Clustering by DCD(low-rank doubly stochastic matrix decomposition) obtains a nonnegative low-rank doubly stochastic decomposition A=UU^T(U≥0) from the graph correlation matrix S by minimizing the criterion of KL(Kullback-Leibler) divergence: KL(A, S), and clustering from U, as the class label matrix. In the method of DCD, because the S is pre-fixed, the initial value of S has a great influence on the clustering result, which leads to its lack of stability. Aiming at this problem, this paper proposed a DCD method based on graph optimization, and integrated the optimization of graph correlation matrix S and DCD in a unified framework, which improved and extended the original DCD. The experimental results show that the graph-optimized DCD has better clustering accuracy and stability than the original DCD.

[2019-02-05] Printed Article /doc/2017.08.0874/v2