Dimension- adaptive latent representation learning with normalized hyperbolic tensor rank for multi-view clustering

In the field of multi-view clustering, latent representation methods have attracted much attention due to their ability to extract reliable feature representations from the underlying structure of raw data. However, existing latent representation methods often rely on pre-set parameter settings for selecting the dimension of the latent space, which lacks a robust theoretical foundation. Additionally, these methods typically fail to fully leverage the high-order correlations between views and overlook the structural features within individual views. To address these limitations, this paper proposes a Dimension-adaptive Latent Representation Learning with Normalized Hyperbolic Tensor Rank (DLRL-NHTR) for multi-view clustering. Firstly, a novel latent representation framework leveraging zero-space projection is introduced, effectively eliminating the reliance on manual dimension selection in previous methods. Furthermore, we propose an improved Normalized Hyperbolic Tensor Rank (NHTR) method to better capture high-order correlations between different views. Meanwhile, we also introduce an Elastic Structural Regularization (ESR) term that effectively captures multi-level structural information within individual views, thus enhancing the overall representation capabilities. Extensive experiments were conducted on ten benchmark datasets of varying types and scales, and the comparative results with 12 state-of-the-art methods demonstrate the superiority of our method in clustering performance.