Introducing the entity and relationship information in the knowledge graph during recommendation is an effective way to alleviate the problem of cold start. The HAN model introduces the attention mechanism-based graph neural networks into heterogeneous graphs for the first time. However, it does not make full use of the high-order neighbor information of nodes. To solve this problem, the study proposes a recommendation model CKG-HAN that integrates the high-order neighbor features of the collaborative knowledge graph. The model employs meta-paths to connect project nodes and divides the collaborative knowledge graph into multiple subgraphs. The high-order neighbor features of each node in the subgraph are aggregated in the node attention layer of the model, and different weights are assigned to node features on different meta-paths by the relation attention layer. Finally, a node embedding representation is obtained which fully integrates semantic information. The Top-K recommendation is performed on the MovieLens-1M data set, and the results show that the model proposed in this study can effectively improve the accuracy of the recommendation results.