Learning and Utilizing a Pool of Features in Non-negative Matrix Factorization

Learning and utilizing a pool of features for a given data is important to achieve better performance in data analysis. Since many real world data can be represented as a non-negative data matrix, Non-negative Matrix Factorization (NMF) has recently become popular to deal with data under the non-negativity constraint. However, when the number of features is increased, the constraint imposed on the features can hinder the effective utilization of the learned representation. We conduct extensive experiments to investigate the effectiveness of several state-of-the-art NMF algorithms for learning and utilizing a pool of features over document datasets. Experimental results revealed that coping with the non-orthogonality of features is crucial to achieve a stable performance for exploiting a large number of features in NMF.