Knowledge tracing is one of the key research areas for empowering personalized education. It is a task to model student’s knowledge state, that is, the mastery level of a knowledge component, based on their historical learning trajectories. In recent years, a recurrent neural network model called deep knowledge tracing (DKT) has been proposed to handle the knowledge tracing task. Literature has shown that DKT generally outperforms traditional methods. However, through extensive experimentation, we have noticed two major problems, which would mislead the interpretation of student’s knowledge state, in the DKT model. Firstly, the model fails to reconstruct the knowledge state with respect to the observed input, and secondly, the predicted performance of students across time-steps...[ Read more ]

Knowledge tracing is one of the key research areas for empowering personalized education. It is a task to model student’s knowledge state, that is, the mastery level of a knowledge component, based on their historical learning trajectories. In recent years, a recurrent neural network model called deep knowledge tracing (DKT) has been proposed to handle the knowledge tracing task. Literature has shown that DKT generally outperforms traditional methods. However, through extensive experimentation, we have noticed two major problems, which would mislead the interpretation of student’s knowledge state, in the DKT model. Firstly, the model fails to reconstruct the knowledge state with respect to the observed input, and secondly, the predicted performance of students across time-steps is not consistent. In this thesis, we introduce regularization terms that correspond to reconstruction and waviness to the loss function of the original DKT model to enhance the consistency in prediction, and evaluate how the regularized DKT model (DKT+) relieves these two problems. Furthermore, the DKT+ model is employed to build predictive models that predict whether the first job of a student out of college belongs to a STEM (the acronym for science, technology, engineering, and mathematics) field. Experiments show that the DKT+ model effectively alleviates the two problems and improves the prediction accuracy of STEM predictors, compared to the original DKT model.