Gaze-Based Prediction of Students' Understanding of Physics Line-Graphs: An Eye-Tracking-Data Based Machine-Learning Approach

Graphs are important formats to measure conceptual knowledge in physics - especially in kinematics. Although there is a broad line of research studying the quality of such tests regarding validity and reliability, little is known about learners' cognitive processes during solving problems with line graphs. In this work, we used a large gaze-data set of 115 high-school students who solved the widely studied Test of Understanding Graphs in Kinematics (TUG-K). We used a supervised machine-learning approach to classify the students' performance using different eye-tracking metrics, specifically the total visit duration and the transitions between areas of interest for one selected item of the TUG-K. Using only five features, the results show that both metrics provide a comparable predictability of students' performance, which is significantly better than a combination of both measures. Surprisingly, only with large data sets, the combination of both metrics achieves the highest predictability. In this sense, the results provide a guide for the selection and combination of features in adaptive learning environments.