A Fairness-Centered Comparison of Machine Learning Models for Student Retention

Machine learning (ML) is increasingly used to support decision-making in high-stakes domains such as healthcare, finance, and criminal justice. Particularly, in education, ML applications have grown substantially, ranging from personalized learning systems to early warning systems for student success. These applications have demonstrated measurable improvements in intervention timing and resource allocation. While ML offers a wide range of benefits in education, it also risks reinforcing societal biases. In this study, we leverage different ML algorithms to predict students at risk of dropping out, hence enabling timely interventions. We further assess the fairness of these models using different group fairness notions such as demographic parity, equality of opportunity, false positive rate and accuracy parity. Our main contribution is to assess which notion of fairness is best suited for the education domain. This research aims to align the potential of ML with ethical considerations, contributing to fairer and more effective educational interventions. Keywords: Machine Learning, Student Dropout Prediction, Fairness, Education.