Multi-view multi-manifold learning with local and global structure preservation

• Published in: Applied intelligence (Dordrecht, Netherlands) Vol. 53; no. 10; pp. 12908 - 12924
• Main Authors: Feng, Wenyi, Wang, Zhe
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.05.2023; Springer Nature B.V
• Subjects: Artificial Intelligence; Computer Science; Data points; Datasets; Feature extraction; Graphical representations; Kernels; Machine learning; Machines; Manifolds (mathematics); Manufacturing; Mechanical Engineering; Processes; Teaching methods; Feature extraction; Multi-manifold learning; LPP; Multi-view learning
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-022-04101-2

Most existing multi-view learning methods adopt a single geometrical model to describe multi-class and heterogeneous data on the original feature space without considering the manifold structure contained in the dataset. This may lose some information when detecting nonlinear forms in real world datasets. Moreover, traditional kernel-based multi-view learning methods pay insufficient attention to preserving the global or local structure of training samples. To address these two problems, this paper proposes a novel method called multi-view multi-manifold learning with local and global structure preservation (MML-LGSP). First, in the feature spaces obtained by multiple empirical kernel mapping, the MML-LGSP uses nonlinear feature extraction to maintain the intrinsic low-dimensional embedding of the original data from multiple views. Second, the discriminant projection matrix is calculated with the between-class graph representing multi-manifold information and the within-class graph representing sub-manifold information. The MML-LGSP method combines multi-manifold learning and multi-view learning based on multiple empirical kernel learning into one unified learning framework. It utilizes the local and global geometrical information around the data point of each view and shows better classification performance. Extensive experiment results of various real-world multi-view datasets demonstrate the superiority our method over the state-of-the-art methods.