Retargeted multi-view classification via structured sparse learning

Multi-view classification is an essential issue in machine learning. Many multi-view classification methods have been proposed by fusing complementary information of multiple views. However, most of them ignore the expectation of a large margin between different classes as they adopt the fixed label matrix for classification. In addition, these methods concatenate views into long vectors to build models, which are of high complexity. To tackle these problems, in this study, we propose a novel multi-view classification model to satisfy the margin maximization criterion and reduce time complexity. In the proposed model, a retargeted constraint is introduced to increase the margin between different classes by relaxing the target matrix, and structured sparse learning is applied to remove redundant features. Moreover, an optimization algorithm is designed to divide the proposed model into several small-scale problems which could be efficiently solved with low complexity. Compared with three classical and six advanced methods, the proposed method achieves superior classification performance on 11 multi-view public datasets. Statistically, the proposed method is significantly better than the six comparison methods (SVM, simpleMKL, LDA, ROMS, LRRR, and LHFS). The visualization results on the MF and Caltech20 datasets also indicate that the proposed method certainly increases the distance between classes. (C) 2022 Elsevier B.V. All rights reserved.