Density peaks clustering algorithm with neighborhood optimization and micro-cluster merging
Date
2025
Authors
Lv, L.
Tan, H.W.
Xiao, R.B.
Pan, J.S.
Wang, H.
Lee, I.
Editors
Advisors
Journal Title
Journal ISSN
Volume Title
Type:
Journal article
Citation
Cluster Computing, 2025; 28(15):1-17
Statement of Responsibility
Conference Name
Abstract
The definition of local density for the Density Peaks Clustering (DPC) algorithm ignores structural differences within the data, making it hard to process datasets with significant density differences; the allocation strategy has chain effects, which can easily lead to a large number of samples in manifold clusters being incorrectly allocated; the value of the cutoff distance is difficult to determine, and its choice affects the accuracy of the clustering results. In view of the above defects, a density peaks clustering algorithm with neighborhood optimization and micro-cluster merging (DPC-NM) is proposed. Firstly, the optimal number of nearest neighbors is determined based on the dataset's structure using the natural neighbor method. This approach adaptively calculates the local density of K-nearest neighbors, avoiding the impact of parameter choice on the clustering results. Secondly, the local density is used to identify representative points and core representative points. The local density is then further optimized by the members of representative points and the K-nearest neighbors of core representative points, enhancing the density gap between clustering centers and non-centers. Finally, considering the samples' neighborhood information and structural characteristics, the micro-cluster similarity is defined to perform micro-cluster merging, thereby avoid incorrect allocation and subsequent chain effects. Experimental results on synthetic and real datasets demonstrate that DPC-NM achieves superior performance compared to DPC and its variants, and performs well in image segmentation.
School/Discipline
Dissertation Note
Provenance
Description
Access Status
Rights
Copyright 2025 The Authors