Abstract: Knowledge tracing involves tracking students’ knowledge state in real-time based on their past answering performance and predicting their future performance, which is crucial for personalized education. In recent years, RNN-based deep knowledge tracing models have gradually become the mainstream research approach in the field of knowledge tracing. However, existing knowledge tracing models suffer from the inability to capture long-term dependencies between sequences and the neglect of the relationship between questions and knowledge points, resulting in insufficient extraction of question features. To address these issues, a knowledge tracing model based on question feature enhancement (QFEKT) is proposed. Firstly, graph convolutional neural network is used to model the relevant features of questions and knowledge points, and contrastive learning is introduced to enhance feature representation. Then, the question features are further enhanced through question matching module and student knowledge state representation module. The question matching module extracts similar questions as a complement to the question features. The student knowledge state representation module combines bidirectional long short-term memory networks with attention mechanism to enhance question features and model students’ knowledge state. Finally, the prediction module integrates similar question features and student knowledge state to predict students’ future performance. Comparative experiments on three publicly available real-world datasets demonstrate that the QFEKT model outperforms other baseline models in knowledge tracing tasks and exhibits significant advantages in predicting students’ future performance.

Key words: knowledge tracing, feature enhancement, graph convolutional neural network, contrastive learning, attention mechanism

摘要： 知识追踪根据学生过去的答题表现实时跟踪学生的知识状态并预测学生未来的答题表现，是实现个性化教学的关键。近年来，基于RNN的深度知识追踪模型逐渐成为知识追踪领域中的主流研究方法。但是，现有的知识追踪模型存在无法捕获序列间长期依赖以及忽略了问题与知识点间关系的问题，导致无法充分提取问题特征。针对上述问题，提出了基于问题特征增强的知识追踪模型QFEKT。使用图卷积神经网络对问题和知识点相关特征进行建模，建模过程中引入对比学习提升特征表示水平。通过问题匹配模块与学生知识状态表征模块进一步增强问题特征：通过问题匹配模块提取相似问题作为问题特征的补充；通过学生问题表征模块将双向长短期记忆网络与注意力机制结合增强问题特征建模学生的知识状态。预测模块融合相似问题特征与学生知识状态预测学生未来答题表现。在三个公开真实数据集上进行对比实验，QFEKT模型与其他基线模型相比可以更好完成知识追踪任务，在预测学生未来答题表现上具有明显优势。

关键词: 知识追踪, 特征增强, 图卷积神经网络, 对比学习, 注意力机制