Multi-view classification via Multi-view Partially Common Feature Latent Factor Learning

In the real world, multi-view data usually consists of different representations or views. There are two key factors in multi-view data: consistency and complementarity. Unlike most multi-view learning algorithms based on nonnegative matrix factorization (NMF), the proposed method can make full use of the consistency and complementarity of data. This paper presents a new semi-supervised multi-view learning algorithm, called Multi-View Partially Common Feature Latent Factor (MVPCFLF) Learning. MVPCFLF is an extended learning form based on Partially Shared Latent Factor (PSLF) learning, which can make full use of common and special features to obtain latent representation. The key idea of MVPCFLF is to increase the constraints on common feature matrix so as to maintain the consistency of common features. The experimental results show that MVPCFLF is more effective than the existing multi-view learning algorithms.