Recurrent memory networks: modeling long short-term user preferences for session-based recommendation

Abstract

In many online systems where recommendations are applied, interactions between a user and the system are usually organized into sessions. This paper studies how to model user preferences in session-based recommendation systems. Existing studies have either assumed that sessions are independent from each other and ignore long-term information from historical sessions, or treat the user's short-term preferences in a session as static, which cannot fully characterize user behavior in practical scenarios. Thus, we propose the recurrent memory network (RMN), which is an RNN-based framework that unifies the users' long-term and short-term preferences in session-based recommendation. The key component of the proposed RMN is preference memory, which stores a user's long-term interests. In addition, in the RMN, we design an intra-session memory reader and inter-session memory writer to facilitate explicit characterization of short-term (i.e., within a session) user preferences variation and long-term (i.e., cross-session) user preference transfer, respectively. The results obtained in extensive experiments on real-world datasets for movie and job recommendations demonstrate that the proposed RMN achieves substantial gains over state-of-the-art baselines.