Multi-Task Multi-Behavior Sequential Recommendation

摘要

Multi-behavior sequential recommendation (MBSR) has garnered significant interest for its potential to mitigate the information overload problem. Most existing MBSR methods focus on recommending items w.r.t. one single target behavior (e.g., purchase). In fact, accurately predicting items to be interacted by other behaviors (e.g., examination and favorite) is also important for user satisfaction and especially long-term user retention. Hence, in this paper, we study an emerging problem called multi-task MBSR (MT-MBSR), which requires to recommend items w.r.t. each behavior. The main challenge of MT-MBSR is how to exploit the synergistic value of different behaviors. As a response, we propose a novel model called attention-based multi-view preference learning (AMPL). Firstly, we design an integrated preference learning (IPL) module, which learns cross-behavior preferences. Secondly, we design a sequence-independent preference learning (SIPL) module, aiming to allow the target behavior to selectively capture the contextual information. Thirdly, we introduce a shared-specific gating mechanism to adaptively fuse the outputs of the above two fine-grained learning modules (i.e., IPL and SIPL) while resolving their inherent incompatibilities, collectively leveraging the synergetic effects. Finally, we design a behavior-independent preference learning (BIPL) module to capture complex sequential dependencies among the items for improved final prediction. We conduct extensive experiments on two public datasets and demonstrate that our AMPL significantly outperforms the baseline models in recommendation accuracy w.r.t each behavior. The datasets and source code are available at https://github.com/ddEdison/AMPL.