A Comparative Evaluation of Federated Learning Algorithms for Privacy-Preserving Academic Prediction on Heterogeneous Data

The rapid growth of educational data enables predictive analytics for academic performance, yet privacy regulations like GDPR and FERPA severely restrict centralized data sharing. Although Federated Learning (FL) has succeeded in privacy-sensitive fields such as healthcare, its application in education remains underexplored, lacking systematic comparative studies of multiple FL algorithms across diverse educational datasets—especially emphasizing recall and ROC-AUC as critical metrics for early identification of students at academic risk. This study fills this gap by evaluating five FL algorithms—Federated Averaging (FedAvg), Federated Proximal (FedProx), Federated Dynamics (FedDyn), Fair Federated Averaging (q-FedAvg), and SCAFFOLD—for privacy-preserving prediction of academic outcomes. Three public datasets were purposefully selected for their representativeness and heterogeneity: Predict Students Dropout and Academic Success (binary dropout prediction with socioeconomic factors), Student Performance (multi-class grade prediction in secondary education), and xAPI-Edu-Data (multi-class performance based on online learning activities). Local neural networks employed Stratified 5-Fold Cross-Validation, while FL algorithms ran for 50 communication rounds. Global models, particularly q-FedAvg and FedProx, consistently surpassed local models, with q-FedAvg achieving 0.7668 accuracy, 0.6813 recall, and 0.8810 ROC-AUC on Predict Students Dropout; 0.8580 accuracy and 0.9871 recall on Student Performance; and 0.7396 accuracy and 0.8815 ROC-AUC on xAPI-Edu-Data. Paired T-tests confirmed significant recall gains for most global models (p < 0.05). These results highlight FL’s ability to handle data heterogeneity and privacy constraints while improving predictive performance, thereby supporting timely educational interventions and enhanced student retention policies.