Density change consistency clustering from density extreme

Abstract

Density-based clustering algorithms can identify clusters with arbitrary shapes and sizes. However, the algorithms still have difficulty detecting clusters with heterogeneous density within or between clusters. To overcome the weakness, we propose an effective clustering algorithm, called ETCD. First, starting from the density extremes, ETCD identifies the points satisfying density change consistency to generate the initial clusters, in which the density differences between neighboring points are slight and the points density decrease from center to boundary. Next, some initial clusters are merged based on density difference between clusters and attachness of clusters borders. To comprehensively demonstrate the performance of ETCD, we benchmarked ETCD on over 40 datasets with 4 classical baselines and 7 State-Of-The-Arts. Experimental results indicate ETCD outperforms the eleven baselines in most cases. ETCD is robust to density change within and between clusters, as well as clusters with different sizes, arbitrary shapes and various compactness.