Position-aware graph neural network for session-based recommendation

Session-based recommendations (SBRs) make recommendations using the current interaction sequence of users. Recent studies on SBRs have primarily used graph neural networks (GNNs) to model the session sequence; however, such methods treat the same items in a session as a single node, thus ignoring differences between items in different positions. Moreover, they do not use other sessions to learn users’ short-term preferences. Therefore, we propose a novel position-aware graph neural network (PA-GNN) for SBRs. First, this model uses a session in the form of a position-aware graph as an input to completely use the position information of the item and apply the attention mechanism to learn users’ long-term interests. Second, it combines other sessions to learn the user’s short-term preferences. Third, it integrates long-term interests and short-term preferences for predictions. The experimental results using three benchmark e-commerce datasets demonstrate that the PA-GNN model performs excellently and is superior to the latest baselines on SBRs.