The importance between source and target nodes is often overlooked in higher-order information, which can introduce excessive noise in complex network scenarios, thereby affecting recommendation performance. To address this issue, we propose a graph neural network-based knowledge graph recommendation model that integrates deep domain information and important domain information, aiming to enhance recommendation accuracy and diversity. Initially, we pre-train the knowledge graph using graph embedding techniques to obtain structural information. Subsequently, we employ a graph convolutional network to delve deeper into the semantic information of the knowledge graph, capturing rich structural and semantic details from both depth and importance perspectives. Finally, we compute the interaction probability between users and items using the inner product of the enhanced vectors and user vectors, facilitating recommendations. Experiments on the Last-FM, Book-Crossing, and MovieLens-20M datasets yield AUC and F1 scores of 83.2% and 75.4%, 74.9% and 66.8%, 97.9% and 93.1%, respectively. Additionally, Recall@50 scores are 34.5%, 11.2%, and 35.0%. Our model outperforms RippleNet, KGCN, LKGR, and other models, indicating that integrating knowledge graph recommendation models with meta-graph neighborhoods effectively improves recommendation performance.