In this study, we aim to predict student off-task behavior in an Intelligent Tutoring System through the determination of antecedent actions and affective and behavioral states at a finegrained level across four time windows – 10 minutes, 5 minutes, 2 minutes and 20 seconds. Using the data obtained from 5,002,991 interaction logs of 7,647 students’ usage of Assistments system within school year from 2004 to 2009, a Java program is created to set up the clips, distill features and create four datasets for model building. A series of multiple linear regression analyses are run using the four datasets to produce models and statistical measures are employed to select the best-fitting model for each time window. Despite the largescale datasets used, the models significantly achieve moderate yet substantial performance for predicting off-task behavior. Out of 23 predictors, the best-fitting models consistently reveal top five antecedents of off-task behavior: boredom, automatic scaffolding, number of correct answers given on first attempt,?when the student takes more than 80 seconds to respond, and when the attempt takes more than 10 seconds and the next action is right. A closer look at these variables may lead to better understanding on why students disengage from the learning environment and establish a baseline for real-time pedagogical interventions.
展开▼
