DSIM: dynamic and static interest mining for sequential recommendation

Sequential recommendation aims to predict the next interaction by mining users’ evolving interest from their historical behaviors. Through comprehensive study, we argue that the evolving interest consists of dynamic interest and static interest, both of which are important for recommendation tasks. However, it is still a challenging task to model the dynamic interest accurately since it may change rapidly. Besides, existing approaches cannot fully exploit users’ static interest. In this paper, we propose a novel dynamic and static interest mining approach called DSIM for sequential recommendation. Specifically, DSIM takes advantage of the time information and learn the dynamic interest precisely with a time-aware neural Hawkes process. Furthermore, a side information-aware self-attention mechanism is designed to leverage side information in a non-invasive way to learn static interest. Finally, DSIM fuses the learned dynamic and static interest adaptively with the gating mechanism and generates the hybrid interest for sequential recommendation. Extensive experiments on three real-world datasets demonstrate that DSIM can effectively mine and model users’ interest, and it outperforms the state-of-the-art baselines in sequential recommendation tasks.