As an important link in the information service of Internet products, the recommendation system has become an important way for users to get information from the huge amount of Internet data. Among existing studies, collaborative filtering has become a widely used method in recommender systems due to its good performance, but it uses a shallow model that cannot learn the deep non-linear features of users and items. In addition, content-based recommendation methods make recommendations by making full use of user registration information and item profiles, but this method also requires effective feature extraction, relying on feature engineering i.e., by manually extracting or design features, which makes the effectiveness and scalability of the method very limited and limits the performance of the recommendation algorithm. In recent years, deep learning has made a great impact in natural language processing, speech recognition, and image processing. This has led to a new breakthrough in the research of recommendation systems. Graph neural network algorithms extend traditional deep learning methods, such as convolution, to the domain of graph data, and combine the idea of data propagation to form deep learning algorithms on graphs, which have achieved good results in social networks, recommendation systems, knowledge graphs, and other fields. In this paper, we focus on the shallow output of the intermediate layer output. A multi-headed attention mechanism is used to fuse the multi-layer output information, so that the shallow structural features provided by the intermediate layer can be better involved in the scoring prediction task, which makes the application of graph neural networks in recommendation systems further improved.