Dual-graph regularized concept factorization for multi-view clustering

Abstract

Matrix factorization is an important technology that obtains the latent representation of data by mining the potential structure of data. As two popular matrix factorization techniques, concept factorization (CF) and non-negative matrix factorization (NMF) have achieved excellent results in multi-view clustering tasks. Compared with multi-view NMF, multi-view CF not only removes the non-negative constraint but also utilizes the idea of the kernel to learn the latent representation of data. However, both of them ignore the local geometric structure in the nonlinear low-dimensional manifold. Furthermore, most of the existing CF-based methods are designed for single-view tasks, which cannot be directly applied to multi-view clustering tasks. To tackle the above shortcomings, we present a new multi-view clustering algorithm, called dual-graph regularized concept factorization for multi-view clustering (MVDGCF). Specifically, we first extend conventional single-view CF to a multi-view version, which can explore the complementary information of multi-view data more effectively. Then we develop a novel dual-graph regularization strategy, which can simultaneously capture the local structure information of the data space and feature space, respectively. Moreover, an adaptive weight vector is introduced to balance the importance of different views. Finally, extensive experiments are carried out on seven datasets. The results show that our method is superior to several popular multi-view clustering methods.