M3SPCL: Multi-stage multi-grained multi-view supervised prototypical contrastive learning.

Most existing multi-view learning methods enhance performance by exploiting view complementarity through multi-stage fusion and enforcing consistency via representation alignment. However, they are largely confined to instance-level modeling and fail to adequately capture category-level semantics, which limits their effectiveness. Recent contrastive learning attempts incorporate category information to improve discriminability, yet they often rely on exhaustive view and sample pairing, leading to prohibitive computational overhead. Consequently, achieving both category-aware effectiveness and computational efficiency remains a fundamental challenge in exploiting dual correlations at the instance and category levels. To this end, we propose M3SPCL, a Multi-stage Multi-grained Multi-view Supervised Prototypical Contrastive Learning framework. In the early stage, M3SPCL adopts a paired view selection strategy to construct contrastive inputs via efficient view concatenation. In the intermediate stage, it integrates supervised and prototypical contrastive learning to enforce fine- and coarse-grained consistency across instance and category levels, where prototype-based matching also reduces dense sample comparisons. In the late stage, an adaptive decision fusion mechanism aggregates view-specific predictions according to their learned contributions. Overall, M3SPCL provides a unified framework that simultaneously enhances effectiveness and efficiency in exploiting dual correlations. Extensive experiments on eight widely used public multi-view benchmark datasets, including image-based datasets (e.g., Handwritten, Scene15, Caltech101) and text-based datasets (e.g., Reuters), demonstrate that M3SPCL consistently outperforms state-of-the-art methods. In particular, M3SPCL achieves improvements of 1.162% in Accuracy, 1.703% in Precision, 1.298% in Recall, and 1.752% in Fscore over competitive baselines, while substantially reducing computational cost and memory usage.