Learning to rank based on feature selection is an effective method in the process of data preprocessing. In this paper, the number of features and the ranking accuracy are taken as two optimization objectives, and a multi-objective algorithm based on decomposition is proposed for feature selection in learning to rank. Then, the feature subset with small number of features and high-ranking accuracy are selected. Finally, the pairwise training set is used to construct the ranking model, and experiments are conducted on the public LETOR benchmark data sets. Comparison with other algorithms, the experimental results demonstrate that the proposed algorithm can obtain more better feature subsets.