Integrating Question Proficiency Level and Historical Knowledge States for Knowledge Tracing

Knowledge tracing focuses on modeling the progression of students’ knowledge states based on their past response records and predicting their next performance. However, most RNN-based knowledge tracing models struggle to capture long-distance dependencies within sequences, leading to a decline in performance. Besides, most models ignore the personalized question proficiency level of students when solving problems. To tackle these challenges, we propose a new knowledge tracing model that incorporates question proficiency levels and historical knowledge states (QHKT). Firstly, the TCN is employed to respectively extract the students’ answering speed and result features at multiple time steps, which are then fused to obtain the personalized question proficiency level of the students. Subsequently, based on the proficiency level, we utilize LSTM network to acquire the refined knowledge states of the students. To alleviate the long-distance dependency problem, a carefully designed sparse attention mechanism is used to aggregate multiple historical knowledge states to obtain a more precise representation of knowledge states, and predictions are made based on this. We extensively validate the proposed method on two public datasets, and the results demonstrate that QHKT outperforms existing models, achieving better knowledge tracing effect.