Prototype-tailored cross-view alignment for partially-aligned incomplete multi-view clustering

Multi-view clustering (MVC) has been widely applied as an effective measure for unsupervised multi-source data analysis. While the mainstream MVC methods generally assume that data are fully observed and perfectly matched, real-world scenarios often diverge from such a presumption, posing a great challenge to instance correlation exploration and degrading the clustering performance. In this paper, we identify the core bottlenecks of dealing with such imperfect multi-view data: (1) how to effectively enhance the view-specific features and (2) how to conduct robust contrastive learning with flexible positive and negative sample selection, and propose a novel PrototypE-Tailored Cross-view AlignmenT (PETCAT) method with fuzzy instance-wise alignment. Specifically, a siamese-encoder-based intra-view feature enhancement module with consideration of high-order sample-wise correlations is proposed to learn representative view-specific features. Meanwhile, an optimal-transport (OT) guided prototype learning module seamlessly aligns the views without introducing extra noise caused by instance mismatching. Afterwards, a cross-view instance alignment module enhances the global geometry with similarity-based positive and negative sample selection. Contrary to current MVC methods that have fixed positive and negative samples, this strategy takes consideration of the intrinsic fuzziness of sample correspondence into consideration, and is more robust to multi-view data noise. Comprehensive experiments on both incomplete and partially-aligned data validate the effectiveness of our PETCAT method. Code is publicly avalilable at https://github.com/WenjueHE/2026-PR-PETCAT.