GTGNN: Global Graph and Taxonomy Tree for Graph Neural Network Session-Based Recommendation

The session-based recommendation aims to predict users’ short-term decisions by analyzing sequences of anonymous users’ historical behavior, often incorporating graph neural networks (GNN) for better performance. However, existing session-based recommendation methods tend to be limited to recommending items that already exist in users’ historical sessions, resulting in inadequate ability to recommend new items that users have never interacted with before, which leads to the so-called “information cocoon” phenomenon. To solve this issue, this paper proposes a global graph and taxonomy tree for graph neural network session-based recommendation (GTGNN). First, GTGNN constructs global and session graphs based on all session sequences. Subsequently, the GNN encoder is used to extract the global-level representation and session-level representation of the items. Second, GTGNN uses an item taxonomy tree to learn user intent from the perspective of attention mechanism and historical distribution data respectively, simulating the decision-making process when interacting with new items. Meanwhile, to solve the problem that GNN cannot learn new items, zero-shot learning is introduced to infer potential representations of new items and recommend new items with higher scores to users by calculating the recommendation scores of the corresponding items. Finally, extensive experiments on two real-world datasets (i.e., Amazon G&GF and Yelpsmall) show that the average enhancement of each evaluation metric on these datasets of the GTGNN reaches 3.25% and 3.02%, respectively.