Multi-view clustering seeks to boost clustering efficacy by uncovering consistent and complementary information shared across multiple views in an unsupervised manner. Among existing approaches, graph-based methods have been widely adopted owing to their efficiency in learning graphs that capture the intrinsic cluster structures of multi-view data. However, existing methods may be limited in original feature spaces where data is nonlinearly inseparable. Furthermore, reliance on Euclidean metrics may overlook the complex, heterogeneous nature of real-world views. These factors can result in suboptimal similarity modeling, possibly compromising the overall clustering accuracy. Furthermore, many existing methods do not fully leverage the high-order correlations across multiple views. To tackle these challenges, we propose a collaborative learning framework for multi-view clustering that integrates kernelized representation metric, graph learning, and weighted tensor low-rank constraint in a unified manner. Specifically, kernel representation learning is first employed to project the original nonlinearly distributed data into high-dimensional space, where the data become more linearly separable while reducing noise and feature redundancy. Additionally, instead of relying on the simple Euclidean distance, we leverage linearity-aware metric to derive similarity measures that better capture the intrinsic relationships among data samples, thereby enhancing the quality of graph construction. Finally, we leverage a weighted tensor to recover the underlying high-order correlations embedded in the affinity matrices, while simultaneously assigning appropriate weights to each view. Results from comprehensive testing on nine multi-view benchmark datasets substantiate that the proposed framework consistently yields superior performance compared to existing state-of-the-art methods. Our code is available at https://github.com/aoli-hrbust/KMTG.
Li et al. (Thu,) studied this question.