Optimization Of Clustering Algorithm Based On Density Peak

In recent years,clustering algorithms based on density peaks have attracted much attention because of their fast search and discovery of cluster centers,their ability to find clusters of arbitrary shapes and their ability to assign data points without iteration.However,the density peak clustering algorithm still has shortcomings and shortcomings,such as inaccurate selection of the center point when processing data sets with uneven density,and prone to associated errors in sample allocation,that is,errors in a certain data point will be followed by problems in the following data points,resulting in poor clustering effect.To solve these problems,two improved schemes are proposed in this paper,namely,density peak clustering algorithm which combines relative local density and nearest neighbor relation and density clustering algorithm based on fuzzy local density and K-nearest neighbor fusion.Both schemes can effectively avoid the defects of density peak algorithm and obtain better clustering results.In the first method,sparse and weight is introduced and relative local density is proposed to avoid errors in selecting density peaks for data sets with large sparse differences,so that appropriate density peaks can be selected to ensure the correctness of central point selection.Then,combining the criterion of nearest neighbor and the distance threshold,the nearest neighbor allocation strategy is proposed,which can effectively suppress the associated errors in allocation.Finally,a modified allocation strategy is proposed based on the intra-class distance mean to improve the accuracy of the algorithm for boundary point clustering.The proposed algorithm is compared with the clustering results of DPC,DPC-Mnd,FKNN-DPC,DBSCAN,OPTICS,AP and KMeans algorithms on the synthetic data set and UCI data set,and the Friedman test shows that the algorithm has the best performance.In the second method,the domain radius is defined to obtain the fuzzy local density of data,which can be used to replace the previous local density,so as to effectively obtain the peak density,avoid the sensitivity of density peak clustering algorithm in the selection of the center point,and improve the correctness of the selection of the center point.Then,the nearest neighbor allocation is carried out according to the nearest neighbor criterion.Finally,the obtained class clusters are fused to further correct the distribution results.The proposed algorithm is compared with the clustering results of FKNN-DPC,DPC-SA,DPC-DBFN,IDPC-FA and DPC algorithms on the synthetic and UCI datasets.On image data set,the proposed algorithm is compared with DPCDBFN,IDPC-FA,DPC-KNN and FKNN-DPC algorithms.Finally,Friedman test shows that the algorithm has the best performance.