Designing a Transferable Predictive Model for Online Learning Using a Bayesian Updating Approach

Predictive modeling in online education is a popular topic in learning analytics research and practice. This study proposes a novel predictive modeling method to improve model transferability over time within the same course and across different courses. The research gaps addressed are limited evidence showing whether a predictive model built on historical data retrospectively can be directly applied to a future offering of the same course or to another different course; lacking interpretable data mining models to improve model transferability over time and across courses. Three datasets from two distinct online courses with one course offered two times over two years were applied using direct transferring of the predictive model and the proposed Bayesian updating technique for model transfer. The results showed that the direct transferring of predictive model to the subsequent offering of the course and to a totally different course did not work effectively. By contrast, the proposed Bayesian updating provided a robust and interpretable approach with improved model transferability results for both situations. This Bayesian updating model can be continuously updated with new collected data rather than building prediction model from scratch every time, which can serve as a new methodological framework to carry experience and knowledge from past and other courses forward to new courses.