A Novel Method for Automatic Parameter Tuning in Density-Based Clustering using Local Density Variations

Abstract

Density based clustering is clustering technique which has the ability to find the arbitrary shaped clusters. One of the most important reason for the success of density-based clustering is its ability to obtain the clusters of arbitrary shapes. DBSCAN is pioneer and one of the most studied algorithms in this category of clustering. The success of DBSCAN attracted the attention of researchers. Due to its ability of forming the arbitrary shape clusters and noise detection, DBSCAN is used as a benchmark algorithm in density-based clustering technique. DBSCAN also have few limitations, one of the major limitations it suffers is the manual input parameters. DBSCAN used two manual values as an input namely Epsilon and MinPts. Epsilon define the searching area and hence plays a very important and decisive role in cluster formation. With very small value of epsilon, one may obtain too many clusters. It will be difficult to detect the noise with small value of epsilon. When the value of epsilon is larger, the size of cluster will be larger. The large size of cluster will affect the ability of algorithm to detect noise. As the epsilon plays a very important role in quality of clusters. The proposed approach addressed this issue by automatic determination of epsilon. The proposed algorithm provides an optimal value of epsilon.