Long and Short-Term User Intent Learning for Sequence Recommendation

Personalized recommendation algorithms accurately mine users’ potential intentions by analyzing historical interaction data, deepening the user-item relationship and screening to filter complex information while effectively improving users’ experience and engagement. However, it is difficult to distinguish between embedded features and interaction relationships, and there is still the problem of lack of timeliness in intent representation. Aiming at the above problems, this paper proposes a sequence recommendation model that integrates users’ long and short-term intentions, which models users’ long and short-term intentions as potential factors over time at the fine-grained intention level of interaction, which accurately locates the users’ real intentions at a certain time. Firstly, a dual-layer gated recurrent network is designed to solve the sequential dependency problem in long-term sequences and to capture the changes in the user’s long-term intention through an intent evolution unit based on the attention mechanism. Secondly, an online meta-learning update strategy is optimized for real-time modeling of short-term intentions to take into account current temporal information and content relevance. Finally, the two types of intentions are adaptively fused to predict the interaction probability at the next time. Extensive experiments on three real-world datasets show that the LSI-Rec model significantly outperforms all baseline models on the recommendation task.