Scalable Incomplete Multi-View Clustering with Structure Alignment

The success of existing multi-view clustering (MVC) relies on the assumption\nthat all views are complete. However, samples are usually partially available\ndue to data corruption or sensor malfunction, which raises the research of\nincomplete multi-view clustering (IMVC). Although several anchor-based IMVC\nmethods have been proposed to process the large-scale incomplete data, they\nstill suffer from the following drawbacks: i) Most existing approaches neglect\nthe inter-view discrepancy and enforce cross-view representation to be\nconsistent, which would corrupt the representation capability of the model; ii)\nDue to the samples disparity between different views, the learned anchor might\nbe misaligned, which we referred as the Anchor-Unaligned Problem for Incomplete\ndata (AUP-ID). Such the AUP-ID would cause inaccurate graph fusion and degrades\nclustering performance. To tackle these issues, we propose a novel incomplete\nanchor graph learning framework termed Scalable Incomplete Multi-View\nClustering with Structure Alignment (SIMVC-SA). Specially, we construct the\nview-specific anchor graph to capture the complementary information from\ndifferent views. In order to solve the AUP-ID, we propose a novel structure\nalignment module to refine the cross-view anchor correspondence. Meanwhile, the\nanchor graph construction and alignment are jointly optimized in our unified\nframework to enhance clustering quality. Through anchor graph construction\ninstead of full graphs, the time and space complexity of the proposed SIMVC-SA\nis proven to be linearly correlated with the number of samples. Extensive\nexperiments on seven incomplete benchmark datasets demonstrate the\neffectiveness and efficiency of our proposed method. Our code is publicly\navailable at https://github.com/wy1019/SIMVC-SA.\n