Tensorized Multi-view Clustering via Hyper-graph Regularization

Abstract

Multi-view clustering intent to separate data into different groups regarding their multiple traits. Existing tensor multi-view clustering techniques can explore the high-order associations of multi-perspective characteristics. However, they suffer from the following issues: (1) data features and local geometric structures in nonlinear subspace are often ignored; (2) the prior knowledge of singular values in the tensor kernel norm is not well utilized. To settle these problems, we propose a novel Markov chain tensor-based approach named Tensorized Multi-view Clustering via Hyper-graph Regularization(TMC-HR). Firstly, the third-tensor based on Markov chain transition probability is constructed and rotated to reduce the model complexity. Secondly, hyper-graph regularization is employed to save the high-order local geometrical structure imbedded in the original space. Thirdly, the weighted strategy is applied to the tensor composed of latent representations to extract the high-order relationships and diverse information between different views. Finally, an effective iterative method is utilized to solve the proposed TMC-HR. We conducted extensive experiments on benchmark datasets corresponding to different types to indicate that TMC-HR performs superior over other multi-view clustering approaches.