Combining user-based and session-based recommendations with recurrent neural networks

Abstract

Recommender systems generate recommendations based on user profiles, which consist of past interactions of users with items. When user profiles are not available, session-based recommendation can be used instead to make predictions based on sequences of user clicks within short sessions. Although each approach can be used separately, it is desired to utilize both user profiles and session information, and other information such as context, when those are available. In this paper, we propose a Recurrent Neural Networks (RNNs) based method that combines different types of information to generate recommendations. Specifically, we learn user and item representations from user-item interaction data and explore a new type of RNN cells to combine global user embeddings with sequential behavior within each session to generate next item recommendations. The proposed model uses an attention mechanism to adaptively regulate the contributions of different input components based on specific situations. The model can be extended to incorporate other input, such as contextual information. Experimental results on two real-world datasets show that our method outperforms state-of-the-art baselines that use only user or session information.