Research On Adaptive Affinity Propagation Clustering Algorithm Based On Neighbor Similarity

Affinity Propagation(AP)is a new type of unsupervised clustering algorithm,which is better for regular data clustering,but there are still some problems:(1)AP algorithm as a central clustering algorithm,similarity measurement is relatively simple,for the characteristics of complex dataset identification is poor,clustering accuracy is not high;(2)The preference in the AP algorithm will directly affect the clustering effect.The selection method has no relevant theoretical guidance and needs to be set by the man,which leads to the decrease of the algorithm.For the above two points,this paper proposes an adaptive AP clustering algorithm based on neighbor similarity.The main improvement measures are as follows:(1)Aiming at the problem that the AP algorithm is difficult to identify the complex data set and the clustering accuracy is low,introducing neighborhood similarity to replace Euclidean distance as a new clustering basis.By analyzing the statistical properties of the data set,a reasonable neighborhood radius is obtained by using the nearest neighbor data set curve and the probability distribution density to calculate the neighborhood similarity,and finally build the neighborhood similarity matrix and clustering,so as to effectively improve the clustering effect of the algorithm on the complex feature data set.(2)Aiming at the problem that the AP preference need to be set manually and inadequate value has a great influence on the accuracy of algorithm clustering,by adjusting the bias preference,automatically determinethe reasonable drop step to automatically adjust the bias preference when the clustering result continues to converge,and appropriate to optimize thepace of decline,so as to improve the adaptability of the algorithm.Finally,the feasibility of the algorithm is validated in the UCI dataset and the Movie-leans dataset and applied to the user recommendation problem.The experimental results show that the clustering accuracy of the proposed algorithm is superior to the similarity algorithm.