文献类型：期刊文献

中文题名：基于轻量级教育大模型的个性化实践学习资料动态推荐

英文题名：Dynamic Recommendation of Personalized Hands-on Learning Materials Based on Lightweight Educational LLMs

作者：翟洁[1];李艳豪[1];陈乐旋[1];郭卫斌[1]

机构：[1]华东理工大学信息科学与工程学院,上海200023

年份：2026

卷号：53

期号：2

起止页码：48

中文期刊名：计算机科学

外文期刊名：Computer Science

收录：;北大核心:【北大核心2023】；

基金：上海高校市级一流课程建设项目(沪教委高[2025]5号);2024年度教育部产学合作协同育人项目;2024年度教育部-华为“智能基座”产教融合协同育人基地一流课程建设项目;国家自然科学基金青年科学基金(62306112)。

语种：中文

中文关键词：轻量级教育大模型;个性化推荐;GCN算法;智能工作流;智能体;强化学习

外文关键词：Lightweight educational large model;Personalized recommendation;Graph convolutional network algorithm;Intelligent workflow;Intelligent agent;Reinforcement learning

摘要：人工智能技术在教育领域的深度应用,已成为国家教育数字化转型的核心战略。在计算机实践教学领域,实践学习资料的精准推荐是提升学生学习效能与质量的重要途径。针对高校教育规模化与学生需求多元化之间的矛盾,提出一种基于轻量级教育大模型的个性化实践学习资料推荐模型LightPLRec(Lightweight Personalized Learning Recommender for Dynamic Practice Materials),旨在依据学生个体特征的动态变化智能推荐个性化的实践学习资料。基于低算力需求的轻量级大模型,通过指令微调和强化学习方法构建了面向个性化实践学习资料推荐的教育大模型SPIR(Student Profile&Interest-based Re-commender)。通过整合多源异构数据,深度融入课程知识体系、学科前沿动态、产业发展趋势、国家战略导向,构建了跨学科、多模态的实践学习资料库,并设计了图转主题文本方法gragh2topic。依托于SPIR大模型的强大赋能和多源资料库的坚实支撑,提出了基于智能工作流的资料推荐方法。设计主题分析方法从学生能力评估结果中提取学生的能力特征,应用图卷积网络算法GCN从学生学习行为数据中挖掘学生的兴趣特征,创建了“能力-推荐智能体”和“兴趣-推荐智能体”,构建了双智能体协同驱动的智能化流程体系,实现了从学生个性化画像智能生成到实践学习资料动态推荐的系列工作流任务;并且构建了个性化资料推荐数据集,在该数据集上验证了所提模型的性能显著优于基线模型。其中,以Qwen2.5-3.0B为基模型训练的LightPLRec模型,在能力推荐与兴趣推荐这两项任务中展现出卓越性能,准确率分别高达0.947和0.939,其表现均优于DeepSeek-V3在同一数据集上的测评结果。该研究为教育大模型的垂直场景应用提供了技术范式,同时通过创建个性化实践学习资料动态推荐模型,为践行“因材施教”理念和培育高素质计算机实践人才提供了创新路径。
The deep integration of artificial intelligence(AI)technology in the education sector has become a core strategy for national educational digital transformation.Within the domain of computer practice teaching,the precise recommendation of practical learning resources serves as a vital pathway to enhance student learning efficacy and quality.Confronting the tension between the scale of higher education and the diversification of student needs,this study proposes a lightweight educational large model-based personalized practice learning resource recommendation framework,named LightPLRec(Lightweight Personalized Learning Re-commender for Dynamic Practice Materials).The model is designed to intelligently recommend tailored practical learning materials in response to the dynamic changes in individual student characteristics.Leveraging a lightweight large model with low computational demands,it constructs the SPIR(Student Profile&Interest-based Recommender)educational large model for personalized practical learning resource recommendation through instruction fine-tuning and reinforcement learning methods.By integrating multi-source heterogeneous data and deeply incorporating the curriculum knowledge system,disciplinary frontiers,industrial development trends,and national strategic orientations,it establishes a cross-disciplinary,multimodal practical learning resource repository and designs the graph2topic method for converting knowledge graphs into thematic text.Empowered by the robust capabilities of the SPIR large model and supported by the multi-source resource repository,it proposes an intelligent workflow-based recommendation method.Specifically,it designs a thematic analysis method to extract student competency features from assessment results,applies the GCN(Graph Convolutional Network)algorithm to mine student interest features from learning behavior data,and creates dual intelligent agents:a“Competency-Recommender Agent”and an“Interest-Recommender Agent”.This constructs a dual-agent collaboratively driven intelligent workflow system,enabling a series of tasks from the intelligent generation of personalized student profiles to the dynamic recommendation of practical learning resources.Furthermore,a persona-lized resource recommendation dataset is constructed,on which the proposed model demonstrates significantly superior perfor-mance compared to baseline models.Specifically,the LightPLRec model trained on the Qwen2.5-3.0B base model demonstrates outstanding performance in both the capability recommendation and interest recommendation tasks,achieving accuracies of 0.947 and 0.939 respectively,surpassing the evaluation results of DeepSeek-V3 on the same dataset.This research provides a technical paradigm for the vertical application of educational large models in specific scenarios.Simultaneously,by creating a dynamic personalized practical learning resource recommendation model,it offers an innovative pathway to implement the principle of“tea-ching students according to their aptitude”and cultivate high-quality computer practice talents.