Enhancing Learning with Label Differential Privacy by Vector Approximation