大语言模型驱动下基于用户画像的在线课程推荐方法研究

邓伟伟¹  卢骏维¹  霍栩俊¹  陈  寒^2*

（1.华南师范大学 经济与管理学院，广东 广州 510006；2.华南师范大学 教师教育学部，广东 广州 510631）

【摘  要】随着在线课程资源的快速增长，学习者面临日益严重的信息过载问题，而现有推荐方法往往难以捕捉其随时间变化的细粒度偏好，且推荐结果的可解释性不足。针对上述问题，本研究提出了一种大语言模型驱动的在线课程推荐方法，该方法首先分析用户按时间顺序排列的历史选课记录，利用大语言模型的上下文理解与结构化输出能力，挖掘用户在课程内容、学科领域、授课教师与所属院校等维度上的细粒度偏好，构建反映兴趣演变的用户画像。随后，提出两阶段语义匹配推荐策略：第一阶段将用户画像中的学科和内容偏好与候选课程属性进行匹配，筛选相关课程；第二阶段结合教师与院校偏好进行复核，生成推荐结论并提供理由。在MOOCCubeX真实数据集上的实验结果表明，本方法在精准率、召回率与F1指标上整体优于对比方法；消融分析显示，内容/学科匹配贡献最大，时序信息、思维链提示以及教师/学校信息具有补充作用。本研究将时序用户画像、候选课程语义属性与分步推荐推理结合，为在线课程推荐提供了可追踪、可解释的实现路径。

【关键词】大语言模型；用户画像；课程推荐

Research on an Online Course Recommendation Method Based on User Profiling Driven by Large Language Models

DENG Weiwei¹, LU Junwei¹, HUO Xujun¹ and CHEN Han^2*

(1. School of Economics and Management, South China Normal University, Guangzhou 510006, China; 2. College of Teacher Education, South China Normal University, Guangzhou 510631, China)

Abstract: With the rapid growth of online course resources, learners are confronted with an increasingly severe information overload problem, while existing recommendation methods often struggle to capture their fine-grained preferences that evolve over time, and the recommendation results are insufficiently interpretable. To address these issues, this study proposes a large language model-driven online course recommendation method. The method first analyzes users’ enrollment records arranged in chronological order, leverages the contextual understanding and structured output capabilities of LLMs to mine users’ fine-grained preferences across dimensions such as course content, subject areas, instructors, and affiliated institutions, and constructs user profiles that reflect the evolution of their interests. Subsequently, a two-stage semantic matching recommendation strategy is proposed: in the first stage, the subject and content preferences in the user profiles are matched with candidate course attributes to filter relevant courses; in the second stage, instructor and institution preferences are incorporated for re-examination, generating recommendation conclusions and providing justifications. Experimental results on the real-world MOOCCubeX dataset show that the proposed method overall outperforms the comparison methods in precision, recall, and F1 score. Ablation analysis reveals that content/subject matching contributes the most, while temporal information, chainof-thought prompting, and instructor/institution information play supplementary roles. By integrating temporal user profiles, semantic attributes of candidate courses, and stepwise recommendation reasoning, this study provides a traceable and interpretable implementation pathway for online course recommendation.

Keywords: large language models; user profiling; course recommendation

下载： 大语言模型驱动下基于用户画像的在线课程推荐方法研究.pdf