Adaptive Local Data Density for Clustering Analysis

Density is a key statistical quantity and is widely applied in clustering analysis, but various existing approaches to computing data density have to be specified by prior information or user's experiences. Moreover, these data densities are computed by global rather than local density threshold, while any cluster consists of a group of locally high-density data in fact. Once data density is incorrectly determined, the related clustering results are guaranteed no longer. In this paper, we propose an adaptive and local method to compute data density by a simple bisecting process. All data samples in a dataset are continually bisected according to their k-nearest neighbor (k-NN) distances and adaptively grouped by their competing relations. Consequently, the computed data density perfectly is unsupervised without prior information or user's experiences. Moreover, it is a local notation and consistent in the clustering process. Experimental results validate the proposed method and its effectiveness.