Multi-view subspace tensorization with attentive clustering embedding

Deep learning-based multi-view clustering (DMVC) has garnered significant attention and achieved remarkable success, primarily due to its powerful nonlinear representation capabilities. However, most existing DMVC methods fall short in effectively modeling the inherent high-order correlations among different views. In contrast, traditional MVC techniques have long recognized the effectiveness of tensor-based approaches in capturing such complex interactions. To bridge this gap, we propose STANCE (multi-view Subspace Tensorization with Attentive Clustering Embedding), a novel DMVC framework that integrates tensor modeling into a deep learning paradigm. Specifically, STANCE first employs view-specific auto-encoders to project raw data into latent feature spaces, enabling robust subspace representations. These representations are then stacked to construct a third-order tensor, upon which a Tensor-based low-rank constraint is imposed to explicitly capture high-order inter-view dependencies. To further enhance consistency across views, STANCE incorporates an attention-based adaptive fusion module that dynamically assigns weights to view-specific representations at the sample level, thereby extracting consistent information from features that are both similar and confident. Comprehensive experiments, encompassing parameter analysis, ablation studies, and visual comparisons, are conducted to thoroughly evaluate the effectiveness of STANCE. The results demonstrate that STANCE significantly outperforms state-of-the-art methods on various benchmark datasets.