融合异构图神经网络的时间卷积知识追踪方法

Temporal Convolutional Knowledge Tracing Method with Heterogeneous Graph Neural Network

知识追踪任务旨在通过建模学生历史学习序列追踪学生认知水平,进而预测学生未来的答题表现.该文提出一个融合异构图神经网络的时间卷积知识追踪模型(Temporal Convolutional Knowledge Tracing Model with Heterogeneous Graph Neural Network,HG-TCKT),将知识追踪任务重述为基于异构图神经网络的时序边分类问题.具体来说,首先将学习记录构建成包含3种节点类型(学生,习题和技能),2种边类型(学生-习题和习题-技能)的异构图数据,异构图描述了学生交互记录中实体类型之间的丰富关系,使用异构图神经网络缓解交互稀疏的问题,引入异构互注意力机制捕捉不同类型节点间的交互关系,提取不同类型节点的高阶特征.将学生节点和习题节点表征拼接,构造边(学生-习题)的表征.最后,使用时间卷积网络捕捉学生历史交互序列的时序依赖关系从而进行预测.在2个真实教育数据集进行实验证明,HG-TCKT相比当前主流知识追踪方法有更好的预测效果.

The knowledge tracking task aims to track students′cognitive levels and predict their future performance by modeling their historical learning sequences.The paper proposes a Temporal Convolutional Knowledge Tracing Model with Heterogeneous Graph Neural Network(HG-TCKT),which rephrases the knowledge tracking task as a temporal edge classification problem based on heterogeneous graph neural networks.Specifically,the learning records are first constructed into heterogeneous graph data containing three types of nodes(students,exercises,and skills)and two types of edges(student-exercise and exercise-skill).The heterogeneous graph describes the rich relationships between entity types in student interaction records,uses heterogeneous graph neural networks to alleviate the problem of sparse interactions,introduces heterogeneous mutual attention mechanisms to capture the interaction relationships between different types of nodes,and extracts high-order features of different types of nodes.The representations of student nodes and exercise nodes are concatenated to construct the representation of the edge(student-exercise).Finally,a temporal convolutional network is used to capture the temporal dependence relationship of students′historical interaction sequences for prediction.HG-TCKT has better prediction performance than current mainstream knowledge tracking methods in two real educational datasets.