Automatic Density Peaks Clustering based on the Cosine Similarity
Aijing Feng · 2022
Density peaks clustering (DPC) is a simple and effective clustering algorithm. However, the clustering results are greatly impacted by the chain reaction due to its special allocation strategy. To solve this problem, we proposed an automatic density peaks clustering based on cosine similarity (ADPC) in this paper. First, data points are seen as vectors and each attribute value is seen as corresponded coordinate based on cosine we proposed to measure similarity, which represents similarity by the data points’ vectors and considers both the direction and the size of points’ vectors. In theory, the cosine similarity varies from 0 to 1, and the value bigger, the result will be better. Such similarity measurement is more appropriate than direct distance. Second, based on this, the improved local density, the improved similarity from the nearest point with higher density are proposed to replace the corresponded concepts of DPC. Third, to make DPC allocation strategy simpler, the improved local similarity is created. At last, a method to specify cut-off distance is introduced. We give out the strategy based on theory analysis, which makes the strategy more convinced. Above all, ADPC could select cut-off distance automatically instead of specifying it by our experience and overcome chain reaction simultaneously. ADPC shows the promotive performance on 4 synthetic datasets and 11 real datasets.