The early detection of learners who are expected to disengage with typical MOOC tasks such as watching lecture videos or submitting assignments is necessary to enable timely interventions aimed at preventing it. This can be done by predicting the decrease of academic engagement indicators that can be derived for different MOOC tasks and computed for each learner. A posteriori prediction models can yield a good performance but cannot be built using the information that is available in an ongoing course at the moment the predictions are required. This paper proposes an approach to build in situ prediction models using such information. Models were derived following both approaches and employed to predict the decrease of three indicators that quantify the engagement of learners with the main tasks typically proposed in a MOOC: watching lectures, solving finger exercises, and submitting assignments. The results show that in situ models yielded a good performance for the prediction of all engagement indicators, thus showing the feasibility of the proposed approach. This performance was very similar to that of a posteriori models, which have the clear disadvantage that they cannot be used to make predictions in an ongoing course based on its data.

MOOC, engagement, supervised machine learning