| Clustering algorithms which are used to divide elements into several categories based on their similar features are an important unsupervised learning technique in data mining.Cluster analysis is widely applied on machine learning,pattern recognition,bioinformatics and image processing.In 2014,Alex Rodriguez et al.proposed a new algorithm,clustering by fast search and find of density peaks(DPC)algorithm in 《Science》.This algorithm exploits the density of data points and their distance to higher density points to discover potential clustering centers.This clustering algorithm is simple and concise,and can get the clustering result by one step,and the clustering effect is satisfactory.However,DPC needs to analysis decision graph and select the potential cluster in the clustering process artificially,which reduces the efficiency of the algorithm.In order to achieve the purpose of automatic clustering,this paper proposes a new method to select potential clustering centers based on the product of density and distance as the new judgment index Z,and use the method of probability and statistics to screen the clustering centers.Because only the potential clustering centers have higher density and larger distance,their Z values are far larger than those of non-clustering center points.Assuming the distribution of Z fits the normal distribution,we can use the method of probability and statistics to determine an upper bound.Points which are above the upper bound are considered as the clustering points automatically.The experimental results show that the probability distribution method can identify the potential clustering points correctly.The method of selecting the clustering center is similar to the method of selecting the potential cluster center of the artificial analysis decision diagram.Compared with other excellent clustering algorithms,the clustering algorithm based on normal distribution has a better performance in dealing with data sets with different shapes and can obtain better clustering results. |
