As an essential part of big data, the recommendation system is widely used because of its practicability. Most of the traditional rating prediction algorithms mainly focus on explicit feedback, but this type of data is usually sparse in the real world. On the contrary, the Bayesian Personalized Ranking algorithm is from implicit feedback. It divides the implicit feedback into positive feedback and negative feedback. And it assumes that the unobserved item is negative feedback, the strict assumption will bring about the negative influence in recommendation applications. Motivated by this, we propose the fusion of predefined similarity and learned similarity based on Bayesian Personalized Ranking algorithm which aims to improve the recommendation. We first fuse the predefined similarity and the learned similarity method to calculate the similarity between items, and then find the items that users may be interested in through the number of item occurrence. To this end, we divide the items into three sets for each user, and provide the pairwise preference assumption. We also extend the idea of the CoFiSet model to our model, and provide a series pairwise assumption. The model parameters are learned by the stochastic gradient descent. We conduct experiments on three real-world datasets to verify the accuracy of FSBPR, and compare FSBPR with the state-of-the-art baseline models. The experimental results indicate that our model significantly improves the accuracy of the recommendation.