Practical Evaluation of Deep Knowledge Tracing Models for use in Learning Platforms

Knowledge tracing (KT) models predict how students will perform on future interactions, given a sequence of prior responses. Modern approaches to KT leverage \textit{deep learning} techniques to produce more accurate predictions, potentially making personalized learning paths more efficacious for learners. Many papers on the topic of KT focus primarily on model performance and do not discuss the practical challenges of implementation. However, understanding the practical aspects of how these models behave is just as important as their predictive performance. Using data from over 500,000 students and over 100 million interactions, we evaluated two deep KT models, a long short-term memory (LSTM) and a self-attentive knowledge tracing (SAKT) model. While global performance metrics for both LSTM and SAKT models are impressive, they also hide important practical flaws. We found significant limitations in their ability to predict responses for new students (i.e., cold start), detect incorrect responses, and maintain sensible predictions independent of question order. Further refinement of the models is needed in these areas in order to enhance their ability to guide a real student's learning path.