Learning Orthogonal Latent Representations for Multi-View Clustering

In the field of multi-view clustering, latent representations are often employed to address the challenge posed by low-quality data. Traditional approaches typically assume that multiple views are fully dependent, directly learning a common latent representation from the observed data. However, this assumption is overly restrictive in real-world scenarios and may overlook valuable information, as the independence of different views can reveal critical view-specific characteristics. To overcome this limitation, we propose learning Orthogonal Latent Representations for Multi-View Clustering (OLR-MVC), which jointly captures both cross-view dependence and independence. Specifically, our model maps multi-view data into shared and private latent spaces using distinct projection bases. To accurately capture both dependence and independence, we enforce orthogonality between the shared and private latent representations while also encouraging pairwise orthogonality among private representations. Furthermore, we leverage the self-expressive property of these latent representations to capture global data structures. Extensive experimental evaluations demonstrate that OLR-MVC outperforms state-of-the-art multi-view clustering methods.