时空相关性融合表征的知识追踪模型

知识追踪通过对知识点的表示来描述习题，以此建模知识状态，最终预测学习者的未来表现。然而目前的研究在知识点的表示方面既没有建模历史知识点对当前知识点产生的时间关系上的影响，又未能刻画习题内部各知识点之间产生的空间关系上的作用。为了解决上述问题，提出了时空相关性融合表征的知识追踪模型。首先，以知识点之间的时间相关程度为基础，建模历史知识点对当前知识点的时间作用；其次，利用图注意力网络建模习题所包含的若干知识点之间的空间作用，得到蕴涵了时空信息的知识点表示；最后，利用上述知识点的表示推导出习题的表示，通过自注意力机制得到当前的知识状态。在实验阶段，与五种相关知识追踪模型在四个真实数据集上进行性能对比，结果表明提出的模型在性能方面有更出色的表现。特别地，在ASSISTments2017数据集中所提模型比五个对比模型在AUC、Acc方面分别提升了1.7%~7.7%和7.3%~2.1%；消融实验证明了建模知识点之间时空相关影响的有效性，训练过程实验表明了提出的模型在知识点的表示及其相互作用关系的建模等方面具有一定的优势，应用实例也可看出该模型优于其他知识追踪模型的实际结果。

Knowledge tracing model of temporal and spatial correlation fusion

Knowledge tracing aims to model the state of knowledge and ultimately predict the future performance of learners by describing exercises through the representation of concepts. However, in terms of the representation of concepts, the current research doesn''t model the influence of historical knowledge concepts on the temporal relationship of the current concepts, nor does it describe the role of the spatial relationship between various concepts in the exercise. In order to solve these problems, this paper proposed a knowledge tracing model characterized by temporal and spatial correlation fusion. First of all, based on the degree of temporal correlation between concepts, it modelled the temporal effect of historical concepts from current concepts. Secondly, it modelled the spatial interaction between several concepts contained in the exercise to obtain the representation of knowledge points containing temporal and spatial information through the graph attention network. Finally, it used the above representation of concepts to derive the representation of the exercises, and generated the current state of knowledge through the self-attention mechanism. In the experimental stage, this paper compared the performance of the proposed model with the five relevant knowledge tracing models on four real datasets. The results show that the proposed model has better performance. In particular, compared to the five comparative models on the ASSISTments2017 dataset, the AUC and Acc are improved by 1.7%~7.7% and 7.3%~12.1%, respectively. At the same time, the ablation experiment proves the effectiveness of modeling the temporal and spatial correlation between concepts, and the training process experiment shows that the proposed model has certain advantages in the representation of concepts and the modeling of their interaction relationships. The application examples can also show that the model has better practical results than other knowledge tracing models.