Nonconvex Tensor Hypergraph Learning for Multi-view Subspace Clustering

Low-rank representation has been widely used in multi-view clustering. But the existing methods are matrix-based, which cannot well capture high-order low-rank correlation embedded in multiple views and fail to retain the local geometric structure of features resided in multiple nonlinear subspaces simultaneously. To handle this problem, we propose a nonconvex tensor hypergraph learning for multi-view subspace clustering. In this model, the hyper-Laplacian regularization is used to capture high-order global and local geometric information of all views. The nonconvex weighted tensor Schatten-p norm can better characterize the high-order correlations of multi-view data. In addition, we design an effective alternating direction algorithm to optimize this nonconvex model. Extensive experiments on five datasets prove the robustness and superiority of the proposed method.