MISRec: Multi-Intention Sequential Recommendation

Learning latent user intentions from historical interaction sequences plays a critical role in sequential recommendation. A few recent works have started to recognize that in practice user interaction sequences exhibit multiple user intentions. However, they still suffer from two major limitations: (1) negligence of the dynamic evolution of individual intentions; (2) improper aggregation of multiple intentions. In this paper we propose a novel Multi-Intention Sequential Recommender (MISRec) to address these limitations. We first design a multi-intention extraction module to learn multiple intentions from user interaction sequences. Next, we propose a multi-intention evolution module, which consists of an intention-aware remapping layer and an intention-aware evolution layer. The intention-aware remapping layer incorporates position and temporal information to generate multiple intention-aware sequences, and the intention-aware evolution layer is used to learn the dynamic evolution of each intention-aware sequence. Finally, we produce next-item recommendations by identifying the most relevant intention via a multi-intention aggregation module. Extensive experimental results demonstrate that MISRec consistently outperforms a large number of state-of-the-art competitors on three public benchmark datasets.