Estimating Student Knowledge from Paired Interaction Data

Estimating students’ knowledge based on their interactions with computer-based tutors has the potential to improve learning by decreasing time taking assessments and facilitating personalized interventions. Although there exist good student models for relatively structured topics and tutors, less progress has been made with more open-ended activities. Further, students often complete activities in pairs rather than individually, with no coding to indicate who performed each action. We investigate whether pair interactions with an open-ended chemistry tutor can be used to predict individual student post test performance. Using L1-regularized regression, we show that student interactions with the tutor are predictive both of the average post-test score for the pair and of individual scores. Towards better understanding pair dynamics in this setting, we also find that for pairs composed of students with similar pre-test scores, we can predict the difference in students’ post-test scores.