Federated and Cross-Domain Student Performance Prediction

Accurate student performance prediction is critical for data-driven educational decision-making; however, it is often hindered by data heterogeneity, privacy constraints, and domain shift across academic contexts. This study investigates student final grade (G3) prediction using three complementary machine learning paradigms: centralized learning, cross-domain generalization, and federated learning with personalization. Experiments were conducted on Portuguese and Mathematics student datasets using traditional regression models, ensemble methods, neural networks, and a personalized federated learning framework based on FedProx and FedBN. In the centralized setting, models capable of capturing non-linear relationships, particularly XGBoost and multi-layer perceptrons, achieved superior predictive performance, with XGBoost attaining an R2 of 0.8308 and the lowest error metrics. In contrast, direct cross-domain application of models trained on Portuguese data to Mathematics outcomes resulted in severe performance degradation, with several models yielding negative R2 values, highlighting the adverse impact of domain shift. To address privacy and heterogeneity challenges, a federated learning simulation was implemented. While the global federated model achieved moderate accuracy, the introduction of local personalization led to substantial performance gains. The personalized client models achieved stronger local predictive performance than the global federated model and showed competitive performance relative to centralized baselines. Learning-curve analysis further indicate that model performance in centralized settings improves with increasing data size but eventually plateaus, whereas cross-domain learning remains constrained despite additional data. In federated learning, predictive performance consistently im-proves across training rounds, demonstrating the effectiveness of iterative collaboration and client-level personalization. Overall, the results suggest that federated learning with personalization offers a competitive privacy-preserving alternative to centralized modeling and provides a clear improvement over direct cross-domain transfer in heterogeneous educational analytics.