Abstract: In recent years, cognitive diagnostic assessment is an area of research that has attracted widespread attention. As we all know, one of the important components in cognitive diagnosis is Q-matrix, because Q-matrix reflects the design of the assessment instrument and is the core element that determines the quality of the diagnostic feedback for the instrument. At present, there are some researches about classification accuracy in DINA model with error existed in Q-matrix. These studies indicate that the quality of the Q-matrix has a great influence on the diagnostic accuracy rate, and also indicate that cognitive diagnosis models constructed around Q-matrix are sensitive to the accuracy of Q-matrix, greatly influenced by Q-matrix, and mostly, the starting point of these research are “if the Q-matrix contains errors, how does it affect the accuracy of parameters estimation and classification accuracy”. Up to now, the most problem is that we haven’t an effective method for validating the Q-matrix at hand. Different diagnostic models have different diagnostic classification accuracy rate, and affected by factors that are not the same. Bayesian networks is one of a widespread concerned model, it has strong processing capacity to uncertainly problem. Starting from another perspective view, uses Bayesian network model which less affected by Q-matrix as diagnosis classification model. Compares Bayesian network with the DINA model in cognitive diagnostic classification accuracy on the base of a Q-matrix which contains errors. Bayesian network classification model is less affected by the Q-matrix than DINA model. Then, two simulation studies are carried out. The first is to study the performance of DINA and Bayesian network classification model when the Q-matrix contains error items, the data is generated under DINA model. To be fair, the data generated in the second research doesn’t base on any specified models, adopts the method introduced by Leighton, Gierl & Hunka(2004). Investigates the effect of different type of Q-matrix (contains a reachable matrix or not), contain different type of error (contain 0, 5, 7, 10, 13, 15 items which have 0, 1, 2, 3 erred calibrate attributes) during classification in different models. The performance of Bayesian network classification model was superior in many cases than DINA model. When Q matrix contained a reachable matrix and 5(or less) error specified items, the performance of DINA model was slightly better than the Bayesian network classification model; but when Q matrix didn’t contained a reachable matrix, or contained more than 5 error specified items, the Bayesian network classification model is better than DINA model.

Key words: cognitive diagnosis, Q matrix, Bayesian Networks, reachable matrix, DINA model

摘要： Q矩阵是认知诊断测验的重要组成部分之一，围绕Q矩阵构建的诊断模型对Q矩阵中包含的错误较敏感。贝叶斯网分类模型是基于网络结点之间的关系构建的模型，将朴素贝叶斯网作为诊断模型，与DINA模型进行比较。模拟实验结果表明：Q矩阵中是否包含可达矩阵和错误界定的项目数量对DINA模型影响较大，对贝叶斯网模型影响较小；项目数量对DINA和贝叶斯网模型影响都较大；样本大小对贝叶斯网模型影响较大，对DINA模型影响较小。模拟研究结果显示，当Q矩阵中不包含可达阵、包含5个以上错误项目或样本数较大时，贝叶斯网分类模型优于DINA模型；而当Q矩阵中包含可达阵和5个(以下)错误项目时，DINA模型优于贝叶斯分类模型。

关键词: 认知诊断, Q矩阵, 贝叶斯网, 可达矩阵, DINA模型