Large search systems are especially necessary for quick response to user queries. At the same time, strict backend delay constraints must be observed when calculating the feature relevance of candidate documents. Feature selection can improve the machine learning efficiency. Considering the characteristics that most of the initial points of fast feature selection in ranking learning start from the single feature, which has the best ranking effect, this paper first proposes an algorithm of generating initial points of fast feature selection by hierarchical clustering, and applies the algorithm to two existing fast feature selection algorithms. In addition, a new method that makes full use of clustering features is proposed to deal with feature selection. Experiments on two standard datasets show that the proposed algorithm can obtain a smaller feature subset without affecting the accuracy and obtain the best ranking accuracy on a medium subset.