Order-Aware Graph Neural Network for Sequential Recommendation

Graph neural networks (GNNs) have gained impressive success in the task of sequential recommendation due to their advantage in obtaining the complex transition patterns of items. However, existing GNN-based sequential recommenders still face some problems: (1) The global order information is lost when converting a sequence into a graph. (2) The long-term dependencies in a sequence are ignored due to the over-smoothing problem in GNNs. In this paper, we propose an order-aware GNN with long-range connections (OAG-LC) for sequence modeling. To capture the global order of a sequence, a novel graph update mechanism is proposed, which evolves the graph embedding recurrently over time rather than concurrently for order preservation. And a novel gate is used to incorporate both order and structural information in the update phase. To model the long-term dependencies of user behaviors, we convert the sequence into a graph via reachability and apply the attention mechanism for information propagation through the long-range connections. Furthermore, the proposed graph construction method differentiated repeated items with their positions for information lossless encoding. We conduct extensive experiments on four public datasets, and the experimental results demonstrate the effectiveness of our proposed model.