“L2 regularization will let the optimization become more stable/conservative for new users when they only have a few reviews for training.”