What Does Time Tell? Tracing the Forgetting Curve Using Deep Knowledge Tracing

Recurrent Neural Network (RNN) based Deep Knowledge Tracing (DKT) can extract a complex representation of student knowledge just using the historical time series of correct-incorrect responses given as input and can predict the student's performance on the next problem. funtoot is a personalized and adaptive learning system used by students to practice problems in school and at home. Our analysis of students' interaction with funtoot showed a time-gap as high as 1 h, 1 day and also 1 week between two problems attempted by a student in a task. In this work, along with the time series of previous correct-incorrect responses, we also encode the time-gap as a feature to investigate its effect on predictions. We call this variant of DKT as DKT-t. We test these models on our dataset and two major publicly available datasets from - Assistments and Carnegie Learning's Cognitive Tutor and analyze the predicted student knowledge by both the models and report our findings. We also show that DKT-t can help us trace the forgetting curve given various response sequences and their knowledge states.