Abstract: Graph neural networks face the challenge of data sparsity and model over-smoothing problem in knowledge graph-based recommendation algorithms, making it difficult for recommendation systems to capture users' long-term preferences. To address this issue, a recommendation algorithm based on deep knowledge graph convolutional networks is proposed. Firstly, multiple views of the knowledge graph are obtained through a random sampling process based on historical interaction data to alleviate the problem of sparse interaction data. Secondly, residual connections are introduced in the knowledge graph convolutional operation to alleviate the over-smoothing problem, thus constructing a deep knowledge graph convolutional network to capture high-order neighbor information and obtain users' potential remote interests. To avoid noisy input brought by the deep graph convolutional network, the graph self-attention mechanism is used to refine the deep feature representations. To verify the feasibility and effectiveness of the model, extensive experiments were conducted on three real-world datasets, and the results showed that the proposed model can effectively alleviate the over-smoothing problem. In these three datasets, the model achieved Area Under Curve(AUC, A_UC) scores of 0.974, 0.812 and 0.732 respectively. Compared to existing methods, the model showed significant improvements in recommendation accuracy.