Domain adaptation based topic modeling techniques for engagement estimation in the wild

In recent years, student engagement estimation has gained focus in the affective computing community. The absence of student monitoring during online MOOC courses makes it challenging to estimate behavioural student engagement during online classes. The non availability of consistent engagement datasets makes it difficult to build cross data automatic behavioural engagement estimation technique. In this paper, we propose an unsupervised topic modeling technique for engagement detection as it captures multiple behavioral cues which are indicators of engagement level such as eye gaze, head movement, facial expression and body posture. We have addressed the various challenges such as less volume of our datasets, large decision unit (annotated for 5 minutes duration) and uneven distribution of different engagement categories with domain adaptation based solution for cross data implementation. We present results on engagement prediction using different clustering techniques such as K-Means and Latent Dirichlet Allocation (LDA) along with different regressors and neural network based attention mechanisms.