练习嵌入和学习遗忘特征增强的知识追踪模型

Exercise embeddings and learning-forgetting features boosted knowledge tracing

现有知识追踪模型大多以概念为中心评估学生的未来表现,忽略了包含相同概念的练习之间的差异,从而影响模型的预测准确性。此外,在构建学生知识状态过程中,现有模型未能充分利用学生在答题过程中的学习遗忘特征,导致对学生知识状态的刻画不够精确。针对以上问题,提出了一种练习嵌入和学习遗忘特征增强的知识追踪模型(exercise embeddings and learning-forgetting features boosted knowledge tracing, ELFBKT)。该模型利用练习概念二部图中的显性关系,深入计算二部图中的隐性关系,构建了一个练习概念异构关系图。为充分利用异构图中的丰富关系信息,ELFBKT模型引入了关系图卷积网络。通过该网络的处理,模型能够增强练习嵌入的质量,并以练习为中心更准确地预测学生的未来表现。此外,ELFBKT充分利用多种学习遗忘特征,构建了两个门控机制,分别针对学生的学习行为和遗忘行为进行建模,更精确地刻画学生的知识状态。在两个真实世界数据集上进行实验,结果表明ELFBKT在知识追踪任务上的性能优于其他模型。

Most existing KT models evaluate students’future performance centered on concepts,overlooking the differences between exercises containing the same concepts,thus affecting the models’prediction accuracy.Moreover,in constructing the students’knowledge state,existing models fail to fully utilize the learning-forgetting features of students during the answering process,leading to an inaccurate modeling of students’knowledge states.To address these issues,this paper proposed an exercise embeddings and learning-forgetting features boosted knowledge tracing model.The model utilized the explicit relationships in the exercise-concept bipartite graph to calculate the implicit relationships within the graph,constructing an exercise-concept relationship heterogeneous graph.To make full use of the rich relationship information in the heterogeneous graph,ELFBKT introduced a relational graph convolutional network(RGCN).Through the processing of RGCN,the model enhanced the quality of exercise embeddings and predicted students’future performance more accurately with an exercise-centric approach.Furthermore,ELFBKT fully utilized various learning-forgetting features to construct two gating-controlled mechanisms,modeling the students’learning and forgetting behaviors respectively,to more accurately model the students’knowledge states.Experiments on two real-world datasets show that ELFBKT outperforms other models in KT tasks.