Automatic Relevance Determination in Nonnegative Matrix Factorization

Nonnegative matrix factorization (NMF) has become a popular technique for data analysis and dimensionality reduction. However, it is often assumed that the number of latent dimensions (or components) is given. In practice, one must choose a suitable value depending on the data and/or setting. In this paper, we address this important issue by using a Bayesian approach to estimate the latent dimensionality, or equivalently, select the model order. This is achieved via automatic relevance determination (ARD), a technique that has been employed in Bayesian PCA and sparse Bayesian learning. We show via experiments on synthetic data that our technique is able to recover the correct number of components, while it is also able to recover an effective number of components from real datasets such as the MIT CBCL dataset.