Linear neighborhood reconstruction constrained latent subspace discovery for incomplete multi-view clustering

• Published in: Applied intelligence (Dordrecht, Netherlands) Netherlands), 2022-01, Vol.52 (1), p.982-993
• Main Authors: Zhao, Jianguo, Lyu, Gengyu, Feng, Songhe
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial Intelligence; Clustering; Computer Science; Datasets; Experiments; Hypotheses; Learning; Machine learning; Machines; Manufacturing; Mechanical Engineering; Methods; Neighborhoods; Processes; Reconstruction; Representations; Subspaces; Surveillance
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-021-02417-z

Multi-View Clustering (MVC) is a widely used paradigm in machine learning, which can effectively explore the complementary information in multiple views to discover the internal patterns of data. Existing MVC methods generally hold a hypothesis that all samples have complete view information, while in many real scenarios, the views information may be incomplete, since collecting complete views for each instance will result in large cost in labor and time. To solve this issue, we establish a novel method named L inear N eighborhood R econstruction constrained L atent S ubspace D iscovery for incomplete multi-view clustering ( LNRLSD ), which jointly integrates latent representation learning, linear neighborhood reconstruction and spectral clustering into a unified framework. Concretely, the latent representation is learned firstly by matrix factorization to fully explore the complementarity of multiple views. Next, different from existing MVC methods that perform data reconstruction globally, we describe multi-view data in a local manner, which leads to a more clear block-diagonal structure for data distribution and makes the learned subspace representation more accurate. Finally, LNRLSD obtains the final clustering results by applying spectral clustering on the subspace representation. In summary, LNRLSD can make full use of both complementary knowledge and the local geometrical structure to improve clustering performance. Extensive experiments on five real-world datasets indicate that compared with several state-of-the-art methods, LNRLSD can achieve superior or comparable performance.