计算机科学与探索 ›› 2024, Vol. 18 ›› Issue (6): 1600-1612.DOI: 10.3778/j.issn.1673-9418.2305096

• 人工智能·模式识别 • 上一篇    下一篇

融合热点与长短期兴趣的图神经网络课程推荐模型

刘源,董永权,陈成,贾瑞,印婵   

  1. 1. 江苏师范大学 计算机科学与技术学院,江苏 徐州 221116
    2. 徐州市云计算工程技术研究中心,江苏 徐州 221116
    3. 江苏省教育信息化工程技术研究中心,江苏 徐州 221116
  • 出版日期:2024-06-01 发布日期:2024-05-31

Graph Neural Network Integrating Hot Spots and Long and Short-Term Interests for Course Recommendation

LIU Yuan, DONG Yongquan, CHEN Cheng, JIA Rui, YIN Chan   

  1. 1. College of Computer Science and Technology, Jiangsu Normal University, Xuzhou, Jiangsu 221116, China
    2. Xuzhou Cloud Computing Engineering Technology Research Center, Xuzhou, Jiangsu 221116, China
    3. Educational Informatization Engineering Technology Research Center, Xuzhou, Jiangsu 221116, China
  • Online:2024-06-01 Published:2024-05-31

摘要: 近年来大规模在线开放课程(MOOCs)平台为用户提供了海量的学习资源,亟需一种有效的个性化课程推荐方法帮助用户解决信息过载问题。现有的课程推荐方法忽略了课程间的时序性且无法较好地捕获课程间的长距离依赖关系,同时面临用户学习兴趣表示和冷启动两个关键问题。基于此,提出一种融合热点与长短期兴趣的图神经网络课程推荐模型(GHLS4CR)。该模型设计无环时序图和无环快捷图两种会话图构建方法来缓解现有方法存在的时序信息丢失和不善于捕获长距离依赖的问题;将用户长短期兴趣进行图级表示,并与热门课程信息进行融合实现个性化推荐,同时缓解冷启动问题。通过在学堂在线(XuetangX)公开数据集MOOCCourse上的大量实验表明,GHLS4CR在个性化课程推荐领域优于FISSA和LESSR等主流推荐模型。与次好的LESSR模型相比,Recall@5提高了13.28%,MRR@5提高了15.50%。

关键词: 课程推荐, 基于会话的推荐, 图神经网络, 长短期兴趣, 冷启动

Abstract: In recent years, massive online open courses (MOOCs) platforms provide users with a wealth of learning resources. Nevertheless, information overload remains a pressing concern, necessitating the development of effective personalized course recommendation methods. The existing course recommendation methods disregard the temporal relationship among courses and are unable to capture long-distance dependencies between them. Simultaneously, personalized course recommendation models designed for interactive sequence modeling are confronted with two key issues: how to extract users’ learning interest representation effectively and how to solve cold-start. Based on this, a graph neural network course recommendation model (GHLS4CR) is proposed, which integrates hot spots and long and short-term interests. This model designs two session graph conversion methods, acyclic timing graph and acyclic shortcut graph, to alleviate the problems of temporal information loss and inability to capture long-distance dependencies in existing methods. This model represents users’ long-term and short-term interests at the graph level, and integrates them with popular course information to achieve personalized recommendations while alleviating cold-start issue. A large number of experiments conducted on the XuetangX public dataset MOOCCourse show that GHLS4CR outperforms mainstream recommendation models such as FISSA and LESSR in the field of personalized course recommendation. Compared with the second best LESSR model, Recall@5 is improved by 13.28%, and MRR@5 is improved by 15.50%.

Key words: course recommendation, session-based recommendation, graph neural networks, long and short-term interests, cold-start