Generating actionable predictions regarding MOOC learners’ engagement in peer reviews

Peer review is one approach to facilitate formative feedback exchange in MOOCs; however, it is often undermined by low participation. To support effective implementation of peer reviews in MOOCs, this research work proposes several predictive models to accurately classify learners according to their expected engagement levels in an upcoming peer-review activity, which offers various pedagogical utilities (e.g. improving peer reviews and collaborative learning activities). Two approaches were used for training the models: in situ learning (in which an engagement indicator available at the time of the predictions is used as a proxy label to train a model within the same course) and transfer across courses (in which a model is trained using labels obtained from past course data). These techniques allowed producing predictions that are actionable by the instructor while the course still continues, which is not possible with post-hoc approaches requiring the use of true labels. According to the results, both transfer across courses and in situ learning approaches have produced predictions that were actionable yet as accurate as those obtained with cross validation, suggesting that they deserve further attention to create impact in MOOCs with real-world interventions. Potential pedagogical uses of the predictions were illustrated with several examples.