Online Nonnegative Matrix Factorization with Temporal Affinity

Online Nonnegative Matrix Factorization aims to learn the nonnegative factorizations of sequential data with temporal learning. Existing methods analyze each instance separately, and independent results are obtained accordingly. In this work, a novel approach to online learning of NMF is proposed, which is able to learn factorizations with temporal affinity of sequential data. Rather than other algorithms, the proposed method attempts to learn the ideal decomposition while reserving the affinities between temporal data. As a consequence, the durable nonnegative patterns can be preserved by the proposed method during sequential learning, and the lost of intrinsic characteristics can be alleviated. Experiments on artificial data sets demonstrate that, the proposed method is able to present fascinating performance compared with the state-of-the-art methods.