Multi-exemplar affinity propagation clustering based on local density peak

Abstract

As the representatives of subclusters in multi-exemplar affinity propagation clustering (MEAP), exemplars are important in generating and merging subclusters. However, MEAP ignores the difference in data point distribution between local neighbourhoods, leading to a situation in which these selected exemplars are sometimes not the most appropriate points to represent subclusters, and the clustering performance is greatly affected. Meanwhile, local density peaks (LDPs) are found to be good representatives in local neighbourhoods. Thus, we propose the multi-exemplar affinity propagation clustering algorithm based on local density peaks (LDP-MEAP), where each exemplar should be selected from LDPs. Since MEAP must regard all points as potential exemplars (PE), we first design the MEAP with settable PE based on the assumption that the similarity values from points to PE are finite and to non-PE are negatively infinite, and we call it MEAP-PE. We then search the LDPs by point similarities to coordinate with MEAP-PE using k nearest neighbours and local density with natural neighbours. Finally, we specify that all LDPs are PE in MEAP-PE and update messages iteratively to obtain clusters. We combine MEAP-PE with LDPs, providing a new approach for AP-based clustering methods with lower computational complexity and better clustering performance. Compared with other LDP-based methods, LDP-MEAP also has better clustering performance. Comparative experiments with many prior approaches on 16 datasets demonstrate the superiority of our proposed method.