Adaptive Density-Based Spatial Clustering of Applications with Noise (ADBSCAN) for Clusters of Different Densities

Abstract

Finding clusters based on density represents a significant class of clustering algorithms. These methods can discover clusters of various shapes and sizes. The most studied algorithm in this class is the Density-Based Spatial Clustering of Applications with Noise (DBSCAN). It identifies clusters by grouping the densely connected objects into one group and discarding the noise objects. It requires two input parameters: <i>epsilon</i> (fixed neighborhood radius) and <i>MinPts</i> (the lowest number of objects in <i>epsilon</i>). However, it can’t handle clusters of various densities since it uses a global value for epsilon. This article proposes an adaptation of the DBSCAN method so it can discover clusters of varied densities besides reducing the required number of input parameters to only one. Only user input in the proposed method is the <i>MinPts</i>. <i>Epsilon</i> on the other hand, is computed automatically based on statistical information of the dataset. The proposed method finds the core distance for each object in the dataset, takes the average of these distances as the first value of <i>epsilon</i>, and finds the clusters satisfying this density level. The remaining unclustered objects will be clustered using a new value of <i>epsilon</i> that equals the average core distances of unclustered objects. This process continues until all objects have been clustered or the remaining unclustered objects are less than 0.006 of the dataset’s size. The proposed method requires <i>MinPts</i> only as an input parameter because <i>epsilon</i> is computed from data. Benchmark datasets were used to evaluate the effectiveness of the proposed method that produced promising results. Practical experiments demonstrate that the outstanding ability of the proposed method to detect clusters of different densities even if there is no separation between them. The accuracy of the method ranges from 92% to 100% for the experimented datasets.