Evaluating the Feasibility of Learning Student Models from Data

Human teachers are expected to improve with experience; they should learn how to handle different types of students, and learn which methods work for which learning difficulties. But how can a computerized teacher improve its skills over time after working with hundreds of students? How can it recognize that certain problems are too difficult for a student or identify which hints should follow other hints? How can it become more skillful over time? Our research is directed toward providing answers to these questions, which may be divided into two distinct issues: 1) creation of student models from student log file data, and 2) creation of optimal policies that query the student model to determine appropriate responses to maximize student learning. This paper is an initial step toward addressing the first component: creating student models from behavioral data. The worth and difficulties of using student log file data to improve tutoring systems has been discussed (Heiner, Beck, & Mostow 2004). Some ideas have been put forth about how to improve the model of the domain by taking into account students’ behavior (Koedinger & Mathan 2004) and also on how to learn Bayesian networks to discover links between behavior and motivation/attitudes/goals (Arroyo & Woolf 2004; Conati, Gertner, & VanLehn 2002). Beck & Woolf (Beck & Woolf 2000a; 2000b; Beck, Woolf, & Beal 2000) demonstrated how past student data can be used to form a population student model. Some of these past stu-