Majorization-minimization Bregman proximal gradient algorithms for nonnegative matrix factorization with the Kullback--Leibler divergence

Nonnegative matrix factorization (NMF) is a popular method in machine learning and signal processing to decompose a given nonnegative matrix into two nonnegative matrices. In this paper, to solve NMF, we propose new algorithms, called majorization-minimization Bregman proximal gradient algorithm (MMBPG) and MMBPG with extrapolation (MMBPGe). MMBPG and MMBPGe minimize an auxiliary function majorizing the Kullback--Leibler (KL) divergence loss by the existing Bregman proximal gradient algorithms. While existing KL-based NMF methods update each variable alternately, proposed algorithms update all variables simultaneously. The proposed MMBPG and MMBPGe are equipped with a separable Bregman distance that satisfies the smooth adaptable property and that makes its subproblem solvable in closed forms. We also proved that even though these algorithms are designed to minimize an auxiliary function, MMBPG and MMBPGe monotonically decrease the objective function and a potential function, respectively. Using this fact, we show that a sequence generated by MMBPG(e) globally converges to a Karush--Kuhn--Tucker (KKT) point. In numerical experiments, we compared proposed algorithms with existing algorithms on synthetic data.