Transfer Learning using Representation Learning in Massive Online Open Courses

In MOOCs predictive models of student behavior support many aspects of learning design, including instructor feedback and timely intervention. Ongoing (target) courses, when their predictive outcomes are yet unknown, must rely upon models trained from the historical data of previously offered (source) courses. It is possible to transfer such models but they still often predict poorly. One reason is features that inadequately capture predictive attributes common to both courses. We present an automated transductive transfer learning approach that addresses this issue. It relies upon straight-forward problem-agnostic, temporal organization of the clickstream data, where, for each student, for multiple courses, a set of specific MOOC event types is expressed for each time unit. It consists of two alternative transfer methods based on autoencoders: a passive approach using transductive principal component analysis and an active approach by adding a correlation alignment loss term. We investigate the transferability of dropout prediction across similar and dissimilar MOOCs and compare with known methods. Results show improved model transferability and suggest that the methods are capable of automatically learning a feature representation that expresses common predictive characteristics of MOOCs.