Abstract: Differential privacy can effectively solve the privacy leakage problem of recommended systems, but introduced noise will reduce the performance of recommended systems. In addition, different users have different sensitivities to privacy protection. So, considering individual needs of users, an algorithm can be designed to reduce the added noise and help improve the performance of the recommended system. In this paper, combining the above two dimensions, a personalized differential privacy-preserving collaborative filtering algorithm was proposed. Firstly, dividing the privacy sensitivity of the ratings locally, the algorithm was designed to use a random flip mechanism to protect the privacy of the privacy-sensitive scores. Then, according to the obtained perturbed data, the algorithm was arranged to use a Bayesian estimation method to reconstruct the joint distribution between items, so as to improve the recommendation accuracy of the algorithm. The theoretical and experimental results show that the algorithm can not only provide better recommendation performance but also protect privacy of users.