Dual Graph regularized NMF with Sinkhorn Distance

Many researchers have paid more attention to the application of non-negative matrix factorization (NMF) in data representation. Recently, some regularization methods can improve the performances by utilizing the data and feature manifold, simultaneously. In this work, a new method, named dual graph regularized NMF with Sinkhorn distance (DSDNMF) is presented. It not only synchronously takes the data structure and feature structure into consideration, but also measures the reconstruction error by adopting the Earth Mover's Distance (EMD) to make full use of the feature correlation. Therefore, DSDNMF can effectively explore the semantic structure information of data in contrast to traditional methods. Besides, we introduce an efficient strategy to optimize our proposed model. Comprehensive experiments on the COIL20 and PIE datasets manifest the superiority of DSDNMF.