AI, governance and federated models for student dropout prediction in education

Educational institutions are using Artificial Intelligence and learning analytics to predict student performance, identify dropout risks, and provide early support to students. These tools improve student outcomes, but institutions usually keep sensitive student data in a centralized location. This raises serious privacy and governance concerns, which can lead to possible non-compliance with the Family Educational Rights and Privacy Act (FERPA), General Data Protection Regulation (GDPR), and Digital Personal Data Protection (DPDP) Act, as well as increased exposure to data breaches. These legal frameworks govern how students’ personal data should be collected, processed, and protected. This work focuses on building accurate early warning AI models across multiple institutions while keeping student records local and aligning with data governance requirements. The methodology compares the performance of centralized models with governance-friendly, federated training using the Open University Learning Analytics Dataset. The best centralized model, a soft-voting ensemble of CatBoost, CNN + MLP, and Transformer + MLP, achieves an at-risk F1 score of 0.80 and an AUC of 0.89. In our re-run with the at-risk class as the positive class, the strongest tree models reach about 0.81 accuracy, 0.81 at-risk F1, and 0.89 AUC under centralized training. A notable finding is that models trained locally on each institution’s own data perform on par with, or slightly better than, centralized training (for example, local-only logistic regression reaches 0.89 AUC), while federated averaging (FedAvg) is the weakest setting (federated logistic regression reaches about 0.75 accuracy, 0.77 F1, and 0.81 AUC). More complex models perform best with centralized or local data but lose ground under federated averaging on this non-IID data. These results make the trade-off clearer: institutions can keep student records local without a meaningful loss in predictive performance, and simple, interpretable models such as logistic regression remain a strong governance-aligned choice.