Abstract: Traditional recommendation methods face challenges in semantic understanding and interest modeling under sparse interaction conditions. Meanwhile, course relationships commonly exhibit strong contextual dependency and insufficient structural constraints. To overcome these limitations, a learning recommendation approach that integrates a Large Language Model (LLM) with a Multimodal Knowledge Graph (MMKG) was proposed. Semantic representations of user behaviors and course content were captured by the LLM to enrich user and item semantics, while the MMKG was constructed from user-course interactions, and a Graph Convolutional Network (GCN) was leveraged to aggregate multimodal features and high-order relational information. A semantic-structural fusion mechanism was further designed to jointly model semantic and graph-based representations. Experimental results show that, across multiple domain datasets and different recommendation ranges, the proposed method consistently outperforms baseline approaches based on Knowledge Graphs (KGs), LLM, and their collaborative modeling in terms of Recall and Normalized Discounted Cumulative Gain (NDCG). Specifically, on the MOOCCube-based learning dataset with different recommendation range settings, Recall is improved by approximately 2%–5% compared with the K-RagRec method, while NDCG maintains its advantage, demonstrating good recommendation accuracy and generalization ability.

Key words: Large language Model (LLM), MultiModal Knowledge Graph (MMKG), learning recommendation, Graph Convolutional Network (GCN), collaborative modeling

摘要： 传统推荐方法在稀疏条件下语义理解和兴趣建模方面存在一定局限，同时课程关系中普遍存在较强的上下文依赖性和结构约束不足等问题。针对上述问题，提出一种融合大语言模型(LLM)与多模态知识图谱(MMKG)的学习推荐方法。该方法利用LLM对用户行为与课程内容进行语义表示学习，从而有针对性地补充用户兴趣与课程内容的语义特征；同时,根据用户-课程关系构建协同MMKG，通过图卷积神经网络(GCN)聚合模态特征和高阶交互关系，并构建语义与多模态特征的融合机制，以提升聚合后结构表示的语义表征能力，实现两者的协同建模。实验结果表明，所提方法在多个领域数据集和不同推荐范围下，与基于知识图谱(KG)、LLM及其协同建模的基线方法相比，在召回率(Recall)和归一化折损累积增益(NDCG)指标上均取得稳定提升。其中，在基于MOOCCube构建的学习数据集上的不同推荐范围设置下，与K-RagRec方法相比，Recall提升约2%~5%，NDCG仍保持一致优势，展现出较好的推荐准确性和泛化能力。

关键词: 大语言模型, 多模态知识图谱, 学习推荐, 图卷积神经网络, 协同建模