FedEAVE: federated evidence alignment and causal diagnosis for privacy-preserving personalized learning

Reliable educational diagnosis requires evidence from multiple sources without exposing sensitive student records or assuming that recommended interventions are effective. We propose FedEAVE (Federated Evidence-Aligned Verification for Education), a federated framework combining contrastive evidence alignment, prerequisite-aware diagnosis, and retrospective intervention verification. Source-specific encoders map assessment, behavioral, peer-interaction, and temporal signals into a shared space; an availability-aware network weights the observed sources by their reliability. A graph-based diagnostic model estimates concept mastery and counterfactual responses under assumptions of an acyclic prerequisite graph and concept-level unconfoundedness. A matched difference-in-differences (DiD) procedure estimates intervention effects and provides feedback for diagnostic and policy updates. Federated optimization employs per-example clipping ( \(C{=}1.0\) ), Gaussian noise ( \(\sigma _\textrm{DP}{=}1.1\) ), and a reported privacy budget of \((\epsilon {=}2.0,\delta {=}10^{-5})\) over \(T{=}50\) rounds. On EdNet, FedEAVE obtains \(84.7\%\) response-prediction accuracy and an MAE of 0.113; it also improves predictive performance on ASSISTments and Junyi Academy. Intervention estimates remain retrospective and require prospective validation.