An improved density peaks clustering based on a new local density

Abstract

Density Peaks(DP) clustering algorithm is shown to be a novel and effective clustering approach. It has been widely used in a range of applications. DP requires computing density $\rho$ and distance $\delta$ of each point. The clustering result is determined by $\rho$ and $\delta$. The $\delta$ of each point is relative to its $\rho$. The local density $\rho$ is computed by counting the number of points within a cutoff distance $d c$. Finally, the clustering result of DP depends on the value of $d c$, i.e., DP is sensitive to the cut-off distance $d c$. We found that a tiny change in $d c$ always results in different clustering results, which decreases the applicability of DP. This is because simply counting the number of points within the $d c$ range is not able to describe the local density of each point very well. To solve the above problem, in this paper, we proposed DP-NLD, an improved density peaks clustering algorithm based on a new local density measurement method. In DP-NLD, we not only count the number of points within $d c$, but also take the distribution of neighbors into account when computing the density of each point. Compared with the original DP, DP-NLD has better robustness. Our experimental results show that DP-NLD has better clustering results than DP and other clustering algorithms.