With the diversified development of data sources, multi-view clustering has become a research hotspot. Most algorithms focus too much on using graph structure to seek consistent representation, but ignore how to learn the graph structure itself. In addition, some methods are usually optimized based on fixed views. In order to solve these problems, this paper proposed a multi-view clustering algorithm based on similarity graph projection learning(MCSGP), which effectively fused the global structure information and local potential information into a consensus graph by using the projection graph, rather than only pursuing the consistency of each view with the consensus graph. By imposing a rank constraint on the graph Laplacian matrix of the consensus graph matrix, this algorithm could naturally divide the data points into the required number of clusters. In the experiments on two artificial datasets and seven real datasets, the MCSGP algorithm shows excellent clustering effect on artificial data sets. At the same time, in the real datasets involving 21 indicators, 17 indicators reach the optimal level, which fully proves the superior performance of the proposed algorithm.

[2023-07-21] Accepted Paper /doc/2023.05.0195/v1

[2024-01-05] Printed Article /doc/2023.05.0195/v2