Matrix completion based on Gaussian parameterized belief propagation

We develop a message-passing algorithm for noisy matrix completion problems based on factorization. The is derived by approximating message distributions of belief propagation with Gaussian that share the same first and second moments. also derive memory-friendly version proposed applying perturbation treatment commonly used in literature approximate passing. In addition, damping technique, which demonstrated to be crucial optimal performance, introduced without computational strain, relationship alternating least squares, method reported certain settings, discussed. Experiments synthetic datasets show while quantitatively exhibits almost performance under settings where earlier optimal, it advantageous when observed are corrupted non-Gaussian noise. real-world emphasize differences between two algorithms.