Assessing student engagement from facial behavior in on-line learning

Area of Science:

• Computer Science
• Educational Technology
• Affective Computing

Background:

• Automatic monitoring of learner engagement in distance education is crucial for personalized support.
• Facial behavior, including action units, gaze, and head poses, offers potential indicators of engagement.
• Existing methods require validation in real-world learning contexts.

Purpose of the Study:

• To investigate the assessment of student engagement levels using facial behavior.
• To propose and evaluate a Long Short-Term Memory (LSTM) network model for engagement prediction.
• To compare model-assessed engagement with students' self-perceived engagement.

Main Methods:

• Utilized the EmotiW 2019 challenge dataset for model training.
• Conducted an experiment with students in an online lecture setting.
• Collected video data of student facial behavior and post-session engagement questionnaires.

Main Results:

• Global engagement prediction from facial behavior showed a weak correlation with subjective self-evaluation.
• Emotional engagement prediction demonstrated a stronger correlation with facial behavior.
• Facial action units and head pose positively correlated with emotional engagement, while gaze showed an inverse correlation.

Conclusions:

• Facial behavior analysis, particularly focusing on emotional dimensions, can offer insights into student engagement.
• LSTM networks show promise for automated engagement assessment in online learning environments.
• Further research is needed to refine models and improve the correlation between objective and subjective engagement measures.