Self-supervised semantic graph propagation for multi-view clustering

Multi-view clustering aims to improve clustering performance by leveraging the intrinsic consistency across these diverse views. Over the years, various multi-view clustering methods have been proposed based on different learning strategies. However, most existing approaches primarily focus on aligning the consistency among shared samples across views, while overlooking the crucial role of pseudo-label graph-based self-supervision. This limitation hinders their ability to effectively uncover discriminative features inherent in the data, especially in unsupervised settings. To this end, we propose a self-supervised semantic graph propagation framework (SemGProp) for multi-view clustering. Specifically, SemGProp first employs a fusion mechanism to integrate information across views, producing a global cross-view feature representation and estimating a common high-confidence pseudo-label graph. Furthermore, we introduce a consistency graph propagation module that propagates the global pseudo-label structure to refine each view-specific similarity graph. This process encourages alignment of feature structures across views by enforcing consistency through a Kullback-Leibler divergence-based graph loss. As a result, the learned representations are regularized by reliable class-level information, ensuring that semantically similar samples exhibit consistent structures across different views. Extensive experiments verify the effectiveness of SemGProp compared to state-of-the-art competitors, demonstrating its ability to enhance the structural consistency of feature representations across views.