Knowledge graph-based recommender systems have attracted increasing attention in recent years. By extracting the semantics of entities and relationships, these recommender systems enable a more comprehensive understanding of user preferences. However, existing methods overlook the impact of low-quality links in knowledge graphs, which leads to user preference propagation bias and affects recommendation accuracy. In this paper, we propose RKR-GAT, a recurrent knowledge-aware recommendation algorithm based on a graph attention network. Specifically, RKR-GAT integrates user-item interactions and the knowledge graph into a heterogeneous graph and learns node embeddings using a Graph Attention Network (GAT). To alleviate the effect of user preference propagation bias, we design a new path filtering module to reject low-quality connections and adaptively retain reasonable reasoning paths between users and items in the knowledge graph. Furthermore, RKR-GAT combines recurrent neural networks with a self-attention mechanism to efficiently analyze the semantics of the paths to generate more accurate recommendations while providing explainability. Ultimately, experimental results on three real-world datasets demonstrate the better performance of RKR-GAT compared to state-of-the-arts baselines.