基于学生知识追踪的多指标习题推荐算法

个性化习题推荐是教育信息化时代的重要课题，传统的习题推荐算法忽略了学生在学习过程中的遗忘规律，未能充分挖掘学生的知识掌握水平和相似学生的共性特征，推荐习题的评价指标单一，推荐习题的新颖度和多样性不足，不能合理地促进学生对新知识的学习或帮助学生查缺补漏。针对上述缺陷，提出一种基于学生知识追踪的多指标习题推荐方法，该方法分为习题初筛和再过滤两个模块，围绕习题推荐的新颖度、难度以及多样性 3 个指标展开研究，首先构造了一个结合学生遗忘规律的知识概率预测（student forgetting behavior based knowledge concept coverage prediction，SF-KCCP）模型，保证推荐习题的新颖性；接着基于动态键值的知识追踪（dynamic key-value memory networks for knowledge tracing，DKVMN）模型精准挖掘学生的知识概念掌握水平，以保证推荐合适难度的习题；最后将基于用户的协同过滤（user-based collaborative filtering，UserCF）算法融入再过滤模块，利用学生群体之间的相似性实现推荐结果的多样性。通过大量的实验证明，所提方法比一些现有的基线模型取得了更好的性能。

Student knowledge tracking based multi-indicator exercise recommendation algorithm

Personalized exercise recommendation was an important topic in the era of education informatization, the forgetting laws of students in the learning process were ignored by the traditional problem recommendation algorithm, which failed to fully tap the students’ knowledge mastery level and the common characteristics of similar students, insufficient, could not reasonably promote students’ learning of new knowledge or help students find and fill omissions.In view of the above defects, a multi-index exercise recommendation method based on student knowledge tracking was proposed, which was divided into two modules: preliminary screening and re-filtering of exercises, focusing on the novelty, difficulty and diversity of exercise recommendation.Firstly, a knowledge probability prediction (SF-KCCP) model combined with students’ forgetting law was constructed to ensure the novelty of the recommended exercises.Then, students’ knowledge and concept mastery level was accurately excavated based on the dynamic key-value knowledge tracking (DKVMN) model to ensure that exercises of appropriate difficulty were recommended.Finally, the user-based collaborative filtering (UserCF) algorithm was integrated into the re-filtering module, and the similarity between student groups was used to achieve the diversity of recommendation results.The proposed method is demonstrated by extensive experiments to achieve better performance than some existing baseline models.