Modeling Students' Metacognitive Errors in Two Intelligent Tutoring Systems

Intelligent tutoring systems help students acquire cognitive skills by tracing students’ knowledge and providing relevant feedback. However, feedback that focuses only on the cognitive level might not be optimal - errors are often the result of inappropriate metacognitive decisions. We have developed two models which detect aspects of student faulty metacognitive behavior: A prescriptive rational model aimed at improving help-seeking behavior, and a descriptive machine-learned model aimed at eliminating attempts to “game” the tutor. In a comparison between the two models we found that while both successfully identify gaming behavior, one is better at characterizing the types of problems students game in, and the other captures a larger variety of faulty behaviors. An analysis of students’ actions in two different tutors suggests that the help-seeking model is domain independent, and that students’ behavior is fairly consistent across classrooms, age groups, domains, and task elements.