Fast hierarchical clustering of local density peaks via an association degree transfer method

Density Peak clustering (DPC) as a novel algorithm can fast identify density peaks. But it comes along with two drawbacks: its allocation strategy may produce some non-adjacent associations that may lead to poor clustering results and even cause the malfunction of its cluster center selection method to mistakenly identify cluster centers; it may perform poorly with its high complex O(n)(2) when comes to largescale data. Herein, a fast hierarchical clustering of local density peaks via an association degree transfer method (FHC-LDP) is proposed. To avoid DPC's drawbacks caused by non-adjacent associations, FHC-LDP only considers the association between neighbors and design an association degree transfer method to evaluate the association between points that are not neighbors. FHC-LDP can fast identify local density peaks as sub-cluster centers to generate sub-clusters automatically and evaluate the similarity between sub-clusters. Then, by analyzing the similarity of sub-cluster centers, a hierarchical structure of sub-clusters is built. FHC-LDP replaces DPC's cluster center selection method with a bottom-up hierarchical approach to ensure sub-clusters in each cluster are most similar. In FHC-LDP, only neighbor information of data is required, so by using a fast KNN algorithm, FHC-LDP can run about O(nlog(n)). Experimental results demonstrate FHC-LDP is remarkably superior to traditional clustering algorithms and other variants of DPC in recognizing cluster structure and running speed. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

FHC-LDP introduces an association degree transfer method to address the drawbacks of DPC caused by non-adjacent associations, enabling fast identification of local density peaks and automatic generation of sub-clusters. By analyzing the similarity of sub-cluster centers, a hierarchical structure of sub-clusters is built to ensure the most similarity within each cluster. FHC-LDP outperforms traditional clustering algorithms and other variants of DPC in recognizing cluster structure and running speed.