Multi-view Subspace Clustering with View Correlations via low-rank tensor learning

With the purpose of boosting clustering performance, the manner of excavating underlying view correlations is an important issue of multi-view subspace clustering (MSC). Nevertheless, regarding most MSC approaches are centered on the view correlations in multiple subspace representations and neglect the view correlations in multiple feature representations. To address this limitation, this paper introduces a method, dubbed Multi-view Subspace Clustering with View Correlations (MSCVC), which excavates underlying view correlations in both multiple feature representations and multiple subspace representations simultaneously. Specifically, the Multi-view Principal Component Analysis (MPCA) is designed to capture the view correlations contained in multiple feature representations by using the orthogonal mapping and low-rank tensor constraint. As a consequence, view correlations in multiple feature representations can be effectively investigated and simultaneously embedded in the learned new feature representations. For view correlations in multiple subspace representations, the new feature representations learned in MPCA are employed and the low-rank tensor constraint is used as well. Experiments are conducted on nine benchmark datasets to demonstrate the progressiveness of our MSCVC compared to several state-of-the-art algorithms.