Knowledge Tracing (KT) has become an important research problem in the educational community with the rise of online learning environments, including intelligent tutoring systems. In this thesis, the two tasks of the KT problem, description and prediction, are identified, and the advantages and disadvantages of existing neural and non-neural KT models are reviewed from the perspectives of those KT tasks. Unfortunately, existing KT models did not excel at both tasks but either showed high prediction performance or provided an intuitive description of students.

The Knowledge Query Networks (KQN) model is proposed to provide an intuitive and descriptive knowledge state of students while achieving the state-of-the-art performance for the prediction task. The architecture of KQN leads...[ Read more ]

Knowledge Tracing (KT) has become an important research problem in the educational community with the rise of online learning environments, including intelligent tutoring systems. In this thesis, the two tasks of the KT problem, description and prediction, are identified, and the advantages and disadvantages of existing neural and non-neural KT models are reviewed from the perspectives of those KT tasks. Unfortunately, existing KT models did not excel at both tasks but either showed high prediction performance or provided an intuitive description of students.

The Knowledge Query Networks (KQN) model is proposed to provide an intuitive and descriptive knowledge state of students while achieving the state-of-the-art performance for the prediction task. The architecture of KQN leads to a distance measure between skills where the distance could be interpreted from the probabilistic perspective. Experiments show that KQN performs better than the previous state-of-the-art KT models for the prediction task while having descriptive states of students. Additionally, through an ablation study, KQN is proven to be stable in learning model parameters.

Additionally, the Controlled Deep Knowledge Tracing (ConDKT) model is proposed in an effort to guarantee consistent probability transitions with the aid of constrained neural networks. Particularly, a novel variant of Long Short-Term Memory is proposed, which controls the movement of hidden states with respect to the input features at the current time step. Through experiments, ConDKT is shown to be comparable to Deep Knowledge Tracing (DKT) for prediction while having explainable network output as opposed to DKT, which has been pointed out by literature for its lack of interpretability.