Student Engagement Prediction in MOOCs Using Deep Learning

The level of peoples' engagement in a certain task is determined using an automated recognition devices (e.g. physiological sensors and pressure-sensing chairs). Even though these tools were being used in previous research works, they were very expensive and intrusive. Presently, the use of RGB video cameras is affordable and has also shown a significant effect in predicting people's engagement in tasks. Statistical tools are providing a strong foundation to model the automatic engagement identification techniques that use video cameras. In this paper, a lightweight MobileNetv2 is used to automatically determine the student engagement in MOOCs for devices with limited resources. All of the layers in the MobileNet V2 architecture have been fine-tuned to improve learning and adaptability. Instead of 1000 classes as in ImageNet, the final layer is adjusted to 3 classes of output at the final classification step. The experimental study is done on open source dataset created by subjects watch videos in online courses. Results from the evaluation phases show that our model performs better than the other two pre-trained networks (ResNet50, Inception V4).