Intention-Centric Learning Via Dual Attention for Sequential Recommendation

In sequential recommendation, it is critical to accurately capture the user's intention with limited session information. Previous work concentrates on modeling a single relationship existing between items in an ongoing session, e.g., sequential dependency or graph structures. They lack the exploration of constituent semantic properties of user intention and also lack effective mechanisms binding important features to deduce user intention accurately. In this paper, we present a novel intention detection-enhanced sequential recommendation model, named DASR, which can capture both sequential dependencies and user intent components. We innovatively introduce slot attention to bind the low-level local features extracted by CNN and design multiple rounds of competitive iteration mechanisms to refine the high-level representation of user intent continuously. Finally, these high-level features cooperate with the global dependencies captured by self-attention to achieve sequential recommendations. Extensive experiments are carried out on three benchmark datasets, the experimental results show that DASR outperforms the state-of-the-art baseline methods up to 6.91%, and 4.73% on Recall@20, and MRR@20, respectively.