When Machine Learning Meets Importance Sampling: A More Efficient Rare Event Estimation Approach