Multi-view clustering with dual tensors

Multi-view clustering methods based on tensor have achieved favorable performance thanks to the powerful capacity of capturing the high-order correlation hidden in multi-view data. However, many existing works only pay attention to exploring the inter-view correlation (i.e., the correlation between views for a same sample) and ignore the intra-view correlation (i.e., the correlation between different samples in a view), such that the high-order information cannot be fully utilized. Toward this issue, we propose an innovative multi-view clustering method in this paper, multi-view clustering with dual tensors (MCDT), which simultaneously exploits the intra-view correlation and the inter-view correlation. Specifically, we first learn a set of specific affinity matrices by using subspace learning in each view. Then, we stack these affinity matrices into a tensor and impose the tensor nuclear norm to exploit the intra-view high-order correlation. Meanwhile, we also rotate this tensor to exploit the inter-view high-order correlation, so as to exploit more comprehensive information hidden in multiple views. Extensive experiments on benchmark datasets demonstrate that the proposed MCDT obtains superior performance in comparison with existing state-of-the-art methods.