Federated Learning as a Regularizer: Enhancing Privacy and Equity in Small-Sample Educational Dropout Prediction

Predicting student dropout is a primary objective for educational institutions, yet the majority of advanced machine learning research focuses on large, centralized datasets. This leaves smaller schools and rural districts behind, as they often lack the data volume necessary to train robust deep learning models without overfitting. Furthermore, privacy regulations (e.g., FERPA, GDPR) prevent these smaller institutions from pooling their sensitive records to create larger datasets. This paper explores a novel application of Federated Learning (FL): utilizing it not merely for privacy, but as an effective regularization technique for small-sample educational data. We validate the Federated Explainable AI (FXAI) framework on the UCI Student Performance dataset ( ), a representative small-scale educational dataset. Our results demonstrate that the federated model consistently outperforms a centralized neural network baseline (AUC-ROC 0.6901 vs. 0.6277). This 10% performance gain suggests that the distributed training process acts as a powerful regularizer, preventing the model from memorizing local noise. Additionally, the integration of fairness constraints reduced the Equal Opportunity Difference (EOD) to 0.0 under the evaluated thresholding regime. This study provides empirical evidence that small schools can form "model consortia" to achieve predictive analytics that are more accurate, fair, and private than what they could achieve individually or via centralized pooling. These findings suggest a practical pathway for small schools to deploy advanced learning analytics without centralizing sensitive student data.