Centralized vs. Federated Learning for Educational Data Mining: A Comparative Study on Student Performance Prediction with SAEB Microdata

Rodrigo Tertulino

REVIEW 3 major objections 6 minor 4 references

Centralized vs. Federated Learning for Educational Data Mining: A Comparative Study on Student Performance Prediction with SAEB Microdata

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read On two million Brazilian student records, a federated model trained without centralizing data reaches 61.23% accuracy versus 63.96% for centralized XGBoost, a 2.73-point privacy cost.

desk verdict A new empirical dataset result with a confounded comparison; the 2.7-point 'privacy cost' is not supported by the experiment as run. read the letter →

arxiv 2509.00086 v1 pith:UTOVIN3I submitted 2025-08-27 cs.LG cs.CY

Rodrigo Tertulino This is my paper

classification cs.LGcs.CY

keywords FederatedLearningFedProxEducationalDataMiningStudentPerformancePredictionSAEBPrivacyLGPDXGBoost

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that privacy-preserving federated learning can be almost as accurate as pooling all student data. Using microdata from Brazil's national basic-education assessment (over two million records), it trains a deep neural network with FedProx across 50 simulated schools, never moving raw records to a central server. The federated model peaks at 61.23% test accuracy, while a centralized XGBoost model reaches 63.96%. The paper reads the 2.73-point difference as the price of legal compliance under Brazil's general data-protection law, and argues this trade-off is acceptable because the realistic alternative is often no predictive model at all.

What carries the argument

The load-bearing mechanism is the FedProx proximal term added to each client's local loss: h_k(w)=F_k(w)+(mu/2)||w-w_t||^2. The term penalizes any school's local model for drifting too far from the global model parameters, which stabilizes training when data distributions differ across schools. Around that mechanism, the experiment uses a median-binarized mathematics-proficiency target, 54 one-hot encoded socioeconomic features, and 50 school clients with at least 20 records each, all orchestrated through federated rounds.

What would settle it

Train the same DNN architecture centrally on the same preprocessed SAEB split. If centralized DNN accuracy matches centralized XGBoost (near 64%), the federated gap is not mostly about privacy; if it matches the federated DNN (near 61%), then federated training itself carries little cost. Repeating the federated experiment with FedAvg instead of FedProx under identical hyperparameters would further isolate the proximal term.

Watch

Extended reading notes

Core claim

The central claim is that on the SAEB dataset the performance-privacy trade-off for educational data mining is small enough to make federated learning a practical tool. A DNN trained with FedProx across 50 school clients reached a peak accuracy of 61.23% (60.32% after the final round) with an AUC of 0.6796, while a centralized XGBoost model on the same preprocessed data reached 63.96%. The paper interprets the gap as the 'cost of privacy' and concludes that collaborative prediction is feasible without centralizing sensitive student records, aligning with LGPD.

Load-bearing premise

The comparison assumes the 2.73-point accuracy difference is caused by the privacy-preserving federated scheme, although the two models being compared also have different architectures.

Editorial extensions

If this is right

Schools can build shared predictive models without transferring student-level records, making analytics feasible under the LGPD.
The 2.73-point gap gives institutions a concrete benchmark for what they sacrifice for privacy, useful for deciding when federated learning is acceptable.
The 61-64% accuracy ceiling implies socioeconomic questionnaire data carries genuine but limited signal; gains will likely require richer or longitudinal data.
Federated models can be used for school-level risk profiles and resource allocation rather than individual high-stakes decisions.
If the result transfers to other jurisdictions, decentralized student data across institutions can be mined without a centralized data lake.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The reported cost of privacy is not isolated: the centralized and federated models differ in architecture as well as training paradigm, so part of the 2.73-point gap may come from using XGBoost versus a two-layer DNN. Training the same DNN centrally would separate those causes.
Because SAEB microdata are publicly available, the privacy guarantee in the experiment is simulated; production use would still need secure aggregation or differential privacy to defend against attacks that recover training data from model updates.
The median split of mathematics proficiency discards information about how far above or below the median a student sits; a regression or multi-class target might change the measured privacy cost.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know: this is a legitimate new application—FedProx-trained DNN vs centralized XGBoost on SAEB microdata—but the central claim is not supported by the design. The paper's own numbers don't allow the 'cost of privacy' interpretation.

On the plus side, it's a straightforward, reproducible study: public dataset, two clear pipelines, code in the appendix, metrics reported per round. The preprocessing is sensible (median split, mode imputation, one-hot). It's a reasonable first benchmark for this specific dataset, and the literature review is honest about the gap.

The big problem: the comparison confounds model family and training population. XGBoost sees all ~2M records; the DNN only sees a random sample of 50 schools with a minimum of 20 records each. So the 2.73-point gap could come from architecture, data size or heterogeneity, or the FL mechanism—you can't separate them. There's no centrally trained DNN baseline. The abstract and discussion attribute the gap to privacy, and that is the load-bearing inference. It doesn't hold.

