Multiview clustering by joint anchor graph and label learning

Multi-view subspace clustering (MSC) has great effective in clustering performance using complementary information across the views and wide applications in various tasks, such as significance detection and information retrieval. Existing MSC methods roughly fall into two categories: self-representation based and anchor graph based methods. The later one stands out due to the high computational efficiency. It usually contains two successive steps, anchor graph learning and the following latent clustering. To implement joint optimization of the two processes and make them facilitate each other, we introduce a label guidance term into an existing model. An alternating direction minimization algorithm is carefully designed to optimize the modified model. The experimental performances on eight popular datasets and the comparison with 6 competitors show that the proposed method obtains higher performance.