FedPRISM: a privacy-preserving federated learning framework for personalized academic recommendation and early intervention in smart education environments
Frontiers in Computer Science · 20.11.2026 · 8
DOI: 10.3389/fcomp.2026.1998673 ↗Kurz gefasst
Fachartikel in der Zeitschrift Frontiers in Computer Science, online erschienen am 20.11.2026. Verfasst von Muhammad Adnan Aslam und 3 weiteren Autor:innen. Die Arbeit nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
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.
Zitieren
APA
Aslam, M. A., Aslam, M. K., Yasin, A., & Ali, Z. (2026). FedPRISM: a privacy-preserving federated learning framework for personalized academic recommendation and early intervention in smart education environments. Frontiers in Computer Science, 8. https://doi.org/10.3389/fcomp.2026.1998673
Vancouver
Aslam MA, Aslam MK, Yasin A, Ali Z. FedPRISM: a privacy-preserving federated learning framework for personalized academic recommendation and early intervention in smart education environments. Frontiers in Computer Science. 2026;8. doi:10.3389/fcomp.2026.1998673
BibTeX
@article{Aslam2026,
author = {Aslam, Muhammad Adnan and Aslam, Muhammad Kamran and Yasin, Amanullah and Ali, Zulfiqar},
title = {FedPRISM: a privacy-preserving federated learning framework for personalized academic recommendation and early intervention in smart education environments},
journal = {Frontiers in Computer Science},
year = {2026},
volume = {8},
doi = {10.3389/fcomp.2026.1998673}
}
EUVIMED