Optimization Research Of Density Peaks Clustering Algorithm Based On Neighbor Searching

The density peaks clustering algorithm can achieve effective assignment of non-centers by only one iteration after manually determining the clustering centers.The algorithm is simple in idea and has a strong practicality,but we still need to address some urgent problems,such as the calculation of local density and minimum-distance based on global parameter,manual determination of clustering centers,and low clustering accuracy when dealing with uneven density datasets and complex manifold structure datasets.In this paper,we propose corresponding improvement strategies for the problems of density peaks clustering algorithm.To address the problems of parameter sensitivity,manual determination of clustering centers and low accuracy when dealing with uneven density datasets,we put forward a new DPC algorithm based on natural-neighbor.Firstly,the pr value is introduced to take place of sample points' local density;Secondly,the clustering centers are identified according to the characteristics of the clustering centers and combined with the concept of natural neighbors;Then the natural neighbor diffusion method is applied to initially cluster the low relative-distance points to form the clustering skeleton,and the high relative-distance points are filled into the clustering skeleton to which most of their natural neighbors belong;Finally,the remaining points are classified into the same cluster with their nearest neighbors.The comparative experimental result shows that the algorithm can get best clustering results on several datasets and has a greater advantage in handling uneven density datasets.To address the problems of parameter sensitivity,manual determination of clustering centers and chain misclassification of sample points when dealing with complex manifold structure datasets,we put forward a new DPC algorithm based on KNN.Firstly,The algorithm defines new formula for calculating the local density as well as the minimum-distance based on KNN;Secondly,the clustering centers are determined by detecting the gap between two points that are adjacent in the decision graph;Then,the non-centers are divided into internal points,boundary points and outlier,and the internal points are assigned to the clusters to which their density-connected center belong;Finally,the boundary points are assigned to the clusters with the highest clustering membership.The comparison experimental result shows that the algorithm has highest accuracy compared to other algorithms,and more greater advantage in dealing with complex manifold structure datasets.