HHSKT: A learner-question interactions based heterogeneous graph neural network model for knowledge tracing

Knowledge tracing (KT) has evolved into a crucial component of the online education system with the rapid development of online adaptive learning. A key component of the online education system, knowledge tracing (KT) assesses the state of knowledge by tracing each learner's learning activities. The deep KT model, however, is unable to completely extract the features of the questions and skills due to the heterogeneity of the knowledge structure and the sparsity of the interaction records. The model's capacity to handle diverse data is also restricted by over parameterization. Additionally, rather than focusing solely on a precise fit, Intelligent Tutoring System (ITS) should stress interpretable feedback to the learner. The deep KT approach's item parameters are still unable to give students useful feedback. This paper proposes to trace learner's short-term attentional knowledge based on heterogeneous hierarchical differentiation, named HHSKT. Hierarchical heterogeneous knowledge structures and short-term memory enhancement will be used to model the effects of different interaction sequences on learners. Specifically, knowledge structure features are extracted by constructing a heterogeneous graph-based graph information augmentation component. Question differentiation parameters are derived by transforming the TrueSkill system. Besides, learners' history-related practices are emphasized by windowing attention. Comparing regression-based and deep-based knowledge tracing experiments shows that HHSKT significantly outperforms the state-of-the-art approach on three real -world benchmark datasets (with an average AUC improvement of up to 3%), demonstrating the superiority of the proposed model.