Using Student Modeling to Estimate Student Knowledge Retention.

The goal of predicting student behavior on the immediate next action has been investigated by researchers for many years. However, a fair question is whether this research question is worth all of the attention it has received. This paper investigates predicting student performance after a delay of 5 to 10 days, to determine whether, and when, the student will retain the material seen. Although this change in focus sounds minor, two aspects make it interesting. First, the factors influencing retention are different than those influencing short-term performance. Specifically, we found that the number of student correct and incorrect responses were not reliable predictors of long-term performance. This result is in contrast to most student-modeling efforts on predicting performance on the next response. Second, we argue that answering the question of whether a student will retain a skill is more useful for guiding decision making of intelligent tutoring systems (ITS) than predicting correctness of next response. We introduce an architecture that identifies two research topics that are meaningful for ITS decision making. Our experiments found one feature in particular that was relevant for student retention: the number of distinct days in which a student practiced a skill. This result provides additional evidence for the spaced practice effect, and suggests our models need to be aware of features known to impact retention.