面向视角非对齐数据的多视角聚类方法

A new challenge for multi-view learning was posed by corrupted view-correspondences. To address this issue, an effective multi-view learning method for view-unaligned data was proposed. First, to capture cross-view latent affinity in multi-view heterogenous feature spaces, representation learning was employed based on multi-view non-negative matrix factorization to embed original features into a measurable low-dimensional subspace. Second, view-alignment relationships were modeled as optimal matching of a bipartite graph, which could be generalized to multiple-views situations via the proposed concept reference view. Representation learning and data alignment were further integrated into a unified Bi-level optimization framework to mutually boost the two learning processes, effectively enhancing the ability to learn from view-unaligned data. Extensive experimental results of view-unaligned clustering on three public datasets demonstrate that the proposed method outperforms eight advanced multiview clustering methods on multiple evaluation metrics.