Abstract:

A clustering algorithm, named constraint projection based affinity propagation (AP), is proposed. The AP algorithm conducts clustering based on similarity matrix, outperforming many traditional clustering algorithms. However, for those datasets with complex structure, the AP algorithm cannot always achieve the ideal results. Firstly, constraints are enlarged. Secondly, the enlarged constrains are used in getting the projection matrix. At last, the clustering result is updated by the enlarged constraints in the space with lower dimension. The result shows that, compared with the comparison algorithms, the proposal is better in both time performance and clustering results.

Key words: semi-supervised;clustering;constraints;projection;affinity propagation