A Unified Spectral Rotation Framework Using a Fused Similarity Graph

Multi-view spectral clustering has recently received a lot of attention. Existing methods, however, have two problems to be addressed: 1) similarity matrices used in clustering omit the high-order neighbor information, reducing embedding accuracy; 2) two independent procedures of embedding and discretization may result in a suboptimal result, lowering the final performance. To address the abovementioned issues, we propose a unified spectral rotation framework for multi-view clustering using a fused similarity graph. The method begins with establishing similarity graphs for each view and constructing first-order and high-order Laplacian matrices for capturing the hidden similarity among different nodes. Then embedding and discretization procedures are integrated into a new framework for performing a spectral rotation to obtain a global clustering result. Finally, a three-step optimization method for obtaining the final clustering labels is proposed. We conduct extensive experiments on a variety of real-world and synthetic datasets to validate the effectiveness of the proposed algorithm. Our method outperforms state-of-the-art methods by 8.0% on average, according to experimental results. The code of the proposed method is available at https://github.com/lting0120/USRF_FSG.git .