Self-adaptive Graph Neural Networks for Personalized Sequential Recommendation

Sequential recommendation systems have attracted much attention for the practical applications, and various methods have been proposed. Existing methods based on graph neural networks (GNNs) mostly capture the sequential dependencies on an item graph by the historical interactions. However, due to the pre-defined item graph, there are some unsuitable edges connecting the items that may be weakly relevant or even irrelevant to each other, which will limit the ability of hidden representation in GNNs and reduce the recommendation performance. To address this limitation, we design a new method called Self-Adaptive Graph Neural Networks (SA-GNN). In particular, we employ a self-adaptive adjacency matrix to improve the flexibility of learning by adjusting the weights of the edges in the item graph, so as to weaken the effect of unsuitable connections. Empirical studies on three real-world datasets demonstrate the effectiveness of our proposed method.