Abstract: To address the problems that existing models use knowledge graph embedding methods without any constraint, resulting in noise penetrating into the underlying information with embedding, meta-path-based recommendation algorithms only use a uniform weight assignment strategy without considering the nuances between meta-paths, and traditional models usually suffer from data sparsity and cold start, a knowledge graph embedded meta-path recommendation algorithm based on heterogeneous attention networks (MRHAN) is proposed, which defines meta paths to capture the complex semantic information between different types of entities and relationships, so as to better utilize the rich heterogeneous information to alleviate the data sparsity and cold start problems. A constraint-based approach, utilizing node relevance, is employed to model semantically related higher-order entities and relationships into unique meta-paths using knowledge graph embedding methods. A hierarchical attention network is used to model node preferences for different neighbors and different meta-paths, learning the weights of nodes for different neighbors and the weights of meta-paths for different recommendation tasks. Experimental results show that the model can effectively learn the representation of knowledge graphs, as well as the importance of node neighbors and meta-paths, and further alleviate the data sparsity and cold-start problems.