A new local density for density peak clustering

Abstract

Density peak clustering is able to recognize clusters of arbitrary shapes, so it has attracted attention in academic community. However, existing density peak clustering algorithms prefer to select cluster centers from dense regions and thus easily ignore clusters from sparse regions. To solve this problem, we redefine the local density of a point as the number of points whose neighbors contain this point. This idea is based on our following finding: whether in dense clusters or in sparse clusters, a cluster center would have a relatively high local density calculated by our new measure. Even in a sparse region, there may be some points with high local densities in our definition, thus one of these points can be selected to be the center of this region in subsequent steps and this region is then detected as a cluster. We apply our new definition to both density peak clustering and the combination of density peak clustering with agglomerative clustering. Experiments on benchmark datasets show the effectiveness of our methods.