No Regularization is Needed: An Efficient and Effective Model for Incomplete Label Distribution Learning