Abstract: The existing knowledge graph recommendation models learn the user's remote potential interest by aggregating the high-order domain information of the entity. There are two problems in these methods. Firstly, the fixed size of the entity neighborhood structure is obtained by calculating the score between the user and the relationship, which cannot make full use of the global information in the knowledge graph. Secondly, the existing models aggregate the neighbor nodes of the entity with the same weight, without considering the preference of the target entity for different sampling neighbors. Based on the above problems, this paper proposes a graph convolution recommendation model which combine neighbor importance sampling and feature cross pooling. The sampling method based on neighbor importance obtains the importance of neighbor nodes by fusing the scores of neighbor nodes and the centrality perception scores, and then introduces the Bi-Interaction pooling layer to carry out the feature crossover and aggregation of the target entity vector and the neighborhood vector to obtain the final entity feature representation. Finally, the improved Sparrow Search Algorithm (SSA) is used to optimize the hyperparameters of Graph Convolutional Neural (GCN) network. This paper verifies the performance of the improved model on three data sets. Compared with baseline models, AUC and F1 indexes are increased by 3.0 % and 2.4 %.