在线教育蓬勃发展逐渐成为炙手可热的新型教学形式，在线教育平台为学生提供了海量的学习材料，进而推动个性化教育的普及和成熟。其中，能否通过认知诊断方法准确诊断学生对知识点的掌握程度进而有针对性地查漏补缺，是个性化教育的关键问题。

现有的认知诊断模型主要分为传统认知诊断模型和神经网络认知诊断模型两类。后者在较大程度上解决了传统方法“无法模拟复杂认知过程导致诊断表现欠佳”的问题。然而，神经认知诊断模型依旧存在以下不足之处：（1）未考虑不同知识点对于正确作答习题的重要程度不同，在模拟认知过程时未纳入知识点权重这一影响因素；（2）忽略知识点之间的潜在教育依赖关系，并且未针对数据集中知识点覆盖率低的问题提出解决方法。

针对上述两个问题，本文从“知识点权重”和“知识点的潜在关联”两个方面改进现有神经认知诊断模型。具体研究内容如下：

研究一提出基于知识点权重的神经认知诊断模型（KW-NCDM），通过“缩放点积”机制，计算知识点嵌入向量和习题嵌入向量之间的相关程度，以量化知识点对于习题的影响权重，并且在两个开源数据集上完成实证研究。结果表明：（1）在三个评价指标上，KW-NCDM模型的预测性能均优于基线模型；（2）KW-NCDM模型的可解释性更好，也更符合认知诊断理论中的单调性假设。

研究二提出基于知识点潜在关联的神经认知诊断模型（Self-NCDM），通过“自注意力”机制充分利用知识点间的潜在相关性来增强知识点特征向量，将知识点特征向量分别与学生因子、习题因子融合以强化神经认知诊断模型。实证研究结果表明：（1）在不同数据集切割比例下，Self-NCDM模型的性能普遍优于基准模型；（2）Self-NCDM模型的诊断结果更加合理，能有效解决弱知识点问题；（3）学生习题交互可视化分析表明Self-NCDM基本符合单调性假设；（4）在智慧教育系统中，Self-NCDM模型能有效区分不同认知水平的学生，因此契合个性化学习的要求。

The vigorous development of online education has gradually become a hot new form of teaching, and online education platforms provide students with a large number of learning materials, thereby promoting the popularization and maturity of personalized education. Among them, whether the cognitive diagnosis method can accurately diagnose students' mastery of knowledge concepts and then check and fill in the gaps in a targeted manner is the key issue of personalized education.

The existing cognitive diagnosis models are mainly divided into two categories: traditional cognitive diagnosis models and neural network cognitive diagnosis models. The latter largely solves the problem of “poor diagnostic performance due to the inability of traditional methods to simulate complex cognitive processes”. However, the neural cognitive diagnostic model still has the following shortcomings: (1) It does not consider the different importance of different knowledge concepts for answering the exercise questions correctly, and does not include the influencing factor of knowledge concept weight when simulating the cognitive process; (2) The potential educational dependence between knowledge concepts is ignored, and no solution is proposed to solve the problem of low coverage of knowledge concepts in the data set.

In order to solve the above two problems, this paper improves the existing neural cognitive diagnosis model from two aspects: “knowledge concepts weight” and “potential association of knowledge concepts”. The specific research contents are as follows:

In the first study, a neural cognitive diagnosis model based on knowledge concept weights (KW-NCDM) was proposed, and the correlation between the knowledge concept embedding vector and the exercise embedding vector was calculated through the “scaled point product” mechanism to quantify the influence weight of knowledge concepts on the exercises, and the empirical research was completed on two open-source datasets. The results showed that: (1) The prediction performance of the KW-NCDM model was better than that of the baseline model in the three evaluation indicators; (2) The interpretability of the KW-NCDM model is better, and it is more in line with the monotonicity assumption in cognitive diagnosis theory.

In the second study, a neural cognitive diagnosis model based on the latent association of knowledge points (Self-NCDM) was proposed, which made full use of the latent correlation between knowledge concepts to enhance the feature vector of knowledge points through the “self-attention” mechanism, and fused the feature vectors of knowledge points with student factors and exercise factors respectively to strengthen the neural cognitive diagnosis model. The empirical results show that: (1) The performance of the Self-NCDM model is generally better than that of the baseline model under different dataset cutting ratios; (2) The diagnostic results of the Self-NCDM model are more reasonable and can effectively solve the problem of weak knowledge concepts. (3) The interactive visual analysis of students' exercises shows that Self-NCDM basically conforms to the monotonicity hypothesis. (4) In the smart education system, the Self-NCDM model can effectively distinguish students with different cognitive levels, so it meets the requirements of personalized learning.