Abstract: Affinity propagation clustering (AP) algorithm has a difficult to get ideal clustering results in a complex manifold structure and non-uniform density datasets．Through studying the low-dimensional manifold structure,this paper drawed out the density auto-adapted "manifold distance kernel " concept(ad-MDK),which takes into account the local density information of data points,also contains the overall structure of the data set globally,making the algorithm can sovle complex distributed data clustering problem．Meanwhile,in order to reduce the manifold distance calculated,parallel algorithm for the proposed algorithm was introduced to the affinity propagation clustering to effectively improve the speed of the algorithm．Experiments on several data sets verify that the proposed algorithm is superior to the traditional AP algorithm performance in dealing with large-scale multi-scale data set.

Key words: Affinity propagation,Manifold distance kernel,Shared nearest neighbor,Parallel computation