Noise-Free Laplacian Learning for Multiple Kernel Spectral Clustering

Multi-view spectral clustering gains popularity and is successfully applied in various fields due to its superior performance. Existing approaches, however, have three drawbacks when enhancing clustering performance with multi-view data: (1) the representation of nodes in the individual view is insufficient because hidden similarity among high-order neighbors is omitted; (2) noise in each view of the data results in poor performance in constructing a uniform representation; and (3) two separate processes of embedding and discretion lead to suboptimal clustering results. To confront the challenges outlined earlier, we propose UFMKSC, a uniform framework for multiple kernel spectral clustering using a noise-free Laplacian matrix. In detail, we first construct similarity matrices with high-order information using the kernel function. Based on constructed matrices, we then calculate their Laplacian matrices and use the most informative eigenpairs that contain cluster information to reduce noise. Finally, we carefully design an efficient iterative optimization algorithm that includes spectral rotation to avoid suboptimal results. Experiments conducted on six public datasets validate the performance of the proposed method.