2DBSCAN with Local Outlier Detection

Abstract

This research is related to designing a new algorithm which is based on the existing DBSCAN algorithm to improve the quality of clustering. DBSCAN algorithm categorizes each data object as either a core point, a border point or a noise point. These points are classified based on the density determined by the input parameters. However, in DBSCAN algorithm, a border point is designated the same cluster as its core point. This leads to a disadvantage of DBSCAN algorithm which is popularly known as the problem of transitivity. The proposed algorithmーtwo DBSCAN with local outlier detection (2DBSCAN-LOD), tries to address this problem. Average silhouette width score is used here to compare the quality of clusters formed by both algorithms. By testing 2DBSCAN-LOD on different artificial datasets, it is found that the average silhouette width score of clusters formed by DBSCAN-LOD is higher than that of the clusters formed by DBSCAN.