Efficient initialization for nonnegative matrix factorization based on nonnegative independent component analysis

Abstract

In this paper, we propose an effective initialization method for nonnegative matrix factorization (NMF). NMF is a powerful unsupervised learning method that extracts meaningful nonnegative features from an observed nonnegative data matrix. Various applications using NMF have been proposed including image recognition, document clustering, and audio source separation. However, the result of such applications always depends on the initial values of the NMF variables because of the existence of local minima. To solve this problem, we propose new initialization methods based on statistical independence between NMF components, where bases and sources estimated by nonnegative independent component analysis are employed. Experimental results show that the proposed initialization method provides faster and better minimization of an NMF cost function than some conventional NMF initialization methods. Also, the proposed method provides better separation performance in audio source separation.