An Improved Deep Model for Knowledge Tracing and Question-Difficulty Discovery

Huan Dai, Yupei Zhang, Yue Yun, Xuequn Shang

Abstract: Knowledge Tracing (KT) aims to analyze a student’s acquisition of skills over time by examining the student’s performance on questions of those skills. In recent years, a recurrent neural network model called deep knowledge tracing (DKT) has been proposed to handle the knowledge tracing task and literature has shown that DKT generally outperforms traditional methods. However, DKT and its variants often lead to oscillation results on a skill’s state may due to it ignoring the skill’s difficulty or the question’s difficulty. As a result, even when a student performs well on a skill, the prediction of that skill’s mastery level decreases instead, and vice versa. This is undesirable and unreasonable because student’s performance is expected to transit gradually over time. In this paper, we propose to learn the knowledge tracing model in a “simple-to-difficult” process, leading to a method of Self-paced Deep Knowledge Tracing (SPDKT). SPDKT learns the difficulty of per question from the student’s responses to optimize the question’s order and smooth the learning process. With mitigating the cause of oscillations, SPDKT has the capability of robustness to the puzzling questions. The experiments on real-world datasets show SPDKT achieves state-of-the-art performance on question response prediction and reaches interesting interpretations in education.