Title: A Unified Bayesian Framework for Sparse Non-negative Matrix Factorization

Abstract: In this work, we study the sparse non-negative matrix factorization (Sparse NMF or S-NMF) problem. NMF and S-NMF are popular machine learning tools which decompose a given non-negative dataset into a dictionary and an activation matrix, where both are constrained to be non-negative. We review how common concave sparsity measures from the compressed sensing literature can be extended to the S-NMF problem. Furthermore, we show that these sparsity measures have a Bayesian interpretation and each one corresponds to a specific prior on the activations. We present a comprehensive Sparse Bayesian Learning (SBL) framework for modeling non-negative data and provide details for Type I and Type II inference procedures. We show that efficient multiplicative update rules can be employed to solve the S-NMF problem for the penalty functions discussed and present experimental results validating our assertions.