Abstract: Aiming at the problems of  ignoring local structure, the high dependence of low rank representation tensor and affinity matrix, 
we proposed a multiple view clustering method based on graph regularization， low rank representation tensor and affinity matrix. Firstly, we proposed a unified framework to learn the graph regular low rank representation tensor and affinity matrix of multiple view subspace. Secondly, furthermore, the correlation of high-order cross views was analyzed by tensor singular value decomposition based on tensor kernel norm, and the local structure embedded in high-dimensional space was preserved by graph regularization. Finally, constrained quadratic programming was used to assign adaptive weights to each view. Experimental results on seven data sets show that the  clustering effect of the proposed method is better.

Key words: multiple view clustering, low rank representation tensor, graph regularization, affinity matrix