Fine-grained similarity fusion for Multi-view Spectral Clustering

Abstract

Recent advancements in multi-view clustering have attracted significant attention. However, many methods suffer from high time complexity or difficulty in tuning their parameters. Moreover, in multi-view clustering, it is very common to allocate weights to different views for ensuring adequate utilization of information from multi-view data. Most methods allocate the same weight to every view, whereas some methods attempt to learn the optimal weight of each view. Since multi-view clustering can be deemed as a task of fusion, we propose a novel method, Fine-grained sImilariTy fuSion for Multi-view Spectral Clustering (FITS-MSC), which can address the problem that exists when assigning the same weight to instances in one view (coarse-grained information fusion): some samples may be corrupted or missing in partial views whereas others remain intact in all views. To obtain promising results, we employ sparse subspace clustering for constructing the initial similarity matrices. Additionally, to address the deficiency of coarse-grained information fusion, we design a fine-grained similarity fusion strategy for obtaining the final consensus affinity matrix. In the fusion process, the local inter-view and global intra-view weight relationships are explored. With only one parameter, FITS-MSC is very practical. Experiments on real-world datasets demonstrate the advantages of our method.