Abstract
Major discoveries in technology and science often rely on mathematical skills. Mathematical knowledge is founded on basic math problem solving such as addition, subtraction, multiplication, and division. Research shows that problem solving is associated with eye movements that index allocation of attention. Machine learning has been used with eye-tracking metrics to predict performance on real-life user efficiency tasks and classic puzzle games. Critically, no study to date has evaluated eye-tracking metrics associated with mathematical operations using machine learning approaches to classify trial correctness and predict task difficulty level. Participants (n = 26, 20-30 years) viewed mathematical problems in three levels of difficulty indexed by 1-, 2-, and 3-digit problems along with four possible answers, while their eye movements were being recorded. Eye-tracking data were acquired with EyeLink Portable Duo SR Research eye-tracker with 1ms temporal resolution (at 1000 Hz frequency) in remote head-free-to-move mode. Results show that trial correctness can be classified with a 0.81 ROC AUC score based on 5 fold cross-validation. Predicting task difficulty level of each trial was attained with 72% accuracy, which is significantly better than the random prediction (i.e., 50%). The most important features for both machine learning models include metrics associated with current pupil fixation, current saccade amplitudes, and current fixation duration. Theoretically, findings contribute to theories of mathematical cognition. Practically, algorithms can contribute to further research in mathematical problem solving and machine learning, which potentially has applications in education in terms of assessment and personalized learning.