With the proliferation of the Internet and the vast amount of information available, recommendation systems have become indispensable for filtering and delivering personalized content to users. Collaborative filtering, one of the most widely used recommendation algorithms, predicts user preferences based on historical interactions. However, as the number of users and items increases, challenges such as sparsity and cold start arise, hindering personalized recommendations. To address these challenges, recommendation algorithms have started incorporating knowledge graphs, structured data repositories, into their models. These algorithms can be classified into embedding-based, rule-based, and aggregation-based approaches. Embedding-based algorithms leverage existing knowledge graph embedding models to obtain entity embeddings, which are then used as input features for recommendation prediction. Rule-based algorithms focus on modeling connections between users and items using paths in the knowledge graph. They can be categorized as meta-path-based or automatic path mining approaches. Aggregation-based algorithms aggregate entities using graph neural networks, such as graph convolutional neural networks (GCNs) or graph attention networks (GATs), to capture relationships between entities at different distances. However, while aggregation-based approaches excel in modeling accuracy, they may lack generalization capability. To overcome this limitation, this paper introduces RKGREC, a recommendation algorithm that combines rule-based modeling with graph neural networks. By leveraging the strengths of both approaches, RKGREC achieves high modeling accuracy through rules while maintaining high generalization capability through entity aggregation. The integration of rules and graph neural networks enhances the semantic understanding of connections between users and items, reducing noise and improving recommendation performance.