Abstract
This work explores non-negative matrix factorization based on regularizedPoisson models for recommender systems with implicit-feedback data. Theproperties of Poisson likelihood allow a shortcut for very fast computation andoptimization over elements with zero-value when the latent-factor matrices arenon-negative, making it a more suitable approach than squared loss for verysparse inputs such as implicit-feedback data. A simple and embarrassinglyparallel optimization approach based on proximal gradients is presented, whichin large datasets converges 2-3 orders of magnitude faster than its Bayesiancounterpart (Hierarchical Poisson Factorization) fit through variationalinference techniques, and 1 order of magnitude faster than implicit-ALS fitwith the Conjugate Gradient method.