Eye-Tracking Indicators of Novice Programmers’ Proficiency: A Machine Learning Approach

This study investigates the efficacy of machine learning algorithms in classifying different levels of programming expertise among 60 first-year undergraduate computer science students from Asian demographic backgrounds using eye-tracking data. Existing studies offer limited detail on the construction and selection of feature sets for machine learning modeling. The study identifies Total Fixation Duration (TFD) and Total Visit Duration (TVD) as robust indicators for machine learning models when distinguishing between high and low performers (two levels) achieving accuracy as high as 76%. However, performance declines notably when classifying expertise into three levels (high, average, and low), with performance dropping below 50%, indicating that binary labels yield more reliable predictions than finer-grained categorization. Our findings suggest that such fixation-based metrics can provide real-time insights into student engagement and potentially cognitive effort, offering opportunities for adaptive instruction and targeted support. Hence, this model can be utilized for real-time screening of novice students during programming tasks in classroom settings, allowing educators to identify students requiring additional support, thereby enhancing programming education. Future research should address study limitations by increasing sample size, diversifying participant demographics, and cross-validating model performance with students’ grades.