Abstract
•New density-based clustering center initialization method called HDCCI is proposed.•The viewpoint with the highest density can help find the real data structure.•The DVPFCM algorithm is proposed based on HDCCI and new viewpoint.•DVPFCM combines the advantages of FCM and PFCM using parameter renewal.•Evaluation indexes and convergence speed of DVPFCM are superior to other algorithms.
Fuzzy clustering algorithms are usually data-driven. Recently, knowledge has been introduced into these methods to form knowledge-driven and data-driven fuzzy clustering algorithms. However, these algorithms still have the problems of sensitivity to clustering center initialization and a lack of robustness, in general. There is a genuine need for a sound acquisition of viewpoints. In this study, a new fuzzy clustering algorithm driven by data and knowledge named Density Viewpoint-induced Possibilistic Fuzzy C-Means (DVPFCM) is put forward. To begin with, we propose a new method to calculate the density radius, which determines the density range of each data point. Based on this, we establish a Hypersphere Density-based Clustering Center Initialization method (HDCCI), which can obtain the initial clustering centers located in the denser region of the dataset. Furthermore, the high density point obtained by the HDCCI method is taken as a new viewpoint and integrated into the clustering algorithm. The new viewpoint helps to speed up the convergence of the algorithm. It can also guide the clustering algorithm to discover the data structure. Finally, on the basis of the HDCCI method, the idea of high-density viewpoint is introduced, and the advantages of FCM (Fuzzy C-Means) and PFCM (Possibilistic Fuzzy C-Means) are combined, and then the DVPFCM algorithm is proposed. Through experimental studies including some comparative analyses, it is demonstrated that the DVPFCM algorithm is better in several different ways in terms of initializing clustering centers and values of some performance indexes. It also exhibits better performance in determining the distance between the computed clustering centers and the reference centers.