An Unobtrusive Approach to Emotion Detection in E-Learning Systems.

This article focuses on detecting learner emotions in e-learning environments using unobtrusive indicators such as keystrokes, mouse clicks, discussion forum sentiment, and academic performance data. The study involved 50 graduate computer science students in a blended learning course, analyzing 24 selected behavioral and academic attributes to classify emotions into happy, angry, stressed, and neutral categories. Machine learning models, particularly logistic regression after hyperparameter tuning, achieved up to 85% accuracy in predicting emotions, with academic indicators shown to significantly improve prediction performance. The research also identified patterns in emotional changes related to assignment submission timing, assessment periods, resource length, and time of day, highlighting the impact of emotions on learning outcomes. Limitations include a homogeneous participant group and the blended learning context, with future work planned to include diverse learners, purely online settings, and multimodal emotion detection methods.