A density-based method for the identification of disjoint and non-disjoint clusters with arbitrary and non-spherical shapes

Abstract

The ability of clustering methods to build both disjoint and non-disjoint partitionings of data has become an important issue in unsupervised learning. Although this problem has been studied during the last decades resulting in several proposed overlapping clustering methods in the literature, most of existing methods fail to look for clusters having arbitrary and non-spherical shapes. In addition, most of these existing methods require to pre-configure the number of clusters in prior, which is not a trivial task in real life application of clustering. To solve all these issues, we propose in this work a new density based overlapping clustering method, referred to as OC-DD, which is able to detect both disjoint and non-disjoint partitioning even when boundaries between clusters have complex separations with arbitrary forms and shapes. The proposed method is based on density and distances to detect highly dense regions and connected groups in data without the necessity to pre-configure the number of clusters. Experiments performed on artificial and real multi-labeled datasets have shown the effectiveness of the proposed method compared to the existing ones.