Abstract
Deep anchor-based multi-view clustering methods enhance the scalability ofneural networks by utilizing representative anchors to reduce the computationalcomplexity of large-scale clustering. Despite their scalability advantages,existing approaches often incorporate anchor structures in a heuristic ortask-agnostic manner, either through post-hoc graph construction or asauxiliary components for message passing. Such designs overlook the corestructural demands of anchor-based clustering, neglecting key optimizationprinciples. To bridge this gap, we revisit the underlying optimization problemof large-scale anchor-based multi-view clustering and unfold its iterativesolution into a novel deep network architecture, termed LargeMvC-Net. Theproposed model decomposes the anchor-based clustering process into threemodules: RepresentModule, NoiseModule, and AnchorModule, corresponding torepresentation learning, noise suppression, and anchor indicator estimation.Each module is derived by unfolding a step of the original optimizationprocedure into a dedicated network component, providing structural clarity andoptimization traceability. In addition, an unsupervised reconstruction lossaligns each view with the anchor-induced latent space, encouraging consistentclustering structures across views. Extensive experiments on severallarge-scale multi-view benchmarks show that LargeMvC-Net consistentlyoutperforms state-of-the-art methods in terms of both effectiveness andscalability.