Abstract: Aiming at the problem that the existing sequence recommendation model ignored the users’ long-term preference and short-term preference, resulting in the recommendation model not being able to  fully play its role and the recommendation effect being poor, we proposed a personalized recommendation model based on the users’ long- and short-term preferences. Firstly, for the characteristics of long and discontinuous long-term preference sequences, BERT （bidirectional encoder representations from transformers） was used to model the long-term preference, for the short-term preference sequences and the short interval time between interaction with the user, which was volatile, vertical and horizontal convolutional networks were used to model the short-term preference, after obtaining the users’ long-term preference and short-term preference, activation functions were used to model dynamically, and then a gated recurrent network was used to balance the long- and short-term preferences. Secondly, for the users’ mis-touching behavior in daily interaction, sparse attention network was used for modeling, and sparse attention network was used to process the users’ behavioral sequences before modeling the long- and short-term preferences. User feature preferences also had an impact on the recommendation results, and user features were extracted by using a multi-head attention mechanism with bias coding. Finally, the results obtained from each part were input into the fully connected layer to obtain the final output result. In order to verify the feasibility of the proposed model, experiments were conducted on Yelp and MovieLens-1M datasets, and the results show that the proposed model outperforms other baseline models.

Key words: sequential recommendation, long-term preference； short-term preference； sparse attention network； , vertical and horizontal convolutional network