Keywords: nonnegative matrix factorization, Kullback-Leibler divergence, dual divergence, intrinsic information, symmetric information divergence, exponential family, EM algorithm, high-dimensional data
A unified approach to non-negative matrix factorization based on generalized dual Kullback-Leibler divergence and intrinsic information, which embeds the exponential family of models within a theoretical framework, is proposed. A family of algorithms is developed using this framework including rigorous proofs of convergence. This approach generalizes existing methods and contrasts with the recently proposed quasi-likelihood approach, thus providing a flexible alternative for non-negative matrix factorization.