Predicting Student Retention Among a Homogeneous Population Using Data Mining

Student retention is one the biggest challenges facing academic institutions worldwide. In this research, we present a novel data mining approach to predict retention among a homogeneous group of students with similar social and cultural background at an academic institution based in the Middle East. Several researchers have studied retention by focusing on student persistence from one term to another. Our study, on the other hand, builds a predictive model to study retention until graduation. Moreover, our research relies solely on pre-college and college performance data available in the institutional database. We use both standard as well as ensemble algorithms to predict dropouts at an early stage and apply the SMOTE balancing technique to reduce the performance bias of machine learning algorithms. Our study reveals that the Gradient Boosted Trees is a robust algorithm that predicts dropouts with an accuracy of 79.31% and AUC of 88.4% using only pre-enrollment data. The effectiveness of the algorithms further increases with the use of college performance data.