A New Method to Construct Neighbor Graph for Non-negative Matrix Factorization on Manifold

Abstract

Recently manifold method has achieved a lot of attentions in face recognition, by constructing a neighbor graph, the algorithms can preserve the geometric structure in the data. So the Graph regularized Non-negative Matrix Factorization (GNMF) which imposed manifold into NMF was achieved grate successful. Experiments show the construction of the neighbor graph is critical for GNMF especially when discriminant information is unknown. In this paper, we propose a new method to construct the neighbor graph for GNMF, by considering the 2D information of pictures, our method would better preserve the geometrical information, experiments show that GNMF with achieves better performance than traditional neighbor graph construct method in GNMF.