An efficient initialization method for nonnegative matrix factorization

Although Non-negative Matrix Factorization (NMF) has been employed in many real applications but it still suffers from three shortcomings in terms of finding a suitable initialization method, choosing an effective cost function in addition with determining the suitable reduced dimension (Factors Rank). The aim of this study is to enhance NMF performance using Fuzzy C-Means Clustering (FCM) as an efficient initialization method for estimating initial factors of NMF. In this paper, we proposed an initialization in which both W and H matrices are identified simultaneously. The proposed method was applied to JAFFE facial expression dataset and the results exhibited superiority of this method compare to other the state-of-art initialization methods on the employed dataset. © 2011 Asian Network for Scientific Information.