With the rapid development of network technology, recommendation systems attract increasing research because of its wide applications in e-commerce. Nevertheless, most existing recommendation models based on graph neural networks do not consider the transitivity of subgraph structures in interactive data. This makes the models unable to capture the complex dependencies and mutual influences between users and items, resulting in the inability to achieve high-quality personalized recommendations. To address the above challenge, we propose a novel recommendation algorithm based on knowledge graph with high-order graph convolutional network, named KG2CN. Firstly, we introduce the subgraph structure on the knowledge graph to capture high-order contextual information between users and items. Secondly, the mined subgraph information and graph convolutional network are combined to learn high-order features of users and items. Finally, the decoder is applied to predict the ratings of target users on the uninteracted items, thereby recommending the Top-K items. Experimental results on the Book-Crossing and Last.FM datasets show that the proposed KG2CN obtains better performance in F1 score and AUC metrics.