AdaptFedAvg: Privacy-Preserving Topic Difficulty Modeling in Heterogeneous Smart Classroom Networks

AdaptFedAvg is a privacy-preserving federated learning framework for topic difficulty modeling in heterogeneous smart classroom networks. Educational institutions generate large volumes of student interaction data, but privacy regulations such as GDPR, FERPA, and India’s DPDP Act restrict centralized data collection. This work proposes AdaptFedAvg, an enhanced Federated Averaging approach that improves performance under non-IID client distributions using three mechanisms: local proximal regularization, server-side momentum, and variance-based client quality weighting. The framework also introduces a seven-class student struggle taxonomy that maps struggling learners to targeted pedagogical interventions. Experimental evaluation on a simulated ten-school dataset under mild, medium, and severe non-IID settings shows that AdaptFedAvg converges within 30–40 communication rounds and achieves lower MAE than standard FedAvg, with especially strong robustness under severe heterogeneity. Results demonstrate the effectiveness of adaptive aggregation for privacy-preserving educational analytics.