Multi-Class Phased Prediction of Academic Performance and Dropout in Higher Education

The application of intelligent systems in the higher education sector is an active field of research, powered by the abundance of available data and by the urgency to define effective, data-driven strategies to overcome students' dropout and improve students' academic performance. This work applies machine learning techniques to develop prediction models that can contribute to the early detection of students at risk of dropping out or not finishing their degree in due time. It also evaluates the best moment for performing the prediction along the student's enrollment year. The models are built on data of undergraduate students from a Polytechnic University in Portugal, enrolled between 2009 and 2017, comprising academic, social-demographic, and macroeconomic information at three different phases during the first academic year of the students. Five machine learning algorithms are used to train prediction models at each phase, and the most relevant features for the top performing models are identified. Results show that the best models use Random Forest, either incorporating strategies to deal with the imbalanced nature of the data or using such strategies at the data level. The best results are obtained at the end of the first semester, when some information about the academic performance after enrollment is already available. The overall results compare fairly with some similar works that address the early prediction of students' dropout or academic performance.