Integrating user short-term intentions and long-term preferences in heterogeneous hypergraph networks for sequential recommendation

Abstract

Sequential recommendation tries to model the binary correlations among users and items in a sequence to provide accurate recommendations. However, user behaviors are influenced by both their intentions and preferences. Existing sequential recommendation models cannot effectively capture the user’s real preferences and intentions just from the interaction data. To tackle this dilemma, we propose IPSRec, an innovative sequential recommendation approach that explicitly mines user intentions and preferences by fusing heterogeneous auxiliary information from multiple sources. IPSRec employs a heterogeneous hypergraph network to fuse and propagate the topological information of users’ short-term intentions and long-term preferences. Subsequently, IPSRec engages a locally reinforced attention structure to learn the migration process of users’ short-term intentions and long-term preferences in sequences. Lastly, we combine the user intentions and preferences for the final recommendation. Experiments conducted on various datasets present the preeminence of our approach over several state-of-the-art methods. Moreover, the results corroborate the validity of the feature modeling for user intentions and preferences in accurately representing users’ purchase behaviors.