An Enhanced Initialization Method For Non-Negative Matrix Factorization
Liyun Gong, Asoke K. Nandi

Abstract:
Non-negative matrix factorization (NMF) is a dimensionality reduction tool, and has been applied to many areas such as bioinformatics, face image classification, etc. However, it often converges to some local optima because of its random initial NMF factors ( and matrices). To solve this problem, some researchers have paid much attention to the NMF initialization problem. In this paper, we first apply the k-means clustering to initialize the factor W, and then we calculate the initial factor H using four different initialization methods (three standard and one new). The experiments were carried out on the eight real datasets and the results showed that the proposed method (EIn-NMF) achieved less error and faster convergence compared with both random initialization based NMF and the three standard methods for k-means based NMF.