Early detection of students at risk of failure from a small dataset

Abstract

Predicting that a student is likely to fail in a course is critical for performing early interventions, prevent dropout and increase performance on distance learning. This work investigates the most promising machine learning model to perform this task using a small (35 samples) dataset that concerns two classes of one undergraduate course subject. The results bring evidence that the implemented ensemble can perform a prediction at the end of the first week of the course, with a mean accuracy of 78%, when presented to unseen data. This paper also investigates the influence of past data on the results of the classifiers by building datasets with different time window configurations.