An improved affinity propagation clustering algorithm for large-scale data sets

Affinity Propagation (AP) clustering does not need to set the number of clusters, and has advantages on efficiency and accuracy, but is not suitable for large-scale data clustering. To ensure both a low time complexity and a good accuracy for the clustering method of affinity propagation on large-scale data clustering, an improved AP clustering algorithm named hierarchical affinity propagation (HAP) is proposed, which clusters data points by using AP algorithm several times on different level data. The data set to be clustered is firstly divided into several subsets, each of which can be efficiently clustered by AP algorithm. Then, the AP algorithm is performed on each subset to respectively select cluster centers of each subset. Further, AP clustering was again implemented on all the local cluster centers to select well-suited global exemplars of whole data set. Finally, to efficiently and accurately cluster data points in a large-scale, all the data points are clustered by the similarities between each data point and the global exemplars. The experimental results on real and simulated data sets show that, compared with the traditional AP and adaptive AP algorithm, the HAP algorithm can greatly reduce the clustering time consumption with a relatively better clustering results.