Discriminative Multi-view latent Space learning based on Marginal Fisher Analysis

In recent years, samples are usually represented with different views of features. Although the features from different views represent the same sample, they are heterogeneous and the information contained in one view feature is unlikely to comprehensively describe all samples. Therefore, in multi-view learning, how to effectively integrate multiple views feature for better depicting the samples is of vital importance. In this paper, we propose a novel supervised multi-view latent space learning algorithm based on Marginal Fisher Analysis, named Discriminative Multi-view latent Space Learning DMSL. In this approach, a latent representation of multi-view feature is obtained, which encodes the information from different views feature by exploiting their complementarities. Moreover, we utilize the graphs which characterize intraclass compactness and interclass separability to improve latent representation discriminability. Thus, DMSL can integrate the information from multiple views feature to construct a compact latent representation and encodes the discriminative information into it simultaneously. Our model can be optimized efficiently by using the Augmented Lagrangian Multiplier with Alternating Direction Minimization (ALM-ADM) framework. Plenty of experiments on diverse datasets demonstrate the effectiveness of the proposed method.