In practical application scenarios, the behavior of users watching movies is random and diverse, and also includes spatiotemporal features. Aiming at the fact that the complex ranking model cannot use a large amount of data for learning and updating in real time, especially the problem of insufficient training data for inactive users, this paper proposes a pre-training-based user embedding algorithm model. In the pre-training stage, the SINE model is used to dig out several intents with the highest user interest, improve the hit rate of user interest, and thus improve the accuracy of Inference. The follow-up test results show that the newly constructed recommendation model has better performance, and the evaluation index AUC is increased by 2.4% compared with the model without pre-training, which proves the effectiveness and feasibility of the new algorithm.
