In the study of recommendation systems,many methods based on predicting ratings have been put forward.However,the rating-predicting methods have some shortages.It pays too much attention to predicting,instead of the nature of recommendation,which is predicting the order of ratings.Thus,we use a pairwise-based learning algorithm to learn our model and take the zero-sampling method to improve our model.In addition,we propose a text modeling method making the recommendations more explicable.It is proved that our system performs better than other state-of-art methods