An Improved Probability Propagation Algorithm for Density Peak Clustering Based on Natural Nearest Neighborhood

TL;DR

This paper introduces an improved density peak clustering algorithm that leverages natural nearest neighborhoods and probability propagation, enhancing robustness and applicability to complex datasets compared to traditional methods.

Contribution

It proposes a nonparametric clustering algorithm based on natural nearest neighborhoods and probability propagation, overcoming limitations of existing density peak clustering methods.

Findings

DPC-PPNNN outperforms DPC, K-means, and DBSCAN in experiments.

The new method is more robust to parameter settings.

It effectively handles complex datasets.

Abstract

Clustering by fast search and find of density peaks (DPC) (Since, 2014) has been proven to be a promising clustering approach that efficiently discovers the centers of clusters by finding the density peaks. The accuracy of DPC depends on the cutoff distance ($d_c$), the cluster number ($k$) and the selection of the centers of clusters. Moreover, the final allocation strategy is sensitive and has poor fault tolerance. The shortcomings above make the algorithm sensitive to parameters and only applicable for some specific datasets. To overcome the limitations of DPC, this paper presents an improved probability propagation algorithm for density peak clustering based on the natural nearest neighborhood (DPC-PPNNN). By introducing the idea of natural nearest neighborhood and probability propagation, DPC-PPNNN realizes the nonparametric clustering process and makes the algorithm applicable for more complex datasets. In experiments on several datasets, DPC-PPNNN is shown to outperform DPC, K-means and DBSCAN.