Multiview Subspace Clustering Based on Adaptive Global Affinity Graph Learning

Multiview subspace clustering is a hot topic in machine learning. Most existing methods perform clustering based on a predefined affinity graph constructed from the neighbor information of the data., which affects the performance of clustering greatly. A method is proposed to solve this problem by constructing adjacency graphs in each of the feature spaces and a common object affinity graph. A sequence of iterations is performed, at each of which the adjacency graphs and the affinity graph are refined. A restriction is also imposed on the rank of the Laplace matrix of the affinity graph, which, according to a well-known theorem, ensures the partition of the graph into several connected components, which, after the completion of the iterations, are considered the required clusters. In the numerical experiments, several test bases from open sources are used. The results are compared with the known methods, and some advantage of the proposed approach is obtained.