Because of the growing number of users and the rapid expansion of information, sparse problem of data matrix in traditional collaborative filtering becomes more seriously. We proposed a new hybrid recommendation system. Firstly, we consider tag information and rating information as constraints under maximum entropy model. Secondly, we define the features of tag information and rating information and calculate the corresponding weights. Finally, we use previously weights to predict probability distribution of target item for current user, then we choose the highest probability as predicted rating. Experiment results show that the proposed method can effectively improve the accuracy of recommendation systems.