High-order attentive graph neural network for session-based recommendation

Recommender systems are becoming a crucial part of several websites. The purpose of session-based recommendations is to predict the next item that users might click based on users' interaction behavior in a session. The latest research on session-based recommendation focuses on using graph neural networks to model transfer relationships between items. However, when the interaction of low-order relationships between adjacent items is insufficient, learning the high-order relationships between non-adjacent items becomes a challenge. Additionally, to distinguish the importance of nodes in the graph, different weights should be assigned to each edge. Therefore, we propose a novel high-order attentive graph neural network (HA-GNN) model for session-based recommendations. In the proposed method, first, we model sessions as graph-structured data. Then, we use the self-attention mechanism to capture the dependencies between items. Next, we use the soft-attention mechanism to learn high-order relationships in a graph. Finally, we update the embeddings of items using a simple fully connected layer. Experiments on two public e-commerce datasets show that HA-GNN has excellent performance.