Multi-view Clustering via Deep Graph Eembedding

Due to the ability to analyze multi-view data with complex structure and sparse value, multi-view clustering has attracted increasing attention in recent years, among which graph-based methods have achieved impressive success. However, most of the current advanced methods obtain the clustering assignment on the unified similarity graph but ignored the latent information contained in the original data, which makes the clustering performance rely heavily on the quality of unified similarity graph. To overcome above drawback, we propose a multi-view deep clustering method based on Graph Convolutional Networks (GCN) called DGEMvC, which can mine deep embedding of the original data and optimize iterative clustering objectives. Firstly, for general type multi-view data, the necessary graph structure required by GCN is completed by the proposed cross-view similarity graph construction strategy. Secondly, the constructed similarity graph assists in mining the deep embedding of original data as auxiliary information. Significantly, we further design a novel strategy to generate a pseudo-label for each data point to realize the unsupervised learning pattern. Dozens experiments on six benchmark multi-view datasets verify that our proposed method is superior to other advanced multi-view algorithms and improves by 7% and 4% in ACC and NMI.