A Federated and Explainable AI Framework for Scalable Learner Performance Prediction in Gamified Educational Platforms
Year: 2026
Type: article
Abstract: Gamified educational platforms have demonstrated moderate positive effects on academic performance (Hedges’ g ≈ 0.50), yet existing approaches suffer from centralized data dependency, black-box predictive models, and the absence of privacy-preserving adaptation. This study proposes a federated and explainable AI framework for learner mastery prediction across simulated distributed institutional nodes. A simulated dataset of 2,500 learner interaction records, generated from a domain-informed parametric model calibrated against published educational statistics, capturing 18 engineered features across knowledge, simulation, engagement, and gamification dimensions, was partitioned across three institutional nodes in a non-IID configuration (mean Jensen–Shannon divergence = 0.019; mean 1-Wasserstein distance = 0.107). The prediction target is defined on an independent summative final examination, y = 1[FinalExamScore ≥ P₇₀], administered after the observation window, so that the label-generating assessment is disjoint from the input feature space. Five supervized classifiers — Logistic Regression, Support Vector Machine, Random Forest, XGBoost, and Artificial Neural Network — were trained under both centralized and federated settings using four classifier-specific federated protocols: Federated Averaging for the linear and neural models, gradient-histogram aggregation for XGBoost, ensemble pooling for Random Forest, and random Fourier feature approximation with Federated Averaging for the SVM. All preprocessing, including SMOTE, is fitted inside each training fold of a nested cross-validation pipeline. The federated XGBoost model achieved 88.4% ± 1.1% accuracy (F1 = 0.870 ± 0.014, AUC-ROC = 0.93 ± 0.010) against 89.6% ± 1.0% under centralized training, a privacy cost of Δ = 1.2% points (paired t-test, p < 0.001). Integrating the Gaussian mechanism with an analytical moments accountant yields formal (ε = 2.0, δ = 10⁻⁵)-differential privacy at a total accuracy cost of 2.8% points, under which membership inference advantage falls to 0.04 and gradient inversion reconstruction error rises to 0.487. A privacy-preserving federated SHAP protocol, in which only 18 aggregate scalars per node are communicated, identifies session frequency (mean |SHAP| = 0.168), module completion rate (0.142), and XP accumulation rate (0.119) as the strongest predictive associations, together accounting for 41.8% of model explanatory power. Supplementary validation on two real public benchmarks (OULAD, 28,785 learners; EdNet, 784,309 learners) reproduces both the federated–centralized gap (Δ = 1.3–1.4%) and the dominance of engagement features. The framework shows consistent performance across simulated institutional nodes (maximum inter-node Δ = 0.8%) and satisfies four formal fairness criteria on the two demographic attributes available. These findings advance the goal of sustainable, privacy-preserving, and equitable educational AI that can scale across institutional and jurisdictional boundaries.
Authors: Aneesha Nigam, Sameer Naugariya, Deepak Thakur
Institution: Lovely Professional University
Language: English
Subfield: Computer Science Applications
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: hybrid
Estimated APC paid: $3,690
OpenAlex ID: W7235884938
