Tensor-based consensus learning for incomplete multi-view clustering

As a challenging task in the field of unsupervised learning, incomplete multi-view clustering can fully utilize multi-view information in the absence of partial views. Nevertheless, most existing methods still suffer from the following two limitations: (1) The missing views are ignored in matrix factorization or self-representation learning, which might not effectively explore the complementary information of incomplete views. (2) They fail to mine the high-order inter-view relationships in consensus graph learning, which can directly affect the accuracy of the learned consensus graph. To this end, this paper proposes an effective consensus learning model, referred to as tensor-based consensus learning for incomplete multi-view clustering (TCLIMC). Specifically, in TCLIMC, the missing views of incomplete instances are reconstructed, and the reconstructed views are devoted to exploring the consensus latent representation shared by all views and self-representation of each view. Meanwhile, the consensus graph learning is enhanced by combining self-representation learning and tensor low-rank constraint. In addition, the consensus latent representation and consensus graph are correlated by graph regularization, which enhances the compactness of the consensus latent representation and the accuracy of the consensus graph. At last, the clustering results on eight incomplete multi-view datasets verify the effectiveness of TCLIMC. The code is publicly available at https://github.com/JSMMu/TCLIMC. •We propose a tensor-based incomplete multi-view clustering framework.•The missing views are reconstructed based on NMF and self-representation.•The tensor low-rank is imposed on the self-representation of all views.•We present an efficient optimization algorithm.