This paper presents an initialization method for Non-negative Matrix Factorization (NMF) and its applications. First, we propose an initialization method that sets the absolute value of elements of the orthonormal matrices obtained by Singular Value Decomposition (SVD) to the initial values of matrices in NMF. It is verified by an experiment that the desirable matrices for NMF are obtained by using our initialization method instead of a random initialization. Second, we propose the applications of NMF using our initialization method such as training data compression and similarity search based on vector space model. With our training data compression method, we can obtain a high-quality compressed training data by utilizing the property of noise reduction in NMF. In similarity search based on NMF, the retrieval accuracy depends on the initial values of matrices, the matrix rank and the number of iterations, so we show how to determine these parameters to improve the retrieval accuracy. It is verified by experiments that our initialization method is effective in these applications.
展开▼