Gaze-based prediction of test performance using machine learning for adaptive learning systems

Prior research has demonstrated that students’ performance on physics test items can be accurately predicted using machine learning algorithms based on their gaze behavior. These gaze data are typically recorded during item completion and capture students’ visual attention to both the verbal item stem and accompanying visual representations, such as graphs. This predictive capability highlights the potential of machine learning as a tool for the development of AI-based adaptive learning systems, such as intelligent textbooks, for improving personalized learning. With this study, we aim to replicate the findings on the predictive power of gaze data recorded during test completion in the field of kinematics and extend them by examining whether gaze data collected during a preceding learning phase can similarly predict performance on the test items. The study involved 88 high school students who first engaged with instructional material on kinematics and then completed test items from the test of understanding graphs in kinematics, a test aimed at assessing students’ conceptual understanding of kinematic graphs. Eye-tracking data were recorded both while students engaged with the learning materials and while they worked on the test items presented after each slide. To analyze these data, we used a support vector machine to predict whether students would answer the test items correctly. We compared prediction accuracies based on gaze data recorded during the learning phase with those based on gaze data recorded during the test phase. The findings indicate that eye-tracking data gathered during the completion of test items were more accurate in predicting students’ performance than data collected during the learning phases. Notably, paying close attention to graphs was one of the key behaviors that distinguished students who answered correctly from those who did not. Students’ gaze behavior on the graph was also most predictive of students’ performance. These findings highlight the potential of gaze-based adaptive systems and the importance of including graphical representations in personalized learning environments. They also suggest that future research should explore how intelligent textbooks can combine visual representations, such as graphs, with short conceptual questions to optimize real-time adaptive support.