By acquiring knowledge from a graph，the recommendation algorithm based on the graph neural network improves the recommendation interpretability. However，with the continuous expansion of the network data scale of the recommended system，the user-item scoring matrix displays a sparsity problem，which makes the graph neural network difficult to learn high quality network node features，and finally leads to the decline of recommendation quality. In this paper，a recommendation algorithm based on heterogeneous graph neural network is proposed by combining graph neural network with heterogeneous information network. This algorithm uses heterogeneous information network to decode multi-source heterogeneous data. And the attention mechanism is introduced into the user and item aggregation process of user-item interaction network and user social network，in order to realize the effective fusion of Node and topology characteristics of user-item interaction and user social networks. The experiment on two continuous sparse datasets show that the recommendation error of the algorithm proposed in this paper is 40% less than that of the baseline method.