Tensorized latent representation with automatic dimensionality selection for multi-view clustering

Abstract

Latent representation has garnered significant attention in the field of multi-view learning due to its ability to capture the underlying structures of raw data and achieve promising results. However, latent representation-based methods often encounter challenges in selecting the dimensionality of the latent view, which limits their applicability. To address this problem, we propose a novel method called Tensorized Latent Representation with Automatic Dimensionality Selection (TLRADS), which can automatically determine the optimal dimensions. In TLRADS, we leverage the cumulative contribution rate of singular values to determine the number of dimensions for each view-specific latent representation. This approach ensures that the chosen dimensions capture a significant portion of the data’s variability while discarding less relevant information. After obtaining the latent representation views, we incorporate the tensor subspace learning technique to capture the underlying structural information more comprehensively. Finally, an efficient iterative algorithm is designed to solve the TLRADS model. Through experimental validation, we demonstrate the effectiveness of the automatic dimensionality selection strategy in TLRADS. Meanwhile, the experimental results on real-life datasets indicate that TLRADS outperforms state-of-the-art multi-view clustering methods.