The aim of single channel source separation is to accurately recover signals from mixtures. In supervised case, non-negative matrix factorization (NMF) is a popular method to separate mixed signals from learned dictionaries. These dictionaries can be produced efficiently by sparse NMF to approximate the input signal as closely as possible. However, previous methods neither consider the structure of the data in terms of the similarity between vertices of the input signal nor use state-of-art variants of NMF that are more efficient than conventional ones. This paper presents a method that incorporate graph regularization constraint into a group sparsity NMF to improve the performance of source separation. Experimental results demonstrate that our method is outstandingly effective for speech separation in two representative scenarios.