Joint Original Space and Latent Space for Multi-view Spectral Clustering

We propose a novel multi-view spectral clustering model, called Joint Original space and Latent space for Multi-view clustering (JOLM). Different from most existing multi-view clustering methods, which usually improve clustering performance by developing original or latent features of multi-view data, the proposed JOLM method integrates both original features and latent features into a framework to improve clustering performance. Specifically, we learn the similarity graph matrix from original multiple features and latent features respectively, and obtain the global graph by minimizing the errors between them, so as to better utilize the rich information from multiple views. An effective iterative algorithm is proposed to optimize the objective function. Finally, abundant experiments show the effectiveness of our proposed method.