Pupil Diameter for Student Engagement Classification and Assessing Content Sequence Effect

There has been a growing interest in improving the understanding of student engagement in online learning and to measure it quantitatively. Our work is focused on measuring student engagement through their pupillary responses when listening to online lectures. The pupillary response is involuntary in nature and impossible to fake. It has been hypothesized that continuous self-report is a reliable source of data annotation when cognitive load measurement is not involved. We collected pupillary responses from 40 participants with a Tobii eye tracker when watching online lecture. The lecture was presented to the participants on a flat-screen and divided into 5 segments based on the content. Each segment was further partitioned by minute to identify engagement labels. The engagement label was finalized by measuring the fixation values. Three sequences were used to present the lecture, where one sequence was presented to 20 participants, and the other two sequences were presented to 10 participants each. Our results showed that introducing higher cognitive load negatively affects overall engagement. Additionally, the average pupil diameter increases in the segment when it follows the segment with the complex content. Our BidirectionalLSTM reached the highest test accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 5 \%}$</tex> compared to Random Forest and Multi-layer Perceptron. This shows that our dataset with a size of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 4. 6 6 M}$</tex> has the potential to be used to train classification models to measure student engagement.