Also: the reported 61.23% is the peak accuracy at round 15, not the final model (60.32% is stated in the RQ2 answer). Using the best-round number in the headline inflates the result. There's a methods inconsistency as well—Section 4.3.2 says 10 rounds and 5 local epochs, while Section 5.2 and Listing 2 say 20 rounds and 10 epochs. Minor but still: no variance or significance testing across runs, and calling the scheme a 'robust privacy guarantee' is an overclaim—FL alone doesn't guarantee privacy without differential privacy or secure aggregation.

These soft spots are central, not cosmetic. The paper is still worth a referee's time because the question matters and the study is reproducible, but it needs major revision: add a same-architecture central baseline trained on the same 50 schools, report final-round accuracy, fix the inconsistency, and temper the privacy language. I'd send it to review conditionally, not desk-reject it.

Referee Report

3 major / 6 minor

Summary. The paper reports a comparative study of centralized vs. federated learning for predicting student performance on Brazilian SAEB microdata. A centralized XGBoost model is trained on roughly 80% of 2,087,904 records and achieves 63.96% accuracy. A DNN trained with FedProx on a simulated federation of 50 schools reaches a peak accuracy of 61.23% during training, with a final accuracy of 60.32% and AUC 0.6796. The authors interpret the 2.73 percentage-point gap between centralized accuracy and federated peak accuracy as a small 'cost of privacy' and conclude that federated learning is a viable LGPD-compliant approach for Brazilian educational data mining.

Significance. If the comparison were internally valid, the paper would provide a useful empirical data point: applying FedProx to a large, real, public educational dataset and quantifying the accuracy trade-off against a strong centralized baseline. The use of a public dataset and the inclusion of condensed source code are strengths. However, the headline quantitative claim is not supported by the experimental design. The centralized and federated pipelines differ in model family (XGBoost vs DNN), in training population (all schools vs a random 50-school subset), and in how the reported accuracy is selected (final model vs best training round). These confounds mean the 2.73 pp gap cannot be attributed to the privacy-preserving mechanism. The paper needs additional baselines and corrected reporting before the central claim can be evaluated.

major comments (3)

[§4.3.1, §4.3.2, §6.1] The central comparison is confounded. The centralized model is XGBoost trained on 80% of all 2,087,904 records, while the federated model is a DNN trained only on a random sample of 50 schools with a minimum of 20 records per client. The 2.73 percentage-point gap reported in §6.1 is therefore the combined effect of model architecture, training data subset, and federated training, not simply the 'cost of privacy' as claimed. A centrally trained DNN on the same 50-school partition, or a federated version of the same XGBoost pipeline, is required to isolate the privacy mechanism.
[Abstract, §5.2, §6.2] The headline accuracy of 61.23% is the peak accuracy at round 15, not the final model. RQ2 in §6.2 reports the final accuracy as 60.32%. Using the best round from the training curve overstates the deployable model and is a form of selection bias. The abstract and RQ3 should either use the final-model accuracy or justify why the peak is the relevant quantity. In addition, the paper reports no variance or significance testing (e.g., multiple seeds, confidence intervals), so a 2–3 percentage-point difference may be within noise.
[§4.3.2 vs §5.2 and Listing 2] The experimental configuration is internally inconsistent. §4.3.2 states the federated analysis was run for 10 rounds with 5 local epochs per client, while §5.2 and Listing 2 use 20 rounds and 10 local epochs. This discrepancy prevents reproducibility and raises doubt about which configuration generated the reported results. The authors should correct the text and ensure the algorithm description matches the code.

minor comments (6)

[Section 2 heading] The heading 'Reletad Works' should be 'Related Works'.
[§3.2] The text says federated learning was 'originally developed by researchers at Google in 2026'; the correct year is 2016, and the cited reference (Hudaib et al. 2025) is a review, not the original source.
[§5.2] The federated model section reports F1, precision, and recall at different rounds rather than for the final model. It would be clearer to present a single confusion matrix and classification report for the final federated model, analogous to Figure 4 for the centralized model.
[Appendix, Listing 2] The code uses a nonstandard API call 'fl.analysis.start_analysis' and the 'forward' pass is omitted. As printed, the code is not directly runnable, which weakens the reproducibility claim.
[Section 6.4] The policy claim that the model can 'generate school risk profiles' is not evaluated in the paper. The experiments only predict individual student classes; school-level aggregation is not assessed.
[Section 1] The sentence 'The present study addresses this gap ...' appears twice in consecutive paragraphs.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported accuracies are measured on a public test set and the comparison, while confounded, does not reduce to a fitted input or self-citation.

