Early dropout prediction via machine learning in professional online courses

Abstract

Despite the advantages of e-learning, this way of learning is prone to dropping out. Previous studies show that machine-learning techniques can be applied to records of interactions between students and the platform to predict abandonment. In this line, this work tries to find predictive dropout models in virtual courses that last between six and sixteen weeks, using Moodle logs from the first two. Models’ sensitivity, specificity and precision were evaluated, but priority was given to the extent to which these models made it easier to avoid attrition through cost-effective retention actions. Specifically, data from several cohorts of four courses with different themes and durations were used. All four dictated by the Secretariat of Extension of the National Technological University of the Argentine Republic, Regional Buenos Aires between February 2018 and October 2019. Different algorithms were used to generate predictive models and optimize them in order to mitigate the economic losses caused by attrition. It was analyzed if any one in particular generated the best models for all courses. It was studied whether it was convenient to build separate models per course or one for the entire data set of the four courses. It was found that it is possible to build successful predictive models and that the algorithm that produced the best models was a neural network in three of the four courses. The model that fit each one separately turned out better.