A SHAP-Inspired Method for Computing Interaction Contribution in Deep Knowledge Tracing

Enrique Valero-Leal, May Kristine Jonson Carlon, Jeffrey S. Cross

Abstract: Deep knowledge tracing (DKT) consists of predicting the probability of correctly answering a test or quiz question using the history of a particular learner’s previous question-answer interactions. The probability of a correct answer is computed using a complex recurrent neural network. In this work, an approach similar to Shapley Additive exPlanations (SHAP) to better understand DKT was used. The number of skills a learner must master to lead to improved learning outcomes in an explainable manner was first reduced. Then, the impact of subsequences rather than every single interaction is studied, as simpler results are expected to be easier to understand. Results help to highlight both subsequences in which the student acquired knowledge and in which its progress stagnated.