With the development of the Internet in recent years, the network information is exploding, people have entered the era of big data, from the lack of information in the past to the great information overload nowadays. It has become an important challenge to filter the information that is beneficial to you in the ocean of information on the Internet. As a subset of the information filtering system, the recommendation system mainly relies on the historical interaction records of users' products and its own attribute information to explore the potential preferences and needs of users, which greatly reduces the time for users to filter information and is helpful for improving user experience and alleviating the information overload problem. Traditional recommendation algorithms represented by collaborative filtering often face the cold-start problem. They usually recommend products for users based on their purchase history or rating information of products, however, the performance of such recommendation algorithms decreases dramatically when the interaction history of users' products is sparse. Heterogeneous graph networks contain multiple types of nodes and edges due to their inclusion. It contains richer semantic information, and more general recommendation algorithms based on heterogeneous graphs are relatively less studied. Therefore, the purpose of this paper is to study graph neural network-based recommendation algorithms, mainly considering the heterogeneity of the network structure and the interaction order information between nodes, and then build a new recommendation model to cope with the cold start problem and improve the recommendation effect.