关键词: 密度峰值聚类, 混合密度, 局部结构, 相似性传递, 标签传播

Abstract: Density peak clustering（DPC） is a simple and effective clustering algorithm, but the gap between assumption and implementation, and the assumption inapplicability remains to be its defects. As a result, DPC has difficulties in determining centroids on data sets with uneven densities, acquires low robustness in non-centroid assignment strategies, and leads to chain reactions. To address this, a density peak clustering algorithm（HS-DPC） combining hybrid density and local structure is proposed in this paper. Firstly, DPC assumes that the centroids are local peaks, but the algorithm is realized as global peaks. So, using relative density and absolute density, a hybrid density calculation formula is proposed to eliminate the inconsistency. Secondly, the similarity between data points is redefined according to the local structure, thus adapting to the data with complex shape. Finally, according to the similarity transfer of the center point, the valid data is searched to build the cluster backbone structure. Referring to the distribution of backbone points of different clusters, the remaining points are assigned according to optimal distance. Thus the boundary clustering is completed without chain reaction. After the experiment on 16 data sets compared to other five clustering algorithms, the results show the effectiveness and robustness of HS-DPC.

Key words: density peak clustering, hybrid density, local structure, similarity transfer, label propagation