Knowledge Interaction Enhanced Knowledge Tracing for Learner Performance Prediction
Knowledge Interaction Enhanced Knowledge Tracing for Learner Performance Prediction
复制标题
DOI：
10.1109/besc51023.2020.9348285
复制
发表时间：
2020-11
期刊：
影响因子：
--
通讯作者：
Wenbin Gan;Yuan Sun;Yi Sun
中科院分区：
文献类型：
--
作者：
Wenbin Gan;Yuan Sun;Yi Sun
在在线学习系统中为学习者提供个性化辅导服务的基本任务之一是预测学习者在未来练习中的表现。为了实现这一目标，有必要通过建模学习者的学习表现来估计和跟踪学习者的知识熟练度（KP）。现有的模型要么无法捕获练习序列中的长期依赖关系，以模拟先前练习对当前练习的影响，要么很难解释结果。为了解决这些问题，本文提出了一个新的模型，称为知识交互增强知识跟踪（KIKT），以估计和跟踪学习者的KP的演变。我们首先提出了一个框架，通过统一记忆网络的强度来增强知识状态的表征和项目反应理论的可解释性来解释学习者的表现。在此框架中，我们跟踪每个学习者对每个知识概念的KP，并使用两种神经网络进一步推断他们的熟练程度和项目特征。此外，我们在模型中加入了知识交互和认知难度，进一步挖掘了练习序列中的长期依赖关系和自适应项目难度。在五个真实数据集上进行的大量实验证明了我们模型的优越性。
One of the fundamental tasks when providing personalized tutoring services to learners in online learning systems is to predict learner performance on future exercises. To achieve this goal, it is necessary to estimate and trace the knowledge proficiency (KP) of learners by modeling their learning performance. The existing models either fail to capture the long-term dependencies in the exercising sequence to model the influence of a previous exercise to the current one or find it difficult to explain the results. To solve these issues, we propose herein a novel model, called the knowledge interaction-enhanced knowledge tracing (KIKT), to estimate and trace the evolution of learners' KP. We first propose a framework by unifying the strength of the memory network to enhance the representation of the knowledge state and the interpretability of the Item Response Theory to explain learner performance. In this framework, we trace each learner's KP on each knowledge concept overtime, and further infer their proficiencies and the item characteristics using two kinds of neural networks. Moreover, we incorporate the knowledge interaction and the cognitive difficulty into our model to further exploit the long-term dependencies and the adaptive item difficulty in the exercising sequences. Extensive experiments conducted on five real-world datasets demonstrate the superiority of our model.