Graph convolutional neural networks have attracted increasing attention in recommendation system fields because of their ability to represent the interactive relations between users and items. At present, there are many session-based methods based on graph neural networks. For example, SR-GNN establishes a user’s session graph based on the user’s sequential behavior to predict the user’s next click. Although these session-based recommendation methods modeling the user’s interaction with items as a graph, these methods have achieved good performance in improving the accuracy of the recommendation. However, most existing models ignore the items’ relationship among sessions. To efficiently learn the deep connections between graph-structured items, we devised a dynamic attention-aware network (DYAGNN) to model the user’s potential behavior sequence for the recommendation. Extensive experiments have been conducted on two real-world datasets, the experimental results demonstrate that our method achieves good results in capturing user attention perception.