Exploring latent connections in graph neural networks for session-based recommendation

Session-based recommendation, without the access to a user’s historical user-item interactions, is a challenging task, where the available information in an ongoing session is very limited. Previous work on session-based recommendation has considered sequences of items that users have interacted with sequentially. Such item sequences may not fully capture the complex transition relationship between items that go beyond the inspection order. This issue is partially addressed by the graph neural network (GNN) based models. However, GNNs can only propagate information from adjacent items while neglecting items without a direct connection, which makes the latent connections unavailable in propagation of GNNs. Importantly, GNN-based approaches often face a serious overfitting problem. Thus, we propose Star Graph Neural Networks with Highway Net- works (SGNN-HN) for session-based recommendation. The proposed SGNN-HN model applies a star graph neural network (SGNN) to model the complex transition relationship between items in an ongoing session. To avoid overfitting, we employ the highway networks (HN) to adaptively select embeddings from item representations before and after multi-layer SGNNs. Finally, we aggregate the item embeddings generated by SGNN in an ongoing session to represent a user’s final preference for item prediction. Experiments are conducted on two public benchmark datasets, i.e., Yoochoose and Diginetica. The results show that SGNN-HN can outperform the state-of-the-art models in terms of Recall and MRR for session-based recommendation.