AUTHOR=Ren Chunhua , Li Chaorong , Yu Yang , Yang Wanan , Guo Ruiqi 

TITLE=Density peak clustering algorithm based on weighted mutual K-nearest neighbors

JOURNAL=Frontiers in Applied Mathematics and Statistics

VOLUME=Volume 11 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/applied-mathematics-and-statistics/articles/10.3389/fams.2025.1598165

DOI=10.3389/fams.2025.1598165

ISSN=2297-4687

ABSTRACT=Ever since Density Peak Clustering (DPC) was published in Science, it has been widely favored and applied in various fields due to its concise and efficient computational theory. However, DPC has two major flaws. On the one hand, it fails to find cluster centers of low-density clusters in datasets with uneven density distribution. On the other hand, its single assignment strategy, which only assigns points to high-density clusters, can lead to incorrect clustering due to a chain reaction. To address these weaknesses, a new density peak clustering algorithm based on weighted mutual K-nearest neighbors called WMKNNDPC is proposed in this paper. WMKNNDPC offers two significant advantages: (1) It introduces the concept of mutual K-nearest neighbors by using K-nearest neighbors and inverse K-nearest neighbors, allowing for the identification of cluster centers in clusters with uneven density distribution through a new local density calculation method. (2) It includes a remaining points assignment method based on weighted mutual K-nearest neighbors, which involves two stages: first, the initial assignment of data points is done by combining mutual K-nearest neighbors and breadth-first search, and second, the membership degree of data points is calculated based on weighted mutual K-nearest neighbors for remaining points assignment. This method allows for efficient assignment based on the local distribution of points, avoiding the disadvantages of using a fixed K-value in DPC-derived algorithms based on K-nearest neighbors. The WMKNNDPC algorithm has been extensively tested on two-dimensional synthetic datasets, real datasets, facial recognition dataset and parameter analysis. The experimental results indicate that our algorithm performs the best on most datasets.