MAGNN-GC: Multi-head Attentive Graph Neural Networks with Global Context for Session-Based Recommendation

Session-based recommendation aims to predict the final preference of anonymous users based on their current session and global context. However, integrating high-dimensional information such as the item features of the current session and global context can lead to insufficient information fusion and imbalanced positive and negative samples during model training. To address these issues, we propose Multi-Head Attentive Graph Neural Networks with Global Context for Session-based Recommendation (MAGNN-GC). We construct global and local session graphs based on all historical session sequences to fully represent the complex transition relationships between items in the session. We use graph convolutional networks to capture the item features in the global context and graph attention networks to capture the item features of the current session. We also integrate the position information of session items with the learned global-level and local-level item embeddings using the multi-head attention mechanism. Additionally, we use the focal loss as a loss function to adjust sample weights and address the problem of imbalanced positive and negative samples during model training. Our experiments on three real-world datasets consistently show the superior performance of MAGNN-GC over state-of-the-art methods.