Modeling Multi-Grained User Interests for Sequential Recommendation

Abstract

Sequential recommendation has garnered considerable attention in recent years, which aims to infer the next desired item for each user based on their historical behavior data. Despite the remarkable progress made by previous methods, we believe that they still have two limitations: 1) they mainly focus on extracting pairwise transition pattern between two items within each individual sequence, how to fully capture the collective influence of a group of items is less explored; and 2) they pay insufficient attention to the intricate dependencies among items across different user behavior sequences, which may lead to suboptimal performance. In response to this, we propose a novel solution for sequential recommendation, termed the multi-grained user interest learning model (MGUI), which captures dynamic user interests by modeling item transition patterns at different levels. Specifically, our designed time interval-aware gating module aims to highlight the individual influence of each recent item on the target behavior, while the devised capsule-guided transformer module aims to capture the collective influence of a group of recent items on the next-to-visit item. These two modules explicitly exploit item transition patterns within a sequence at the individual- and group-level. Furthermore, we reorganize all users’ behavior sequences into a personalized item transition graph and devise a graph transition propagation module upon this graph to capture global-level dependencies among items across different sequences. Finally, we design a user-specific mixture-of-experts strategy to dynamically determine the importance of each module for each user. Extensive experiments across four recommendation scenarios demonstrate the superiority of MGUI over recent state-of-the-art recommendation models. In-depth analyses verify the effectiveness of each part in MGUI.