full rationale

The paper's central comparison is empirical: XGBoost is trained on an 80/20 split of 2,087,904 SAEB records and evaluated on the held-out test set (Section 4.3.1, Listing 1); the FedProx DNN is trained on a 50-school partition and evaluated on a separate global test set each round (Section 4.3.2, Listing 2). The 63.96% and 61.23% figures are measured values, not quantities defined by the model's inputs. The 2.73 pp 'cost of privacy' is simply the arithmetic difference between these two measured accuracies. No parameter is fitted to the test set and then renamed as a prediction; the abstract explicitly labels 61.23% as 'peak accuracy', and RQ2 also reports the final-round accuracy (60.32%). The paper invokes no self-citations, no uniqueness theorem, and no ansatz smuggled through a citation. The main validity threat—that the gap may reflect the XGBoost-vs-DNN architecture or the smaller 50-school training population rather than the FL mechanism—is a confound in the experimental design, not a circularity. Likewise, selecting the best round from the training curve is selective reporting, not a derivation that reduces to its inputs. Hence no circular step meets the evidentiary bar.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new theory or entities; it uses existing algorithms and a public dataset. The central claim depends on a set of hand-chosen hyperparameters and on the assumption that the FL simulation faithfully represents a real multi-school setting.

free parameters (6)

FedProx proximal parameter mu = 0.1
Set by hand to control local drift; not tuned. It directly influences the federated update rule (Eq. 4).
Number of federated rounds = 20 (results/code) vs 10 (method text)
The paper reports peak accuracy at round 15, but the method section states 10 rounds; the discrepancy affects convergence.
Local epochs per client = 10 (code) vs 5 (method text)
Inconsistent reporting; affects local model quality and the final global model.
DNN hidden layer sizes = 64 and 32
Chosen without a search; model capacity influences the federated accuracy.
Random sample of 50 schools = 50
Randomly sampled from schools with at least 20 records; no repetition or justification.
Train/test split = 80/20 stratified
Standard split; no validation set is used to select hyperparameters, and the global test definition for FL is not fully specified.

assumptions (5)

domain assumption The SAEB microdata accurately represents the target population and the selected 11 features are reliable predictors.
The paper treats public data as ground truth without validating feature quality or measurement noise.
domain assumption Federated learning simulation under Flower faithfully models a real multi-school deployment.
No real distributed infrastructure is used; all clients are partitions of a single dataset on one machine.
ad hoc to paper Binarizing the target at the median yields a meaningful below/above-average classification task.
This choice creates a balanced target but discards proficiency magnitude and is not independently benchmarked.
domain assumption Keeping raw data local is sufficient to guarantee LGPD compliance.
The paper assumes that not centralizing data satisfies the privacy law, without discussing model-update leakage or other LGPD obligations.
domain assumption Accuracy is the primary performance metric for the comparison.
The authors rely on accuracy (and selected secondary metrics) without reporting cost-sensitive or significance analysis.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Centralized vs. Federated Learning for Educational Data Mining: A Comparative Study on Student Performance Prediction with SAEB Microdata." pith.science (2026). https://pith.science/paper/UTOVIN3I

@misc{pith2026250900086,
  author       = {Pith},
  title        = {Pith review of: Centralized vs. Federated Learning for Educational Data Mining: A Comparative Study on Student Performance Prediction with SAEB Microdata},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UTOVIN3I}},
  note         = {Machine review of arXiv:2509.00086}
}

read the original abstract

The application of data mining and artificial intelligence in education offers unprecedented potential for personalizing learning and early identification of at-risk students. However, the practical use of these techniques faces a significant barrier in privacy legislation, such as Brazil's General Data Protection Law (LGPD), which restricts the centralization of sensitive student data. To resolve this challenge, privacy-preserving computational approaches are required. The present study evaluates the feasibility and effectiveness of Federated Learning, specifically the FedProx algorithm, to predict student performance using microdata from the Brazilian Basic Education Assessment System (SAEB). A Deep Neural Network (DNN) model was trained in a federated manner, simulating a scenario with 50 schools, and its performance was rigorously benchmarked against a centralized eXtreme Gradient Boosting (XGBoost) model. The analysis, conducted on a universe of over two million student records, revealed that the centralized model achieved an accuracy of 63.96%. Remarkably, the federated model reached a peak accuracy of 61.23%, demonstrating a marginal performance loss in exchange for a robust privacy guarantee. The results indicate that Federated Learning is a viable and effective solution for building collaborative predictive models in the Brazilian educational context, in alignment with the requirements of the LGPD.

Figures

Figures reproduced from arXiv: 2509.00086 by the authors.

