期刊
INFORMATION SCIENCES
卷 576, 期 -, 页码 157-172出版社
ELSEVIER SCIENCE INC
DOI: 10.1016/j.ins.2021.06.069
关键词
Multi-view learning; Dictionary learning; Sparse representation
资金
- Natural Science Foundation of China [62076074, 61876044, 61672169]
- Guangdong Basic and Applied Basic Research Foundation [2020A1515010670, 2020A1515011501]
- Science and TechnologyPlanning Project of Guangzhou [202002030141]
Multi-view learning explores information from different perspectives, with dictionary learning being advantageous for classification, but there is limited research on combining the two. The proposed MVDL-CV method enhances multi-view classification by learning specific dictionaries and utilizing regularization between them for improved discriminative representation.
Multi-view learning can be considered as a kind of classification method which explores common and unique information among different views. For dictionary learning, it can identify informative features by learning sparse representation of samples and has great advantages for classification. However, there are few researches on the problem of multi-view learning with dictionary learning. In order to improve the performance of multi-view classification, we propose a new multi-view dictionary learning with consensus of view(MVDL-CV). First of all, we learn a particular dictionary for each view and obtain the sparse representation of the sample. Then, by utilizing the regularization term between two dictionaries in consensus, we can determine the similarity of samples and obtain the discriminative sparse representation, which can be helpful to construct the improved classifiers. Further, we obtain the solution of the model through an alternating convex opti-mization method and present the convergence analysis of MVDL-CV. In the experiments, we compare the proposed method with previous multi-view learning methods, and the experimental results show that MVDL-CV is a feasible and competitive method. (c) 2021 Elsevier Inc. All rights reserved.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据