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.
Users
Please
log in to take part in the discussion (add own reviews or comments).