Research On Affinity Propagation Clustering Algorithm For Probabilistic Undirected Graph Model

Affinity propagation(AP)clustering algorithm is becoming popular for its universal usage as a new kind of unsupervised clustering algorithm in recent years.Nevertheless,its application encounters the following problems.(1)The AP algorithm has a low clustering accuracy when dealing with complex data sets with noises and outlines.(2)The possibilities to make the data points of each sample to be a candidate exemplar are the same,which cannot effectively the prior knowledge to speed up clustering.(3)The preference of AP algorithm has a great impact on the clustering effect,but the selection method has no theoretical guidaince,and needs to be adjusted according to the experimental results,the adaptive clustering ability.(4)The number of clusters obtained by the traditional AP algorithm is usually more than real result,the accuracy of the results will lose meaning.Aiming at the four points above,this paper proposesd an affinity propagation clustering algorithm based on mean density optimization and probabilistic undirected graphical model.(1)According to the low complexity of AP algorhm in complex data structures,this paper introduced the combined kernel function of support vector machine,mapping data from low dimensional Euclidean space to Hilbert space,which improves the recognition ability of the algorithm to the nonlinear and feature complexity data.In order to reduce the interference of the noise information,using the Gauss filter operator to smooth the data set is helpful to improve the clustering accuracy.For the problem of automatic selection preference,the probability of using the probabilistic undirected graph model to estimate the sample data as the clustering center is proposed,and the probabilities are introduced into the preference of AP algorithm as the prior knowledge of clustering,and the clustering efficiency of the algorithm is accelerated.Then,the cluster mergling method is used to further improve the clustering accuracy.(2)In this paper,we used tlhe mean density optimization strategy to identify and mark the outlines,so that it does not participate in the iterative AP algorithm,reducing the error of the outline data iin the AP iterative process of accumulation,which accelerates the enmergence of the cluster center and improved the efficiency of clustering.After clustering,according to the shortest distance from the outlines to the center of the cluster,the classification is accomplished.Experimental results on the UCI data sets and synthetic data set show better clustering efficiency and accuracy of the proposed algorithm against several other similar algorithms.