Bias-Aware Machine Learning for Student Dropout Prediction: Balancing Accuracy and Fairness

Student dropout remains a persistent global challenge with serious social and economic consequences. Early identification of at-risk learners enables timely support, which can improve retention while promoting fairness in educational outcomes. This study presents a bias-aware machine learning framework for student dropout prediction that jointly evaluates predictive performance and fairness across demographic subgroups. Six machine learning models are benchmarked using academic, demographic, and socioeconomic features. Model performance is assessed using Accuracy, F1-score, Precision, and Matthews Correlation Coefficient, while fairness is evaluated across gender, marital status, and displacement groups. Initial results show that CatBoost achieves the strongest overall performance before class balancing; however, subgroup analysis reveals systematic disparities affecting vulnerable populations. To address these biases, the Synthetic Minority Oversampling Technique is applied. After rebalancing, XGBoost delivers the best performance, achieving substantial improvements in predictive accuracy alongside marked reductions in subgroup disparities. In particular, dropout detection for displaced students improves significantly, narrowing fairness gaps across all evaluated groups. The findings demonstrate that data-level bias mitigation can enhance both accuracy and equity in educational predictive systems. This work provides empirical evidence that fairness-aware machine learning can support more reliable and inclusive early warning systems for student retention.