An Alternative Parameter Free Algorithm to DBSCAN Method by Using Data Point Positioning Analysis (DBSCAN-DPPA)

Abstract

Density-based spatial clustering of applications with noise (DBSCAN) is a powerful unsupervised clustering method for its ability to manage noises and arbitrary cluster shapes without the need to pre-determine the total clusters. However, the performance of DBSCAN is dependable to right choice of its two initial parameters – MinPts and Eps. Much research had been done to overcome the challenges by reducing the dependencies of these two parameters or automatically determine the values. This paper will review some of these various techniques related to reducing the dependencies of two DBSCAN initial parameters and subsequently proposes the new algorithm called DBSCAN-DPPA which do not require the initial setting of the two parameters by using Data Point Positioning Analysis. The algorithm is simple as it does not need any parameters to be initially assigned manually as it performs analysis on the positions of data points to determine the 1-NN and Max-NN. The algorithm is applied on 13 benchmark datasets that have been applied in many clustering algorithms. The performance of the algorithm is visually compared with the three-dimensional graph plotting at various angles to prove the number of clusters. The results show that the modified DBSCAN-DPPA algorithms are comparable to the performance of the traditional DBSCAN algorithm such that it manages to detect arbitrary cluster shapes, identify the number of clusters and manage the data sets with noises.