关键词: 近邻传播聚类, 多聚类中心近邻传播聚类, 基于密度的聚类, 流形结构, 相似性度量

Abstract: When dealing with arbitrary shape data set with manifold structure, multi-exemplar affinity propagation cannot obtain good clustering results. To overcome this shortcoming, this paper designs a brand new similarity measure based on the idea of manifold learning. This similarity can amplify the similarity between data points of the same manifold and reduce the similarity between data points of different manifolds. As a result, the similarity matrix can reflect the internal manifold structure of the data set precisely. Based on this similarity matrix, this paper proposes the novel manifold structure based multi-exemplar affinity propagation, which can solve the problem mentioned above effectively and also improve the efficiency of this algorithm. It obtains promising results both on artificial datasets and USPS handwritten digits datasets. The simulation results show that the new method outperforms traditional MEAP algorithm.

Key words: affinity propagation, multi-exemplar affinity propagation, density-based clustering, manifold structure, similarity measure