Self-Refining Spherical Consensus Embedding for Constrained Multi-View Clustering

Constrained multi-view clustering aims to integrate external prior knowledge and complementary information from multiple views to enhance clustering performance. However, existing approaches typically employ Euclidean distance to learn view-specific and consensus embeddings, which often fail to capture the intrinsic geometric structure of high-dimensional data. Moreover, the dependence on limited external constraints or pre-defined anchors usually leads to suboptimal generalization and sensitivity to anchor quality. To address these limitations, we propose a novel deep constrained multi-view clustering framework, namely SeSCE. Specifically, we encode view embeddings into a spherical space, leveraging pairwise constraints to maximize intra-class compactness and inter-class separability. Crucially, a confidence-aware pseudo-constraint mining mechanism is designed to distill reliable pairwise constraints from high-confidence predictions iteratively. This effectively bridges the gap between unsupervised feature learning and discriminative clustering by progressively sharpening cluster boundaries. Finally, a globally aware attention mechanism is introduced to facilitate adaptive multi-view fusion. Extensive experiments demonstrate the superiority of our algorithm over state-of-the-art methods.