In order to improve the accuracy of personalized recommendation system, a contentbased weighted granularity sequence recommendation algorithm is proposed. The item is turned to be granularities by analyzing the relationship between the item properties. The item characteristic matrix is obtained by calculating the contribution degree of each granularity. Then, according to the information of user behaviors, the user granularity sequence is generated and the granularity mapping is performed. The user preference matrix is extracted by using Apriori algorithm. Finally, the item characteristic matrix is multiplied by the user preference matrix, and the result is introduced into the improved sigmoid function to predict the preference probability, thus completing the TopN item recommendation. Experiments on MovieLens dataset show that the contentbased of weighted granularity sequence recommendation algorithm achieves an accuracy rate of 72.27%, which is higher than the accuracy rate of the current popular recommendation algorithms. In terms of efficiency, its recommended time is less than that of the collaborative filtering recommendation algorithm with the same number of users. The F1score is 0.393, fully verifying that the overall performance of the proposal is better than that of other recommendation algorithms.