Dual-attentional time-aware fusion networks for knowledge tracing

Knowledge tracing (KT) is a crucial technique to predict students’ future performance by observing their historical learning processes. A challenging aspect of this process is that the KT model should be flexible and adaptive to reflect student-specific temporal behaviors and this is also in the case when the available student-specific data are highly sparse and non-uniformly sampled. To address this problem, we proposed a dual-attentional time-aware knowledge tracing model, i.e., DaTaKT to improve the prediction performance of the original self-attentive knowledge tracing model by capturing time-aware patterns. Specifically, our DaTaKT model utilizes a dual-attentional mechanism to capture relations between exercises and student responses from both temporal and question perspectives. Furthermore, we design a discrimination factor to simultaneously represent question-centric information and avoid the data sparsity issues. The proposed time-aware KT model is evaluated on three real-world educational KT datasets with a wide range of deep learning based KT baselines. The results demonstrate the benefits and superior performance of our approach on the prediction tasks for non-uniform interaction sequences. Furthermore, we conduct both ablation studies and quantitative analysis to show the effectiveness of time-related factors and the superior prediction outcomes of DaTaKT.