Research On Adaptive Density Peaks Clustering Algorithm Based On Natural Neighbor

Density peaks clustering algorithm is a new clustering algorithm which can quickly cluster by finding cluster centers in decision graph.The algorithm assumes that the cluster center of each cluster has the maximum density and the cluster centers of any cluster are far away from each other.Through calculating density and distance of each data point,the decision graph is generated.Selecting the cluster centers in this graph according to its characteristics,and then completing the nearest distribution of the remaining points.This algorithm stands out with its advantages such as few parameters,being able to deal with arbitrary shape clusters,being simple and fast.However,there are still some problems in this algorithm,such as the sensitivity of the result to unique parameters,the single density formula,the need for human intervention to select cluster centers,and high complexity of the algorithm.Based on the analysis and summary of different clustering algorithms,an improved algorithm is proposed in this thesis for the above shortcomings.The main research results are as follows:(1)The performance of the density peaks clustering algorithm is sensitive to the choice of the cut-off distance(9(9₍(8(8).The selection of(9(9₍(8(8) is based on subjective experience.Therefore,this thesis introduces the natural neighborhood algorithm,which can adaptively get the local neighborhood of each data point,and does not need artificial input parameters,so it can better solve the problem of parameter sensitivity;at the same time,according to the neighborhood of the data points,further considering the distribution of the surrounding data points,redefining the local density formula to extract the core points,so that the density formula can be applied to more complex datasets.(2)Because the density peaks clustering algorithm needs to select the cluster center artificially,when there are multiple density peak points in the cluster,it is very difficult to identify the center from the decision graph.To solve this problem,this thesis introduces the idea of biological colony growth into the density clustering algorithm,through simulating the strategy of bacterial reproduction to find the maximum density point in the remaining core point set each time then expanding it first and finding the complete structure of each sub cluster adaptively.And then according to the definition of boundary point set of nearest neighbor pairs proposed in this thesis,the sub-cluster that may belong to the same cluster is merged.The whole process does not need human intervention,but also can avoid the influence of pseudo cluster center on clustering results through automatic clustering.(3)This thesis proposed an adaptive density peak clustering algorithm based on natural neighbor Na N-ADPC.Na N-ADPC algorithm uses the natural neighbor algorithm to calculate the density of points,extracts the core point set according to the density threshold,clusters the core point by priority expansion to find the complete cluster structure,and finally allocates the remaining points according to the natural neighbors.The algorithm does not need to set parameter artificially,and can deal with complex datasets well.In addition,the algorithm avoids large-scale distance calculation and reduces the complexity of the algorithm.(4)Through a large number of experiments on different types of twelve artificial data sets and eight real data sets,it is verified that the clustering results of the Na N-ADPC algorithm proposed in this thesis are better than other six compared algorithms.Finally,the Na N-ADPC algorithm is applied to face recognition,and the results show that the algorithm can recognize facial images of multiple people in different angles.