Decentralized Clustering by Finding Loose and Distributed Density Cores

Abstract Centroid-based clustering approaches fail to recognize extremely complex patterns that are non-isotropic. We analyze the underlying causes and find some inherent flaws in these approaches, including Shape Loss, False Distances and False Peaks, which typically cause centroid-based approaches to fail when applied to complex patterns. As an alternative to current methods, we propose a hybrid decentralized approach named DCore, which is based on finding density cores instead of centroids, to overcome these flaws. The underlying idea is that we consider each cluster to have a shrunken density core that roughly retains the shape of the cluster. Each such core consists of a set of loosely connected local density peaks of higher density than their surroundings. Borders, edges and outliers are distributed around the outsides of these cores in a hierarchical structure. Experiments demonstrate that the promise of DCore lies in its power to recognize extremely complex patterns and its high performance in real applications, for example, image segmentation and face clustering, regardless of the dimensionality of the space in which the data are embedded.