面向序列推荐的大模型知识注入方法研究
Research on a Knowledge Injection Method of Large Language Models for Sequential Recommendation
白文涵 1贾彩燕1
作者信息
- 1. 北京交通大学计算机科学与技术学院，北京 100044
- 折叠
摘要
在互联网和移动应用快速发展的背景下，推荐系统已成为缓解信息过载、提升用户体验的重要技术手段。近年来，序列推荐通过建模用户行为序列中的兴趣演化规律，在电商、短视频和新闻等场景中取得了显著进展。然而，现有序列推荐模型大多依赖用户-物品交互数据进行学习，物品表示通常基于ID embedding，缺乏丰富的语义信息，难以刻画物品之间潜在的语义关系，并在数据稀疏或新物品出现时表现受限。大语言模型在语义理解和知识表达方面具有显著优势，为推荐系统引入外部知识提供了新的可能。已有研究尝试利用大语言模型生成语义表示以增强推荐性能，但仍存在融合机制复杂或语义信息利用不足的问题。针对上述不足，本文提出一种面向序列推荐的大模型知识注入方法KIRec。该方法首先利用大语言模型对物品文本进行属性抽取，获得高维语义嵌入；随后通过降维映射使语义表示适配序列推荐模型的表示空间；最后将语义嵌入与原始物品嵌入进行融合训练，从而增强物品表示能力并提升推荐效果。实验在公开数据集上进行，并与多种序列推荐模型及典型LLM-based方法进行对比。结果表明，所提出方法在HR和NDCG等指标上均取得更优性能，同时保持较低的模型复杂度，验证了该方法的有效性与通用性。
Abstract
With the rapid development of the Internet and mobile applications, recommender systems have become an essential technique for alleviating information overload and improving user experience. In recent years, sequential recommendation has achieved significant success in scenarios such as e-commerce, short-video platforms, and news services by modeling the evolution of user interests from historical interaction sequences. However, most existing sequential recommendation models rely heavily on user-item interaction data, where item representations are typically learned from ID embeddings. Such representations lack rich semantic information and fail to capture the latent relationships among items, leading to limited performance in sparse data scenarios or when new items emerge.Large Language Models (LLMs) have demonstrated strong capabilities in semantic understanding and knowledge representation, providing new opportunities for introducing external knowledge into recommender systems. Recent studies attempt to leverage LLMs to generate semantic representations for enhancing recommendation performance. Nevertheless, these methods often suffer from complex integration mechanisms or insufficient utilization of semantic knowledge.To address these limitations, this paper proposes KIRec, a knowledge injection framework for sequential recommendation based on large language models. Specifically, we first employ an LLM to extract attribute-level semantic information from item textual content to obtain high-dimensional semantic embeddings. Then, a dimensionality reduction mapping is applied to adapt these semantic representations to the embedding space of sequential recommendation models. Finally, the semantic embeddings are fused with the original item embeddings to jointly train the recommendation model, thereby enhancing item representation and improving recommendation accuracy.Extensive experiments are conducted on public benchmark datasets, comparing our method with several sequential recommendation models and representative LLM-based approaches. The results demonstrate that the proposed method consistently improves recommendation performance in terms of HR and NDCG while maintaining relatively low model complexity, verifying its effectiveness and general applicability.关键词
人工智能/序列推荐/大语言模型/知识注入/推荐系统Key words
Artificial Intelligence/Sequential Recommendation/Large Language Models/Knowledge Injection/Recommendation Systems引用本文复制引用
白文涵,贾彩燕.面向序列推荐的大模型知识注入方法研究[EB/OL].(2026-03-26)[2026-03-27].http://www.paper.edu.cn/releasepaper/content/202603-258.学科分类
计算技术、计算机技术
评论