Enhancing link prediction through adversarial training in deep Nonnegative Matrix Factorization