acceptodds
Withdrawn from ICLR 2027

FedDCOL:Federated Dynamic Continual Oversampling Learning for Imbalanced Tabular Data Classification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated learning (FL) enables multiple clients to train a global model without sharing raw local data. However, clients usually face the problem of imbalanced class distributions, which may bias the global classifier toward the majority class. Existing data level approaches typically use static oversampling strategies, making it difficult to dynamically adjust minority class generation during federated training. To address this limitation, this study proposes Federated Dynamic Continual Oversampling Learning (FedDCOL). FedDCOL first characterizes the local data structure based on class overlap degree and then partitions local samples into safe, boundary, and noisy regions. Based on this structural information, FedDCOL develops a Structure-Constrained Diffusion Model (SC-Diff). By incorporating a base diffusion objective, a prototype relation constraint and an inter-class margin constraint, SC-Diff learns the data distribution while preserving the discriminative structure of the minority class. During federated training, SC-Diff is continually updated across communication rounds to adapt minority class generation, while structure selection guided by class overlap information retains reliable generated samples for classifier training. This transforms one-shot oversampling into a dynamic continual oversampling process. Experimental results on seven real world imbalanced tabular datasets show that FedDCOL outperforms six SOTA methods, achieving improvements of , , and in AUC, F1-score, and G-mean over the strongest baselines, respectively. The particularly pronounced gains in F1-score and G-mean highlight the effectiveness of dynamic continual oversampling for minority class learning in federated tabular classification.

Primary area:
generative models
Keywords:
Federated Learning, Class Imbalance, Dynamic Continual Oversampling, Structure-Constrained Diffusion

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_JyypOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FedDCOL, a federated oversampling framework for imbalanced tabular classification. The method first defines a class-overlap score using neighborhood label consistency and relative distances to class centers, based on which local samples are partitioned into safe, boundary, and noisy regions. It then develops SC-Diff, a tabular diffusion model augmented with a prototype-relation constraint and an inter-class margin constraint. During federated training, SC-Diff is repeatedly updated and used to generate minority-class candidates, which are filtered according to the same structural criterion before being incorporated into classifier training. Experiments on seven tabular datasets report improvements in AUC, F1-score, and G-mean over six baselines

Critical strengths

--Minority-class scarcity is indeed a meaningful challenge in federated tabular learning, and dynamically updating a generative oversampler during federated optimization is a reasonable direction beyond one-shot preprocessing.

--The authors report results on seven datasets using AUC, F1-score, and G-mean, perform Friedman/Nemenyi testing, include parameter sensitivity studies, and provide component-level ablations

Critical weaknesses

+FedDCOL combines several existing ingredients, including neighborhood-based overlap estimation, prototype modeling, tabular diffusion, margin/prototype regularization, minority-class oversampling, and federated gradient aggregation. The main additional idea is to continually update the diffusion model and regenerate minority samples across communication rounds. However, the current experiments do not convincingly isolate whether the gains come from dynamic updating itself, the use of diffusion instead of other generators, the structural constraints, or the candidate filtering strategy. Stronger controlled comparisons with static diffusion, periodically retrained oversampling, and diffusion without structural constraints are needed.

+The proposed overlap score combines same-class KNN consistency and relative class-center distance through a specific exponential transformation, but the functional form and thresholds are largely heuristic. More importantly, FedDCOL applies KNN and Euclidean class-center distances directly after concatenating min-max normalized numerical variables with one-hot categorical variables. Such geometry may not reliably reflect semantic similarity in heterogeneous tabular data. Since the safe/boundary/noisy partition is central to the entire framework, the authors should justify this metric more rigorously and compare it against alternatives such as Gower distance, learned embeddings, or other tabular similarity measures.

+The paper defines the prototype set using embeddings of all safe samples rather than a small number of representative class prototypes. As a result, the relation vector has dimensionality proportional to the number of safe samples, which can make the prototype-relation constraint increasingly expensive as local dataset size grows. This also makes the term “prototype” somewhat misleading. In addition, the nearest-prototype margin formulation may be unstable when many reference embeddings are used. The paper should provide complexity analysis and compare against class centers or clustered prototypes.

+All experiments use only five clients with full participation in every round, while the default client partition uses a Dirichlet concentration of 1. This does not adequately test realistic federated settings with more clients, partial participation, unequal data sizes, heterogeneous imbalance ratios, or client dropout. At the same time, FedDCOL repeatedly trains and communicates a diffusion model, generates candidates, evaluates their structure, and updates the classifier, yet no systematic communication, runtime, memory, or scalability analysis is provided. These omissions weaken the practical federated-learning claims of the paper.

