Density Peaks Clustering Based on Local Minimal Spanning Tree

Abstract

The fake center is a common problem of density-based clustering algorithms, especially for datasets with clusters of different shapes and densities. Clustering by fast search and find of density peaks (DPC) and its improved versions often ignore the effect of fake centers on clustering quality. They usually have a poor performance even the actual number of centers are used. To solve this problem, we propose a density peaks clustering based on local minimal spanning tree (DPC-LMST), which generates initial clusters for each potential centers first and then introduce a sub-cluster merging factor (SCMF) to aggregate similar sub-clusters. Meanwhile, we introduce a new strategy of representative points to reduce the size of data and redefine local density ρi and distance δi of each representative point. Furthermore, the hint of γ is redesigned to highlight true centers for datasets with clusters of different densities. The proposed algorithm is benchmarked on both synthetic and real-world datasets, and we compare the results with K-means, DPC, and the three state-of-the-art improved DPC algorithms.