With the growing interest in improving recommendation accuracy, recent studies have increasingly turned to Knowledge Graph (KG)-driven recommender systems, recognizing their potential to overcome the shortcomings of traditional methods. However, many KG-based methods rely solely on node embeddings and fail to provide personalized and interpretable recommendations for diverse users. In order to tackle the aforementioned drawbacks, we introduce Knowledge-Aware Reasoning with User-Centeric Subgraph Network for Recommendation (KR-UCN), a Graph Neural Network (GNN)-based subgraph learning framework that combines path-based and embedding-based recommendation strategies to achieve more effective results. In KR-UCN, a distinct user-item subgraph is generated per interaction, structurally extract the key signals from user behavior patterns and the auxiliary knowledge embedded in the interaction and knowledge graph, while filtering out irrelevant noise through Personalized PageRank and the attention weights of the Graph Transformer. By calculating the values propagated along paths from user nodes to item nodes within the U-I subgraphs, and continuously updating user and item embeddings through alternating GNN and Graph Transformer layers, KR-UCN can provide accurate, efficient, and interpretable recommendations. Experimental findings indicate that KR-UCN outperforms existing KG-based and collaborative filtering (CF) approaches.