Privacy-preserving federated learning for interpretable student at-risk prediction across schools

Area of Science:

• Educational Data Mining
• Machine Learning
• Privacy-Preserving Technologies

Background:

• Early-warning systems for at-risk students often use predictive models trained on sensitive educational data.
• Centralized learning raises privacy, data sovereignty, and auditability concerns, especially when sharing student data across institutions.

Purpose of the Study:

• To present a federated learning framework (FL-AtRisk-DP-PBT) for multi-school at-risk student prediction.
• To integrate Federated Averaging (FedAvg), client-side Differential Privacy (DP), and a Proof-Before-Train (PBT) protocol.
• To evaluate the framework's utility, privacy, and auditability compared to centralized and standard federated approaches.

Main Methods:

• Developed FL-AtRisk-DP-PBT, a federated learning framework using FedAvg, client-side DP, and PBT.
• Employed a single interpretable global logistic-regression classifier.
• Evaluated on three datasets: a primary student cohort (14,003 students in 10 simulated schools), xAPI-Edu-Data, and Students Performance in Exams.

Main Results:

• On the primary dataset, FL-AtRisk-DP-PBT achieved 98.00% accuracy, F1=0.9802, AUC=0.9992, closely matching centralized (99.14% accuracy) and FedAvg (98.61% accuracy) models.
• On xAPI and Exams datasets, FL-AtRisk-DP-PBT reached 93-94% accuracy, F1≈0.92-0.93, and AUC≈0.97-0.98.
• Feature-importance analysis showed FL-AtRisk-DP-PBT preserves similar interpretation patterns to baseline models, with PBT causing minimal metric changes.

Conclusions:

• Interpretable federated at-risk prediction can achieve competitive utility while preserving student data privacy.
• The FL-AtRisk-DP-PBT framework offers privacy-preserving and verifiable training mechanisms.
• Results suggest the feasibility of decentralized, privacy-conscious predictive modeling in education.