关键词: 模式识别, 聚类分析, 密度峰值, 类簇生长, 边界分配

Abstract:

Aiming at the problems that the density peak clustering algorithm has a poor clustering effect on variable density datasets and that the "domino" phenomenon will occur in the sample assignment process, a density peak clustering algorithm combining cluster growth and boundary assignment strategy is proposed. The algorithm uses the local k-nearest neighbor information to calculate the sample density and relative distance, and then obtains the sample decision value. Based on the distance, density and neighbor relationship between samples, the attraction degree and growth radius are defined. Combined with the decision value, the cluster centers are selected in turn, and the cluster growth strategy is proposed. Starting from each cluster center, this strategy grows the current cluster by using the attraction degree and the growth radius to obtain the initial clustering result, on the basis of which the adjacency degree is defined by using the nearest neighbor and distance information between the assigned clusters and the unassigned samples, and the boundary assignment strategy is proposed. The assignment strategy divides each unassigned sample into the most appropriate cluster by the adjacency degree, and updates the assigned and unassigned sample sets continuously until all the samples are assigned to obtain the final clustering result. Compared with 7 algorithms on 16 synthetic datasets and 10 UCI datasets, experimental results show that the proposed algorithm is superior to the comparison algorithms in adjusted rand index, normalized mutual information and adjusted mutual information on most datasets. At the same time, the statistical test results show that the proposed algorithm and the comparison algorithm have statistically significant differences. The proposed algorithm has a better clustering effect.

Key words: pattern recognition, cluster analysis, density peak, cluster growth, boundary assignment