基于认知诊断的个性化试题推荐方法  

Cognitive Diagnosis Based Personalized Question Recommendation

作　　者：朱天宇[1] 黄振亚[1] 陈恩红[1] 刘淇[1] 吴润泽[1] 吴乐[2] 苏喻[3] 陈志刚[1] 胡国平 

机构地区：[1]中国科学技术大学计算机科学与技术学院,合肥230027 [2]合肥工业大学计算机与信息学院,合肥230009 [3]安徽大学计算机科学与技术学院,合肥230039 [4]科大讯飞股份有限公司,合肥230088

出　　处：《计算机学报》2017年第1期176-191,共16页Chinese Journal of Computers

基　　金：国家杰出青年科学基金(61325010);国家自然科学基金(61403358);中国科学院青年创新促进会会员专项基金(会员编号2014299);安徽省自然科学基金(1408085QF110)资助~~

摘　　要：面向学生的个性化试题推荐是智能教育领域重要的研究课题,现有的试题推荐工作大多采用协同过滤的方法或基于认知诊断的方法.然而,协同过滤的试题推荐方法往往忽略了学生的学习状态(知识点掌握情况);基于认知诊断的方法只能建模单个学生的学习状态,不能利用相似学生的共性特征.针对以上问题,文中提出一种基于学生知识点掌握程度的协同过滤试题推荐方法.该推荐方法分为3步:第1步结合认知诊断模型,根据学生已有的答题情况和试题知识点的关联对学生的试题掌握水平进行建模;第2步将学生的试题掌握水平用于概率矩阵分解预测学生的答题情况;第3步根据得分预测和试题难度向学生进行相应的试题推荐.该推荐方法同时考虑了被推荐学生学习的个性和群组学生学习的共性,在保证试题推荐解释性的同时提高了试题推荐的可靠性.最后,文中通过大量对比实验证明了该方法在进行学生试题推荐时能够保持精确性和可解释性.Personalized question recommendation for students is a significant research direction in the domain of intelligent education.Current studies depend on either collaborative filtering based methods or use the cognitive diagnosis models.Unfortunately,collaborative filterings ignore the knowledge states（e.g.skill proficiency）of students and cognitive diagnosis models can only model the knowledge state for a single student while the common features of similar students are neglected.To solve the problems above,we propose a three-step personalized question recommendationmethod（PMF-CD） which combines the complementary advantages of Probabilistic Matrix Factorization and Cognitive Diagnosis.Firstly,we model students＇ proficiency on each question by cognitive diagnosis models based on students＇ scores and question-skill correlation matrix.Secondly,we predict students＇ performance by probabilistic matrix factorization combined with students＇ question proficiency.Finally,we recommend questions to students according to their predicted performance and the difficulty of questions.PMF-CD takes both individual and common study status of students into account,which ensures the interpretability and effectiveness of recommendation.Extensive experiments on real-world datasets demonstrate that PMF-CD performs well,and the recommendation is also interpretative.

关 键 词：协同过滤 个性化试题推荐 概率矩阵分解 认知诊断 知识点掌握 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]