A novel density peak clustering algorithm based on squared residual error

• Main Authors: Parmar, Milan, Wang, Di, Tan, Ah-Hwee, Miao, Chunyan, Jiang, Jianhua
• Other Authors: School of Computer Science and Engineering
• Format: Conference or Workshop Item
• Language: English
• Published: 2018
• Subjects: Clustering; Density Peak Clustering; DRNTU::Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/89594; http://hdl.handle.net/10220/47062
• Institution: Nanyang Technological University

The density peak clustering (DPC) algorithm is designed to quickly identify intricate-shaped clusters with high dimensionality by finding high-density peaks in a non-iterative manner and using only one threshold parameter. However, DPC has certain limitations in processing low-density data points because it only takes the global data density distribution into account. As such, DPC may confine in forming low-density data clusters, or in other words, DPC may fail in detecting anomalies and borderline points. In this paper, we analyze the limitations of DPC and propose a novel density peak clustering algorithm to better handle low-density clustering tasks. Specifically, our algorithm provides a better decision graph comparing to DPC for the determination of cluster centroids. Experimental results show that our algorithm outperforms DPC and other clustering algorithms on the benchmarking datasets.