In the era of information explosion, information overload has become a challenge for people to find information that interests them. Recommendation systems, as a kind of information filtering system, can understand users’ interests based on their personal data or historical behavior records, and are widely used in Web applications such as e-commerce, search, and streaming websites. However, traditional GNN-based recommendation algorithms can only handle regular topological graphs composed of a single type of nodes, while the data in today’s network is not composed of only a single type of nodes. In addition, traditional GNNs only fuse the first-order neighborhood features of nodes and cannot obtain the deeper structural relationships of nodes in the network. Therefore, when the data set is sparse and each node has only a very small number of neighbors, the recommendation quality of the traditional GNNbased recommendation algorithm decreases significantly. Heterogeneous graph neural networks provide new directions and ideas for recommendation research, and bring new challenges to data mining methods of heterogeneous graphs. Although the recommendation algorithm based on heterogeneous graph neural network is mainly based on the similarity of meta-paths, this method cannot effectively extract and utilize the rich structural and semantic information in heterogeneous graphs for recommendation, although the performance of the recommendation model can be improved by using the similarity of meta-paths. We proposed a joint multi-feature representation of information in heterogeneous information networks. For both user/item interaction and user socialization domains, the representation of node features in the case of sparse network connectivity is enhanced by using multi-order topological structure information in heterogeneous information networks. A graph neural network recommendation method oriented to attention mechanism is proposed. The recommendation quality of the recommendation model is enhanced in the case of sparse scoring matrix through the effective fusion of multidimensional representation vectors of users, items, ratings, and social. By conducting experiments on multiple public data sets, it is verified that the recommendation performance of our model in this paper significantly better than the baseline methods.