VarDenGrid: A New Variable Density Clustering Algorithm with Grid-Based Initialization

Among all clustering methods, the density-based clustering algorithm is one of the most powerful methods for discovering arbitrary-shaped clusters. However, almost all the renowned density-based algorithm requires user input(s), which are sometimes difficult to determine. Additionally, some of them like DBSCAN are unable to get variable density clusters. We introduce a new algorithm VarDenGrid in the paradigm of density-based clustering that will not require any user input(s) and also will be able to obtain variable density clusters. VarDenGrid starts with grid-based initialization of prospective clusters. Then the initial clusters are modified based on a novel density-based clustering algorithm proposed here. Each prospective cluster determines its density (Epsilon) by the maximum of the nearest neighbor distance of the objects in the cluster. We compared the results obtained using the new algorithm to that of well-known state-of-the-art density-based clustering algorithms DBSCAN and OPTICS on 15 real-world data sets from the UCI data repository and noted that performances of VarDenGrid favorably compare with these well-known state-of-the-art algorithms.