A min–max manifold regularization framework for multi-view clustering

Integration of manifold learning (i.e., local geometric structure learning) with clustering has been found effective for correctly identifying subgroups in the data. The manifold learning problem becomes more challenging in a multi-view data setting, as data represented across multiple views may be sampled from different underlying manifolds. Learning an accurate manifold can significantly improve the performance of multi-view data clustering. We propose a novel multi-view manifold learning algorithm using min–max operators. This ‘Min-Max Manifold’ is embedded within a proposed NMF-based framework designed to learn consensus information from multi-view data and preserve it in a low-rank representation. In contrast to deep multi-view clustering methods that enforce latent-space alignment, our method directly constructs a parameter-free consensus manifold that preserves majority-consistent geometric structure while retaining essential complementary information. Our framework further incorporates a kernel-based diversity constraint to better exploit view-specific complementary information when constructing the consensus representation and imposes an orthogonality constraint to make the learned consensus representation directly clustering-oriented. Extensive experimental results and visualizations show the effectiveness of the proposed method in finding accurate clusters and its superiority over other state-of-the-art methods.