CDBSCAN: Density clustering based on silhouette coefficient constraints
Guo Jin-Heng, Jiaxiang Lin, Zhang Zhen-Chang, Ling Han-Yu
Abstract
Aiming at the problem that the edge points are difficult to be accurately divided in the DBSCAN algorithm, a density clustering algorithm based on silhouette coefficient constraints (CDBSCAN) is proposed. The CDBSCAN improves the clustering accuracy with the silhouette coefficient as the criterion. Firstly, the data is preliminary classified by the DBSCAN algorithm and the number of data points in the formed clusters is calculated. Then both noise data and clusters with fewer data are listed as potential noise data. Subsequently, each data point in the noise set is classified again according to the silhouette coefficient. Finally, experiments are conducted on both synthetic and public datasets, and the result shows that CDBSCAN has better clustering results, especially in the discrimination of data points on the clustering edge.