Density propagation based adaptive multi-density clustering algorithm.

Yizhang Wang,You Zhou,Wei Pang,Xiangtao Li

doi:10.1371/journal.pone.0198948

Abstract

The performance of density based clustering algorithms may be greatly influenced by the chosen parameter values, and achieving optimal or near optimal results very much depends on empirical knowledge obtained from previous experiments. To address this limitation, we propose a novel density based clustering algorithm called the Density Propagation based Adaptive Multi-density clustering (DPAM) algorithm. DPAM can adaptively cluster spatial data. In order to avoid manual intervention when choosing parameters of density clustering and still achieve high performance, DPAM performs clustering in three stages: (1) generate the micro-clusters graph, (2) density propagation with redefinition of between-class margin and intra-class cohesion, and (3) calculate regional density. Experimental results demonstrated that DPAM could achieve better performance than several state-of-the-art density clustering algorithms in most of the tested cases, the ability of no parameters needing to be adjusted enables the proposed algorithm to achieve promising performance.

Highlights

Clustering has been a promising technique in data mining and pattern recognition [1,2,3]
Manual intervention and domain knowledge is still required to obtain an appropriate threshold in density peaks (DP), and a user may require a significant amount of time to learn how to configure parameters properly [9]
Number of classes 2 3 2 2 2 2 3 2 7 2 3 correct clustering results for low dimensional data, (2) performance comparison with other density clustering algorithms, and (3) whether DPAM is robust on high-dimensional data

Summary

Introduction

Clustering has been a promising technique in data mining and pattern recognition [1,2,3]. In addition to the above algorithms, recently Rodriguez and Laio proposed a novel fast density clustering algorithm by searching density peaks (DP) [8], and they adopted a global threshold to calculate the local density of each point. Based on the above consideration, in this research, we propose an adaptive density clustering method to address the above issue. The chosen parameter values may be invalid when clustering new datasets with different characteristics. This necessitates an adaptive approach which can automatically achieve density clustering. The DPAM algorithm is proposed to automatically extract potential spatial data structure without manual adjustment of parameter values.

Related work

Experimental results and discussion

Conclusions and future work