Multi-View Subspace Clustering by Combining ℓ2,p-Norm and Multi-Rank Minimization of Tensors

Abstract

In this article, based on the self-represented multi-view subspace clustering framework, we propose a new clustering model. Based on the assumption that different features can be linearly represented by data mapped to different subspaces, multiview subspace learning methods take advantage of the complementary and consensus informations between various kind of views of the data can boost the clustering performance. We search for the tensor with the lowest rank and then extract the frontal slice of it to establish a well-structured affinity matrix. Based on the tensor singular value decomposition (t-SVD), our low-rank constraint can be achieved. We impose the ℓ2,p-norm to flexibly control the sparsity of the error matrix, making it more robust to noise, which will enhance the robustness of our clustering model. With combining ℓ2,p-norm and tensor multi-rank minimization, the proposed Multi-view Subspace Clustering(MVSC) model can effectively perform clustering with multiple data resources. We test our model on one real-world spoon dataset and several publicly availabe datasets. Extensive evaluation methods have proved that our model is effective and efficient.