+Although the paper includes ablation studies and sensitivity analysis, it does not directly address the key causal question of whether continual dynamic oversampling outperforms comparable static or periodically updated generative oversampling approaches. Sensitivity analysis is performed exclusively on the BUPA dataset even though the method involves many critical hyperparameters. Furthermore, the manuscript suffers from multiple writing and formatting issues. Representative examples include misspellings such as “CLass Overlap Structure Modeling”, “Mini‑bacth”, “Multionmial”, and “Blances samples”, as well as redundant phrasing “Equation equation 6”. Some labels in Figure 1 are also rendered too small for clear readability. More rigorous experimental decomposition and comprehensive proofreading are required.

Questions for authors

--How do the authors justify the specific formulation of the class-overlap score?

--How reliable is the proposed overlap modeling for mixed-type tabular data?

--Why are all safe-sample embeddings treated as prototypes in the prototype-relation constraint?

Reviewer_ynSpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper attempts to address class imbalance in federated tabular classification and the inability of static oversampling methods to adapt to ongoing training. The authors propose FedDCOL, a dynamic continual oversampling framework that combines class overlap structure partitioning with a structure-constrained diffusion model to improve minority-class classification performance.

Critical strengths
  1. FedDCOL continually generates new minority class samples during federated training, rather than just once before training.
  2. It selects more reliable generated samples based on their class overlap degree.
  3. Good results were achieved on 7 real datasets, and the effectiveness of the method was verified through ablation experiments.
Critical weaknesses
  1. Lack of key comparative experiments, without direct comparison with classical oversampling methods such as SMOTE and existing tabular diffusion generation methods.
  2. Although the average performance is better, the improvement compared to the strongest baselines FedANAG and FedWGAN has not reached statistical significance.
  3. The additional computational and communication costs and training time are not evaluated, leaving practical feasibility unclear.
  4. Experiments only cover 5 clients, binary classification, and full client participation scenarios, lacking validation in more complex environments.
  5. The existing ablation experiments cannot distinguish the effects of continuously updating the generative model and regenerating samples each round.
Questions for authors
  1. Can you separately evaluate the effects of continuously updating the generative model and regenerating samples in each round?
  2. Compared to existing methods, how much computation, communication, and time overhead does FedDCOL incur?
  3. Is the method still effective in scenarios with more clients, partial client participation, and more severe class imbalance?
  4. Without differential privacy or secure aggregation, how do you evaluate and mitigate the potential risk of privacy leakage from client gradients?
Reviewer_97iPOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces FedDCOL, a framework for addressing class imbalance in federated tabular data classification. First, a class overlap degree is computed to partition the local data into samples classified as safe, boundary, or noisy. Using this partition, the authors introduce a structure-constrained diffusion model (SC-diff), which combines a diffusion loss with a relation constraint and an interclass margin constraint. This allows the model to learn the data distribution while preserving the structure of the minority class. During federated training, SC-diff is continually updated between communication rounds and used to generate minority-class samples, which are filtered according to their class overlap degree before being added to the classifier training data. The authors evaluate the framework on several imbalanced datasets and compare its performance against a number of federated learning baselines.

Critical strengths
  1. Well-motivated problem. Class imbalance in federated learning is an important problem, particularly when minority-class observations are scarce on individual clients.

  2. Empirical evaluation. FedDCOL achieves good average performance across the seven datasets, particularly in terms of F1-score and G-mean. The paper also includes ablation studies and sensitivity analyses to evaluate the proposed method.

Critical weaknesses
  1. Limited methodological novelty. The proposed method combines existing ideas, and it is unclear how its contribution differs substantially from existing approaches.

  2. Limited evidence of improvement. Although FedDCOL achieves the best average results, the improvements over the two strongest baselines, FedANAG and FedWGAN, are not statistically significant. The gains observed in the ablation studies are relatively small and largely driven by the Parkinson's dataset. It is also unclear how the hyperparameters and classification thresholds were selected and whether they were tuned in a comparable way across methods.

Questions for authors
  1. Novelty of the proposed method. What is the main novelty of the dynamic continual oversampling compared with existing methods that update a generative model during federated training? Could the authors compare their method with a simpler approach that updates a standard diffusion model without the additional structural constraints and sample selection?

  2. Experimental validation. How much of the improvement is due to the proposed method rather than oversampling itself? Could the authors include comparisons with simpler methods, such as class-weighted FedAvg and local SMOTE, and other approaches, such as Fed-OLF? Could they also clarify how the hyperparameters and classification thresholds were selected and whether the same tuning procedure, using only validation data, was applied to all methods?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.