Based On Affine Propagation Clustering Algorithm To Improve Research

With the rapid development of information science, more and more data information has appeared in all areas of people’s life. How to better solve data information retrieval problem which with large-scale and high dimensional, not only can division objects according to the similarity between samples, and can quickly cluster, in order to solve this problem, some scholars put forward affinity propagation algorithm. There are some literatures have proved that, the standard affinity propagation algorithm is slightly better than classical k-means algorithm not only in the cluster scale, and in clustering time. The traditional affinity propagation algorithm most used Euclidean distance calculations similarity between the sample points, in which all features are treated as equally. Use which distance measurement can better response characteristics of the sample space, this article is to solve this problem, used several different distance function to calculation similarities, then compared the cluster results, and finally do some relevant improvement on how to improve the speed of the clustering algorithm.In this thesis, I described the standard affinity propagation algorithm, and the basic principles of the iterative algorithm, and then I analyzed several important parameters. Most of the current improved affinity propagation algorithm calculation similarities between the sample points using Euclidean distance function, so sometimes it can’t reflect the real features of the space of samples. In allusion to deal with different space features, This thesis suggests three distance functions, such as Euclidean distance、Manhattan distance and feature distance, then severally use them to calculation similarities between the sample points of different dimensions data sets, and the gray matrix of images, then use the results to cluster. The results suggest that, for different space features, we should use different distance function to calculate the similarity matrix, so as to get a better quality cluster result.The traditional affinity propagation algorithm uses sequential search method to find a preference which corresponding to the highest clustering quality, this method would spend a long time, especially at the worst case, it will search the whole space of preference. In the pursuit of clustering quality, in order to improve the speed of the clustering, chapter4of this paper improved algorithm of scanning mode of preference, we use binary scanning instead of order scanning. The experiments results show that the improved algorithm can effectively reduce the time of searching optimal clustering results.