Density peaks clustering algorithm integrating manifold distance and mutual nearest neighbors

•A new density metric calculated using mutual nearest neighbors is proposed to eliminate density disparities between clusters.•The improved manifold distance is proposed to identify weak connections between clusters, allowing true cluster centers to be accurately identified in the decision graph.•Di...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Pattern recognition Jg. 172; S. 112554
Hauptverfasser: Deng, Cai, Zhang, Qinghua, Zhou, Xinran, Zhang, Siyang, Wang, Guoyin, Xu, Weihua
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier Ltd 01.04.2026
Schlagworte:
Connectivity
Density peaks clustering
Mutual nearest neighbors
Proportional weights
Proportional weights
Connectivity
Mutual nearest neighbors
Density peaks clustering
ISSN:0031-3203
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:•A new density metric calculated using mutual nearest neighbors is proposed to eliminate density disparities between clusters.•The improved manifold distance is proposed to identify weak connections between clusters, allowing true cluster centers to be accurately identified in the decision graph.•Distinct weights are designed for the different types of neighbors. Density peaks clustering (DPC) is an efficient density-based clustering algorithm widely applied in various fields. However, it has several limitations, such as the inability to accurately identify the cluster centers of sparse clusters and the continuous misassignment of non-center points, which is often referred to as the “domino effect”. To address these issues, a novel density peaks clustering algorithm integrating manifold distance and mutual nearest neighbors (DPC-MDMN) is proposed. First, the harmonic density based on the mutual nearest neighbors is introduced to eliminate density differences between any two clusters. Second, a new relative distance is defined using the expanded connectivity-based manifold distance. This approach determines the connections between different clusters by analyzing the density variations along the paths of points, effectively amplifying the distance between different cluster centers and enabling more accurate identification of the cluster centers in the decision graph. Finally, a two-step assignment strategy is designed. The information of mutual nearest neighbors is used in the first step to cluster highly similar points. In the next step, different neighbor types are taken into account, with proportional weights using for the assignment. Experiments conducted on 12 synthetic datasets and 19 real-world datasets demonstrate that the DPC-MDMN algorithm performs better, particularly on the datasets with weak connectivity, where the true cluster centers can be clearly identified in the decision graph.
ISSN:0031-3203
DOI:10.1016/j.patcog.2025.112554