Attentive or Not? Toward a Machine Learning Approach to Assessing Students’ Visible Engagement in Classroom Instruction

Abstract Teachers must be able to monitor students’ behavior and identify valid cues in order to draw conclusions about students’ actual engagement in learning activities. Teacher training can support (inexperienced) teachers in developing these skills by using videotaped teaching to highlight which indicators should be considered. However, this supposes that (a) valid indicators of students’ engagement in learning are known and (b) work with videos is designed as effectively as possible to reduce the effort involved in manual coding procedures and in examining videos. One avenue for addressing these issues is to utilize the technological advances made in recent years in fields such as machine learning to improve the analysis of classroom videos. Assessing students’ attention-related processes through visible indicators of (dis)engagement in learning might become more effective if automated analyses can be employed. Thus, in the present study, we validated a new manual rating approach and provided a proof of concept for a machine vision-based approach evaluated on pilot classroom recordings of three lessons with university students. The manual rating system was significantly correlated with self-reported cognitive engagement, involvement, and situational interest and predicted performance on a subsequent knowledge test. The machine vision-based approach, which was based on gaze, head pose, and facial expressions, provided good estimations of the manual ratings. Adding a synchrony feature to the automated analysis improved correlations with the manual ratings as well as the prediction of posttest variables. The discussion focuses on challenges and important next steps in bringing the automated analysis of engagement to the classroom.