PUB - Publikationen an der Universität Bielefeld

Currently, the majority of multi-view spectral clustering approaches fuse complementary information by performing consistent spectral embedding learning based on pre-constructed similarity graphs and achieve clustering results using a single-view clustering method. Therefore, the whole clustering strategy involves three independent steps, including similarity matrix learning, relaxing a binary cluster indicator matrix (BCIM) into a continuous spectral embedding matrix (CSEM), and discretizing the CSEM back into a BCIM. Consequently, the final clustering results are not optimal since these three independent learning processes only achieve their optimality without considering the coupling between them. To tackle this concern, this study presents an innovative discrete strategy for spectral clustering, leading to the development of a joint learning model referred to as Binary Multi-view Spectral Clustering (BMSC). The suggested BMSC effectively integrates BCIM, CSEM, and consistent graph learning. Precisely, the learning of the consistent similarity graph is dynamically adjusted according to the BCIM feedback to explore complementary information in different views. In turn, the dynamic consistent similarity graph will also facilitate learning the BCIM with view-specific CSEM and obtain the joint optimal clustering results. The suggested BMSC has been shown to be superior to other state-of-the-art methods via experimental findings conducted on several benchmark multi-view datasets.