Multiple Attention and Adaptive Fusion Item Category graph for Session-based Recommendation

Session-based recommendation focus on predicting the next clicked item with limited information from anonymous users. Most existing session-based recommendation methods rely on graph neural network and self-attention techniques. However, these methods overlook the potential synergies between the two that could enhance recommendation performance. Furthermore, self-attentive networks lack the capability to consider dependencies among different pieces of information, limiting their capacity to learn contextual item representations within a sequence. We introduce MHAW-GNN, a novel model that leverages an adaptive weight fusion mechanism and normalization process to combine remote relationships and global data. The model dynamically adjusts the importance of these factors based on user preferences. MHAW-GNN incorporates multi-head attention to capture remote dependencies among items and categories within a session, enhancing contextual relational understanding through graph neural networks. By utilizing FFN, the model refines features to address both over-smoothing issues in item representation and overfitting challenges. Extensive experiments on two real-world datasets consistently demonstrate that MHAW-GNN surpasses existing state-of-the-art methods.