Research On Affinity Propagation Clustering Algorithm And Its Applications

Affinity propagation(AP)clustering algorithm is a new widely used clustering algorithm, it can quickly and efficiently deal with large-scale multiple types of data, and which successfully applied to data mining in various fields by the domestic and foreign scholars. Nevertheless, AP still faces two major problems:(1) although it has great preferment on spherical or convex data, obviously when the data structures has become more complex include manifold, large-scale sparse data, reasonable clustering results cannot obtain;(2) Preferenceof AP has close correlation with number of clustering. Manual extraction of Preference is time-consuming and laborious, and Preference searching technology is imperative, which is consistent with the data structure. In this paper, here the two issues were studied in depth in this paper.In order to capture the reasonable Preference, we propose stability threshold AP(STAP) and constraint rules distributed AP(CRDAP) on the basis of stability model, establishing a reasonable framework for the Preference search technology. In view of the large sparse data, properties distribution similarity-based AP(PDS-AP) was proposed. Matching min-cluster hierarchical clustering(MMHC) algorithm for manifold data was proposed. The main contributions of this paper as followings:(1) Preference search technology of stability threshold is proposed on the basis of stability model. Simulation experimental results show that stability threshold AP has higher accuracy, faster speed than original.(2) We put forward CRDAP to construct constraint rules close to stability of clustering results. Simulation results show that the clustering algorithm based on the constraint rules can obtain the real class number of the experimental data sets.(3) We proposea new algorithm of PDS-AP based on attribute similarityin the face of sparse data. The Jaccard coefficient is introduced into the traditional algorithm to reconstruct the similarity matrix which is consistent with the sparse matrix.(4) Traditional AP algorithm cannot identify the manifold data structure,we put forward the MMHC algorithm based on minimum cluster matching using results of AP, then aggregation operation of the smallest clusters by hierarchical MMHC algorithm. Simulation experiments prove thatit can effectively identify separated manifolds.