Virtual label guided multi-view non-negative matrix factorization for data clustering

Abstract

Non-negative matrix factorization (NMF) has attracted widespread attention due to its good performance and physical interpretation. However, it remains challenging when handling multi-view data for clustering. On one hand, the current multi-view NMF methods do not fully utilize the virtual label information that can be learned in the clustering process. On the other hand, they usually perform the procedures of learning latent representation and clustering individually. To solve these problems, we develop a novel multi-view clustering model, named virtual label guided multi-view non-negative matrix factorization (VLMNMF). Specifically, we learn the virtual label information of each view, which is used to guide the learning of the latent representation of data. Then, we integrate the latent representation learning and clustering process of the data into a joint framework. A multi-view graph Laplacian is further imposed on the learned low-dimensional representation, which can well preserve the local geometric structure of multi-view data. Experiments on several benchmark datasets illustrate the efficacy of the proposed method.