| Popis: |
Nowadays, Massive Open Online Courses (MOOCs) are adopted by students worldwide. One of the main critical issues often associated with MOOCs is the dropout phenomenon. In other words, the percentage of students abandoning a MOOC-based study path is considered still too high. Therefore, an increasing number of scientific works, coming from several heterogeneous communities (e.g., computer science, data science, statistics, education) propose approaches trying to mitigate such a problem. The majority of the aforementioned works focus on machine learning methods to define classifiers able to be trained and, subsequently, to predict students who are going to abandon a course before it ends. Among such approaches, the ones achieving the best performance use enriched sets of features (to train their models) and produce results that cannot be used to easily clearly characterize the different behaviors of dropping-out and non-dropping-out students. The present work proposes the design of a novel process to train a set of dropout predictors leveraging on a reduced set of features. The underlying idea is to exploit weekly data in order to classify, with acceptable levels of precision, students who are likely going towards dropout or not. In cases of uncertainty, the classification decision is deferred to the next week, when new data is available. Such an approach, which takes care and is aware of the course timeline, offers several advantages. The first one is the chance to build a real-time educational decision support system able to support decision as sufficient information is available (as the time goes on). The second one is to preserve resources and avoiding wasting them with students erroneously classified at risk of dropout. The third one is to allow explicit characterization of dropout-conducing behavior by using a rule mining approach. |
