Passively Classifying Student Mood and Performance within Intelligent Tutors

It has been long recognized that successful human tutors are capable of adapting instruction to mitigate barriers (e.g., withdrawal or frustration) to learning during the one-to-one tutoring process. A significant part of the success of human tutors is based on their perception of student affect (e.g., mood or emotions). To at least match the capabilities of human tutors, computer-based intelligent tutoring system (ITS) will need to “perceive” student affect and improve performance by selecting more effective instructional strategies (e.g., feedback). To date, ITS have fallen short in realizing this capability. Much of the existing research models the emotions of virtual characters rather than assessing the affective state of the student. Our goal was to determine the context and importance of student mood in an adaptable ITS model. To enhance our existing model, we evaluated procedural reasoning systems used in virtual characters, and reviewed behavioral and physiological sensing methods and predictive models of affect. Our experiment focused on passive capture of behaviors (e.g., mouse movement) during training to predict the student’s mood. The idea of mood as a constant during training and predictors of performance are also discussed.