Multi-view Spectral Clustering With Adaptive Local Neighbors

Some existing multi-view spectral clustering methods calculate a fixed similarity matrix. However, the fixed similarity is hard to reflect comprehensive and correct relationship between data points and may not be optimal for learning sample embedding, leading to the suboptimal result. In this paper, we propose a novel multi-view spectral clustering (MSCAN) algorithm to address these issues. Specifically, the proposed algorithm generates the dynamic similarity matrix with adaptive neighbors for each view instead of using the fixed similarity matrix. It enhances the representative capacity of the similarity matrix and accuracy of the sample embedding. Meanwhile, the knowlege trnasfer is imposed to the sample embedding for all view, such that the sample embedding learns shared information betwwen the feature embeddings to keep consistency. Additionally, the feature embedding of each view learns from the shared sample embedding to obtian complementary information. An alternating iterative optimization algorithm is presented to optimize the objective function. Experimental results on several benchmark databases show the superior performance compared to the previous state-of-the-arts.