MGCN4REC: Multi-graph Convolutional Network for Next Basket Recommendation with Instant Interest

Abstract

Sequential patterns involved in users’ historical behaviors have received extensive attention in recommendation system, which is important to represent item-level preferences. The existing works often combine the long-and short-term patterns to capture user’s preferences. But the short-term preferences modeled by the recent behavior patterns cannot clearly indicate the users’ instant interest. In this paper, we propose a sequential recommendation model MGCN4REC based on multi-graph to learn the representation of users and items and then model preferences and instant interests simultaneously. Firstly, this paper utilizes multi-graph convolutional network (MGCN) to learn users and items embeddings from multi-graph. Secondly, to aggregate preferences and instant interests, we use the attention mechanism to find the degrees of dependencies on these two features. Finally, this paper conducts experiments on real data sets of Amazon to evaluate the performance of MGCN4REC model, and the results show that our model outperforms the current state-of-the-art sequential recommendation methods over 15% on the metrics.