Multi-behavior Attention Mechanisms Graph Neural Networks based on Session Recommendation

Under current behavior, session-based recommendation attempts to predict the next item our users will purchase. Previous research approaches on session-based recommendation are based on users' target behaviors, however, they ignore the problem that accurate session vector identification cannot be obtained from the limited user target behaviors. For this purpose, this paper proposes a multi-behavioral attention mechanism model based on graph neural networks, which can perform a deeper representation of multi-behavioral interaction sequences for session-based recommendations, and at the same time can make more accurate predictions for long sessions. Specifically, using the target and auxiliary behavior sequences of the user, we construct a multi-relational item graph. Secondly our model uses GNN to extract item-to-item relationships. After that each behavior sequence learns the dependencies by applying a spatial attention mechanism to fuse the target and auxiliary behavior sequences, and finally to predict the next item interacting with the target behavior. Through experiments on two publicly available datasets, it is demonstrated that our model performs well on HR and MRR compared to other representative recommendation models.