IMDENCLUE: an Improved DENCLUE Clustering Based on Kernel Density Estimation Optimisation and Cluster Merging Strategy

As a representative of density-based method, DENCLUE discovers the density-attractors of data using Hill Climbing (HC), if there exists a path between significant density-attractors, data points belonging to those attractors are grouped as one cluster. However, it is very difficult to select the appropriate influence parameter and the density-attractor significance for the DENCLUE. In this paper, we propose a kernel density estimation (KDE) optimization and a cluster merging strategy for improving the performance of the traditional DENCLUE. Asymptotic mean integrated square error (AMISE) of standard KDE is introduced to calculate the optimal window width. It can help to obtain the correct global possibility density, which will lead to accurate HC results. Fine reconstruction of data as well as outlier detection will be accomplished by an optimized threshold firstly. Then, boundary density between two fine clusters is evaluated for merging purposes, which can reorganize fine-grained groups. Experiment results show that we can improve the clustering performance by using the optimized KDE and the merging strategy. Compared with the traditional DENCLUE and DBSCAN, our method gains a higher clustering performance on the data with arbitrary shapes and sizes.