Clustering algorithm based on characteristics of density distribution

Abstract

Density-based clustering algorithms, which are important algorithms for the task of class identification in spatial database, have many advantages such as no dependence on the number of clusters, ability to discover clusters with arbitrary shapes and handle noise. However, clustering quality of most density-based clustering algorithms degrades when the clusters are of different densities. To address this issue, this paper brings forward a clustering algorithm based on characteristics of density distribution--CCDD algorithm. Firstly, it divides data space into a number of grids. Secondly, it re-divides data space into many smaller partitions, according to each grid's one-dimensional or multi-dimensional characteristics of density distribution. Finally, it uses an improved DBSCAN algorithm, which chooses different parameters according to each partition's local density, to cluster respectively. The experimental results show that CCDD algorithm, which is superior in quality and efficiency to DBSCAN algorithm, can find clusters with arbitrary shapes and different densities in spatial databases with noise.