This study proposed a personalized recommender system based on LLM and explored how to better capture and utilize user preferences to increase recommendation accuracy and explainability. Traditional recommender methods, such as collaborative filtering, content-based, and hybrid approaches, often overlook the relationships between user preference attributes. This study leverages LLM, considers the relationships between user preference attributes, and combines it with a self-attention mechanism to achieve fine-grained scoring of user preferences. The system delved into user preferences through dialog, adjusting the recommendation content in each turn of the dialog. The experimental results demonstrated significant improvements in both the recommendation accuracy and the quality of the generated dialog content. Specifically, on the ReDial dataset, Recall@10 improved by 2.9%, and the informativeness increased by 3%. The study also emphasized the importance of attribute sorting order, further enhancing the system's explainability.