Density Peaks Clustering Based on Weighted Local Density Sequence and Nearest Neighbor Assignment
Density Peaks Clustering Based on Weighted Local Density Sequence and Nearest Neighbor Assignment
复制标题
基于加权局部密度序列和最近邻分配的密度峰聚类
DOI：
10.1109/access.2019.2904254
复制
发表时间：
2019-01-01
期刊：
影响因子：
3.9
通讯作者：
Yao, Shuang
中科院分区：
文献类型：
--
作者：
Yu, Donghua;Liu, Guojun;Yao, Shuang
密度峰值聚类（DPC）是一种基于密度的聚类算法，具有良好的聚类性能，包括准确性，自动检测簇的数量，并确定中心点。然而，DPC的局部密度强烈地依赖于必须预先指定的截止距离;此外，该策略将每个剩余点分配到同一个簇中，作为其最近的较高密度的邻居，按局部密度的降序排列，这很可能导致簇标签错误传播。为了克服这些局限性，我们提出了一种改进的DPC，通过引入加权局部密度序列和两阶段分配策略，称为DPCSA。许多以前的改进DPC算法忽略了额外的复杂性，而DPCSA结合最近邻动态表，以提高聚类效率。对包括Olivetti人脸在内的12个人工数据集和11个真实数据集的实验结果表明，DPCSA的聚类性能明显上级DPC和热扩散DPC（HDDPC），略上级模糊加权k近邻密度峰值聚类（FKNNDPC）.此外，DPCSA的计算效率比FKNNDPC和HDDPC更高，但低于DPC。DPCSA的源代码可在https://www.github.com/Yu123456/DPCSA上获得。
Density peaks clustering (DPC) is a density-based clustering algorithm with excellent clustering performance including accuracy, automatically detecting the number of clusters, and identifying center points. However, the local density of DPC strongly depends on the cutoff distance which must be prespecified; in addition, the strategy assigns each remaining point to the same cluster as its nearest neighbor of higher density in descending order of local density, which is likely to cause cluster label error propagation. To overcome these limitations, we propose an improved DPC by introducing weighted local density sequence and two-stage assignment strategies, called DPCSA. Many previous improved DPC algorithms neglect additional complexity, whereas DPCSA incorporates the nearest neighbor dynamic table to enhance clustering efficiency. The experimental results for 12 artificial and 11 real-world datasets, including Olivetti face, verify that the DPCSA clustering performance is significantly superior to DPC and DPC via heat diffusion (HDDPC), and slightly superior to fuzzy weighted k-nearest neighbors density peak clustering (FKNNDPC). In addition, the DPCSA is more computationally efficient than FKNNDPC and HDDPC, but less than DPC. The source code of DPCSA is available at https://www.github.com/Yu123456/DPCSA.