A Majorization-Minimization Algorithm for Nonnegative Binary Matrix Factorization

This paper tackles the problem of decomposing binary data using matrix factorization. We consider family mean-parametrized Bernoulli models, a class generative models that are well suited for modeling and enables interpretability factors. factorize parameter an additional Beta prior on one factors to further improve model's expressive power. While similar have been proposed in literature, they only exploit as proxy ensure valid Bayesian setting; practice it reduces uniform or uninformative prior. Besides, estimation these has focused costly inference. In this paper, we propose simple yet very efficient majorization-minimization algorithm maximum posteriori estimation. Our approach leverages whose parameters can be tuned performance completion tasks. Experiments conducted three public datasets show our offers excellent trade-off between prediction performance, computational complexity, interpretability.