Federated Learning of AnDE Classifiers
Pablo Torrijos, Juan C. Alfaro, José A. Gámez, José M. Puerta
cs.LG
Quick verdict
Presents FedAnDE, the first federated framework for discriminative AnDE (n>=1); solid empirical gains over DP generative NB, though privacy claims lack formal backing.
Summary
The paper introduces FedAnDE, a federated learning framework for discriminatively trained Averaged n-Dependence Estimators (AnDE), generalizing federated weighted Naive Bayes (NBw) to any dependency order n. Each client optimizes weights by L-BFGS-B on conditional log-likelihood while its generative parameters Theta stay local; only weight vectors, argued to be semantically meaningless, are exchanged with a server averaging them over 10 rounds. A generative AnDE baseline with Laplace-noise differential privacy (epsilon=1.0) is also implemented. On 12 discrete OpenML datasets with 5-100 clients and 5-fold CV repeated five times, the authors report that discriminative models with n>=1 consistently beat federated NB, that A2DEw is best overall, and that the DP generative baseline degrades sharply with more clients and noise while discriminative models stay robust. Key limitations: the privacy claim has no formal guarantee or leakage quantification, the discriminative-versus-DP-generative comparison is asymmetric, no significance tests are reported, and no non-IID splits or federated logistic regression baselines are evaluated. Code is released on GitHub.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 4
● low 2
4 strengths · 0 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- FedAnDE is the first framework for federated training of AnDE classifiers, extending federated weighted NB (NBw) to any dependency order n, with clients sharing only discriminative weights.
- Sharing discriminative weights improves privacy relative to sharing generative counts because the transmitted parameters lack semantic meaning.
- On 12 discrete datasets, federated discriminative models with n>=1 consistently outperform federated Naive Bayes, and A2DEw attains the highest accuracy across all tested federation sizes.
- Differentially private aggregation (epsilon=1.0) degrades the generative baseline substantially, increasingly so for higher n and more clients, while discriminative training remains robust.
Strengths
First federated treatment of AnDE classifiers, with a well-motivated generalization from NBw
The paper genuinely fills a gap: prior federated probabilistic work covered only n=0 (NBw, ref [5]), and Sections 2-3 give a clean extension to arbitrary n with a per-structure weight vector wP and parallelizable per-P federated training.
Transparent, fully specified protocol with public code and datasets
Section 4 fully specifies T=10 rounds, L=5 L-BFGS-B steps, epsilon=1.0, the Laplace mechanism, and derives the l1-sensitivity Delta(n,d)=1+|Pn|*(1+d-n); code (github.com/ptorrijos99/BayesFL) and OpenML dataset IDs are public, and Appendix A gives per-dataset plots for all 12 datasets.
Consistent and large empirical effects across 12 datasets and 5 federation sizes
Insightful analysis of the DP noise / client-count trade-off
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a compact conference paper (ECML PKDD 2025 proceedings version, CCIS vol. 2841, pp. 464-471) introducing FedAnDE, presented as the first framework for federated training of Averaged n-Dependence Estimator (AnDE) classifiers. The core move is discriminative: extending federated weighted Naive Bayes (NBw) to arbitrary order n, so each client keeps its generative probability tables Theta local and transmits only CLL-optimized weight vectors. Three central claims are made: (i) weight-only sharing improves privacy over count sharing; (ii) higher n improves federated accuracy, with discriminative A2DEw best throughout; (iii) the differentially private generative baseline degrades more, especially at larger C. The claims are mostly supported by the reported experiments, though claim (i) is asserted rather than demonstrated.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 0 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is AnDE in machine learning?
AnDE (Averaged n-Dependence Estimators) are Bayesian network classifiers that generalize Naive Bayes by averaging over models in which each attribute depends on the class plus up to n other attributes. n=0 recovers Naive Bayes and n=1 recovers AODE; larger n captures richer attribute dependencies at combinatorial cost.
What does the FedAnDE paper propose?
It trains Averaged n-Dependence Estimator classifiers across distributed clients without sharing raw data. Each client keeps its probability tables local and sends only discriminative weight vectors to a server, which averages them; this is claimed to leak less than sharing raw probability counts.
Does FedAnDE provide differential privacy?
No formal guarantee is given. The authors argue that weight vectors lack the direct semantic meaning of probability counts and are therefore harder to exploit, while the generative baseline is protected with epsilon-differential privacy (epsilon=1.0). The weights could still encode information about local data, and the paper does not quantify this.
How accurate is federated AnDE compared to Naive Bayes?
On 12 discrete OpenML datasets, the weighted second-order model (A2DEw) achieved the highest mean accuracy at every tested client count (5 to 100 clients), and models with n>=1 consistently beat federated Naive Bayes. Adding Laplace-noise differential privacy hurt the generative baselines noticeably more.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 are axis labels and captions only in the extracted text; exact plotted values could not be verified
- Per-run variance and per-dataset numeric tables are not present in the text
- The GitHub repository was not executed or inspected; reproducibility claims rest on the paper's statements
- Published version: ECML PKDD 2025 CCIS vol 2841, pp. 464-471; this review covers the arXiv author version
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.