Abstract
Bayesian Non-negative Matrix Factorization (NMF) is a promising approach forunderstanding uncertainty and structure in matrix data. However, a large volumeof applied work optimizes traditional non-Bayesian NMF objectives that fail toprovide a principled understanding of the non-identifiability inherent in NMF--an issue ideally addressed by a Bayesian approach. Despite their suitability,current Bayesian NMF approaches have failed to gain popularity in an appliedsetting; they sacrifice flexibility in modeling for tractable computation, tendto get stuck in local modes, and require many thousands of samples formeaningful uncertainty estimates. We address these issues through aparticle-based variational approach to Bayesian NMF that only requires thejoint likelihood to be differentiable for tractability, uses a novelinitialization technique to identify multiple modes in the posterior, andallows domain experts to inspect a `small' set of factorizations thatfaithfully represent the posterior. We introduce and employ a class oflikelihood and prior distributions for NMF that formulate a Bayesian modelusing popular non-Bayesian NMF objectives. On several real datasets, we obtainbetter particle approximations to the Bayesian NMF posterior in less time thanbaselines and demonstrate the significant role that multimodality plays inNMF-related tasks.