A Density-Peak-Based Clustering Method for Multiple Densities Dataset

Abstract

Clustering methods in data mining are widely used to detect hotspots in many domains. They play an increasingly important role in the era of big data. As an advanced algorithm, the density peak clustering (DPC) algorithm is able to deal with arbitrary datasets, although it does not perform well when the dataset includes multiple densities. The parameter selection of cut-off distance dc is normally determined by users’ experience and could affect clustering result. In this study, a density-peak-based clustering method is proposed to detect clusters from datasets with multiple densities and shapes. Two improvements are made regarding the limitations of existing clustering methods. First, DPC finds it difficult to detect clusters in a dataset with multiple densities. Each cluster has a unique shape and the interior includes different densities. This method adopts a step by step merging approach to solve the problem. Second, high densities of points can automatically be selected without manual participation, which is more efficient than the existing methods, which require user-specified parameters. According to experimental results, the clustering method can be applied to various datasets and performs better than traditional methods and DPC.