基于大语言模型强化的个性化推荐算法研究

Research on Personalized Recommendation Algorithm Enhanced by Large Language Models

摘要: 随着互联网平台内容规模的持续扩张，用户在海量候选项中快速找到符合自身兴趣的信息变得越来越困难，个性化推荐成为提升用户体验的重要手段。现有推荐方法大多以用户–项目交互数据为核心进行偏好建模，但在用户兴趣动态变化、交互数据稀疏以及多源异构信息利用不足等问题上仍存在明显局限。针对上述问题，本文提出了一种结合大语言模型语义增强、知识图谱结构建模与深度强化学习决策的个性化推荐方法。该方法将推荐过程建模为序列决策问题，在用户侧结合行为序列表示与语义画像构建状态表示，在项目侧联合学习文本语义表示与知识图谱结构表示，并通过深度Q网络完成候选项目的价值评估与推荐决策。在Yelp和Amazon-Book两个公开数据集上进行了实验，并与多种代表性推荐方法进行了对比。结果表明，所提出的方法在HR@K以及NDCG@K等指标上均取得了较优表现，验证了所提框架的有效性。

Abstract: With the continuous expansion of the content scale on internet platforms, it has become increasingly difficult for users to quickly find information that matches their interests from the vast number of options. Personalized recommendations have become an important means to enhance user experience. Most existing recommendation methods mainly model preferences based on user-item interaction data. However, there are still significant limitations in addressing issues such as dynamic changes in user interests, sparse interaction data, and insufficient utilization of multi-source heterogeneous information. To address these problems, this paper proposes a personalized recommendation method that combines semantic enhancement of large language models, knowledge graph structure modeling, and deep reinforcement learning decision-making. This method models the recommendation process as a sequence decision-making problem. On the user side, it combines behavioral sequence representation and semantic profiling to construct state representation. On the item side, it jointly learns text semantic representation and knowledge graph structure representation, and completes the value assessment and recommendation decision through a deep Q-network. Experiments were conducted on two public datasets, Yelp and Amazon-Book, and compared with several representative recommendation methods. The results show that the proposed method achieves superior performance in metrics such as HR@K and NDCG@K, verifying the effectiveness of the proposed framework.