted from metric spaces containing feature vectors, which objective is to minimize the overlap between clusters. The proposed clustering method is a combination of DBSCAN and range query (Rq) techniques, and introduces the k-nearest neighbors query (kNNq) as an upper bound of the neighborhood observed by the Rq. The overlap factor is a ratio which considers the number of elements which are in the neighborhood of more than one cluster. Over the input parameter combinations and test spaces, the clustering method presented better or equivalent results to those obtained by DBSCAN, both on the identified clusters and on the evaluation metrics.