Answers are grounded in this paper and cite their sources.
Educational risk models are often assessed by aggregate accuracy, although false alarms and leakage in model-selection procedures can materially change their practical value. We present FedPRISM as a reproducible, privacy-aware federated simulation for at-risk screening and preliminary academic guidance, evaluated on SAPEx-D (494 students; 86 encoded predictors). A stratified 80/20 split (395/99) is used; scaling and F-ANOVA selection are fitted within training folds, and a logistic stacking model is trained on five-fold out-of-fold detector scores. The primary input comprises 30 selected non-personality predictors plus all five Big-Five traits (35 dimensions); an ablation uses the 30 non-personality predictors alone. Four approximately equal IID clients are simulated rather than mapped to actual departments. On the held-out set, the stacked model achieves accuracy 0.828, precision 0.825, recall 0.767, F1 = 0.795, false-alarm rate (FAR) = 0.125, and ROC-AUC = 0.889. The Random Forest alone is similar (accuracy 0.818, F1 = 0.775, FAR = 0.107, AUC = 0.894); paired tests do not establish a performance difference. Ablation results indicate that the Random Forest supplies most of the discriminative signal: removing it lowers accuracy to 0.636 and F1 to 0.438. Removing the autoencoder reduces accuracy to 0.816, F1 to 0.781, and ROC-AUC to 0.875, while removing the Isolation Forest yields accuracy of 0.821, F1 of 0.786, and ROC-AUC of 0.881. The ablation analysis did not reveal a statistically significant incremental predictive benefit for the Big-Five personality features examined in the analyzed context, however their contribution might differ depending on the datasets and educational contexts. A separate RDP-accounted Gaussian-noise analysis evaluates client-level privacy for the FedAvg autoencoder component only; it does not cover the forests, stacker, recommendations, or the coordinator's view, and no cryptographic secure aggregation is implemented. The study is a single-institution, retrospective screening experiment: the data do not establish prospective warning time, recommendation uptake, or intervention benefit. These results support further evaluation, not a claim of deployment readiness or end-to-end differential privacy.
No takes yet. Share an insight, caveat, or question.
Aslam et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: