随着教育信息化的不断推进和人工智能技术的发展，知识追踪模型成为了 教育大数据时代的一项关键研究。其可以通过分析学习者的行为数据来预测其知识状态和学习进度，为教育研究者提供了宝贵的参考资料。作为一种创新的在线教育模式，其在应用过程中也表现出了一些局限性，例如可解释性不足以及学习特征的缺失。针对可解释性问题，本项目在深度学习知识追踪模型 (Deep Knowledge Tracing, DKT)模型的基础上引入了注意力机制，同时结合项 目反应理论(Item Response Theory,IRT)引入了学习情况的相关参数，从而使模 型在决策过程中提供更加清晰、易于理解的解释。本文具体的研究内容如下： 其一，本研究在模型中设计了两种不同的注意力模块——缩放点积注意力 和加性注意力，通过对比实验观察不同注意力机制的性能差异，分析其在处理不同数据集时识别关键信息的能力。实验结果表明，采用了注意力机制的模型相比之前的深度知识追踪模型在各数据集上的性能均有较为明显的提升。 其二，本文考虑了项目反应理论的相关特征参数——学生能力，题目难度及区分度，并将其以嵌入参数的形式引入模型中。为了验证添加参数的有效性 和模型的可解释性，本研究中设计了一系列的消融实验，观察改变特征参数的 个数之后模型性能的变化。消融实验的结果表明，在引入全部参数的时候，模型性能达到最佳。

As educational informatization continues to advance and artificial intelligence technology develops, knowledge tracing models have become a key research area in the era of educational big data. These models can predict a learner's knowledge state and progress by analyzing their behavioral data, providing valuable references for educational researchers. As an innovative online education approach, it has also shown some limitations in application, such as insufficient interpretability and a lack of learning characteristics. To address these issues, this project incorporates an attention mechanism into the deep learning knowledge tracing (DKT) model, and integrates it with the Item Response Theory (IRT) to include parameters related to learning situations. The specific research content of this paper is as follows: Firstly, we have designed two different types of attention modules in the model—scaled dot-product attention and additive attention. By conducting comparative experiments, we observed the performance differences between these attention mechanisms and analyzed their ability to identify key information in different datasets. The experimental results showed that models with attention mechanisms perform significantly better across datasets compared to previous deep knowledge tracing models. Secondly, we considered relevant feature parameters from Item Response Theory—student ability, question difficulty, and discrimination—and integrated them into the model as embedding parameters. To verify the effectiveness of adding parameters and the interpretability of the model, we designed a series of ablation experiments to observe the changes in model performance after varying the number of feature parameters. The results of the ablation experiments indicated that the model performed best when all parameters were introduced.