**Figure 1.** Centralized ML Workflow Obj(t) = Xn i=1 l(yi , yˆ (t) i ) +X t k=1 Ω(fk) (1) Where l(yi , yˆ (t) i ) is the loss function that measures the discrepancy between the true label yi and the prediction yˆ (t) i for the i-th instance. The term PΩ(fk) is a regularization component that penalizes the complexity of the models. For decision trees, the complexity is defined as: Ω(f) = γT + 1 2 λ||ω||2 (2) Here, T is the number… view at source ↗

**Figure 2.** Federated Learning (FedProx) Workflow 1. Initialization: The server initializes a global model and sends its parameters (weights) to a selection of clients. 2. Local Training: Each selected client trains the received model on local data for a few iterations. 3. Update Communication: The clients send only their updated model parameters, not the data, back to the server. 4. Secure Aggregation: The server aggregates th… view at source ↗

**Figure 3.** The data preprocessing pipeline, from raw data selection to the final model-ready dataset [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Confusion matrix for the centralized XGBoost model on the test set [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

**Figure 5.** Figure 5: Top 15 most important features identified by the centralized XGBoost model [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

**Figure 6.** Figure 6: Global performance metrics of the federated DNN model over 20 communication rounds, [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

**Figure 7.** Figure 7: Receiver Operating Characteristic (ROC) curve for the final federated DNN model, evaluated on [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

4 extracted references · 3 canonical work pages

[1]

Alhazeem, E., Alsobeh, A., & Al-Ahmad, B. (2024). Enhancing software engineering ed- ucation through ai: An empirical study of tree-based machine learning for defect prediction. Proceedings of the 25th Annual Conference on Information Technology Education, 153–156. https://doi.org/10.1145/3686852.3686881 An, T., Ma, L., Wang, W., Yang, Y., Wang, J., & Che...

work page arXiv 2024
[67]

L., & Wu, Z

https: //doi.org/10.3390/mti7060067 Lin, H., Wen, X., Ye, G. L., & Wu, Z. (2023). Relationship between governance structure and performance based on the application of xgboost algorithm.Proceedings of the 2022 4th International Conference on Big-Data Service and Intelligent Computation, 73–79. https://doi.org/10.1145/3578339.3578352 Marcolino, M. R., Port...

work page arXiv 2023
[495]

A., Khuwaja, P., Trivedi, H., & Pervaiz, K

https://doi.org/10.1007/s43621-025-01388-2 Khowaja, S. A., Khuwaja, P., Trivedi, H., & Pervaiz, K. (2023). Federated learning for healthcare: A survey.Multidisciplinary Digital Publishing Institute, 7(6),

work page doi:10.1007/s43621-025-01388-2 2023
[9840]

https://doi.org/10.1038/s41598-025-93918-1 Mastour, H., Dehghani, T., Moradi, E., & Eslami, S. (2025). Explainable artificial intelli- gence for predicting medical students’ performance in comprehensive assessments. Scientific Reports, 15(1), 23752. https://doi.org/10.1038/s41598-025-07460-1 McMahan, H. B., Moore, E., Ramage, D., & y Arcas, B. A. (2016). ...

arXiv 2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Alhazeem, E., Alsobeh, A., & Al-Ahmad, B. (2024). Enhancing software engineering ed- ucation through ai: An empirical study of tree-based machine learning for defect prediction. Proceedings of the 25th Annual Conference on Information Technology Education, 153–156. https://doi.org/10.1145/3686852.3686881 An, T., Ma, L., Wang, W., Yang, Y., Wang, J., & Che...

work page arXiv 2024

[2] [67]

L., & Wu, Z

https: //doi.org/10.3390/mti7060067 Lin, H., Wen, X., Ye, G. L., & Wu, Z. (2023). Relationship between governance structure and performance based on the application of xgboost algorithm.Proceedings of the 2022 4th International Conference on Big-Data Service and Intelligent Computation, 73–79. https://doi.org/10.1145/3578339.3578352 Marcolino, M. R., Port...

work page arXiv 2023

[3] [495]

A., Khuwaja, P., Trivedi, H., & Pervaiz, K

https://doi.org/10.1007/s43621-025-01388-2 Khowaja, S. A., Khuwaja, P., Trivedi, H., & Pervaiz, K. (2023). Federated learning for healthcare: A survey.Multidisciplinary Digital Publishing Institute, 7(6),

work page doi:10.1007/s43621-025-01388-2 2023

[4] [9840]

https://doi.org/10.1038/s41598-025-93918-1 Mastour, H., Dehghani, T., Moradi, E., & Eslami, S. (2025). Explainable artificial intelli- gence for predicting medical students’ performance in comprehensive assessments. Scientific Reports, 15(1), 23752. https://doi.org/10.1038/s41598-025-07460-1 McMahan, H. B., Moore, E., Ramage, D., & y Arcas, B. A. (2016). ...

arXiv 2025