SOD-FE: a supervised outlier detection and feature engineering approach for student dropout prediction

Abstract

Abstract Student dropout in higher education creates academic and socio-economic challenges for institutions and students. Effective early prediction models are essential to identify at-risk students and implement timely interventions. This paper proposes SOD-FE, a supervised machine learning approach that combines label-aware outlier detection with feature engineering to enhance dropout prediction. The approach integrates interquartile range (IQR) based outlier detection with mutual information and Pearson correlation to identify and mitigate the impact of outliers before constructing the final model. Then, a feature selection strategy is applied to refine the dataset. The approach is evaluated through experiments on two real-world datasets (Portugal and Slovakia) utilizing five classification algorithms, including Random Forest (RF) and Extreme Gradient Boosting (XGB). Performance improved markedly with the RF classifier, achieving F1 scores of 98.09% and 98.33% on the Portugal and Slovakia benchmark datasets, respectively, under 5-fold cross-validation. Furthermore, the SOD-FE approach outperformed the baseline RF models by 7.46% and 3.37% in F1 score across the two datasets, demonstrating highly competitive results that align with or exceed the highest reported performances in recent literature by up to 7.33%. The proposed approach also incorporates explainable AI techniques (SHAP) to enhance model transparency and support data-driven educational policy. These findings show the significant potential of the SOD-FE method for improving student retention and early intervention systems in educational institutions.