acceptodds
Under review as a conference paper at ICLR 2027

DataFed: Bid-Agnostic Auctions for Hierarchical Client Selection in Online Federated Edge Learning

Abstract

Dynamic client availability and resource variations motivate online federated edge learning (OFEL), where real-time client selection requires truthful auctions to align clients’ bids with their actual costs. However, existing auctions face three key challenges: (i) two-sided incomplete information in bid-agnostic settings, (ii) hierarchical objective misalignment between cloud and edge servers, and (iii) lack of long-term optimization in dynamic environments. To address these challenges, we propose DataFed, a bid-agnostic auction mechanism that integrates dynamic Bayesian bid distribution estimation with Stackelberg-based client selection. Specifically, we formulate hierarchical client selection as a dynamic Bayesian Stackelberg game with a multi-armed bandit objective to minimize cumulative expected regret. With a known bid distribution, we establish a Stackelberg equilibrium for the normalized proxy model. When the distribution is unknown but fixed, we derive estimation error bounds under independent proxy feedback. With sufficient exploration, we prove sublinear cumulative expected regret relative to the proxy benchmark. By adaptively selecting clients with favorable utility-cost trade-offs, DataFed improves learning efficiency and stabilizes online training in dynamic environments. Experiments show that DataFed reduces mean cumulative regret by 23.0% on average across four real-world datasets and improves mean cloud utility by 7.9% through ES response anticipation.

Primary area:
learning theory
Keywords:
Online Federated Learning, Bid-Agnostic Auction, Stackelberg Game, Online Learning with Sublinear Regret

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_LX4qOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

DataFed addresses hierarchical client selection in online federated edge learning. The cloud first chooses the winner count, edge servers choose candidate counts, and a reverse Vickrey auction then allocates participation using submitted bids. The method estimates an unknown bid distribution from historical payment thresholds and coordinates decisions through Stackelberg response anticipation. The paper proposes an equilibrium construction for a normalized proxy model and analyzes estimation error and sublinear regret under independent proxy feedback and a prescribed exploration policy. Experiments evaluate cumulative regret on four classification datasets and jointly report accuracy and payment on two, with additional synthetic-bid, coordination, drift, and approximation comparisons.

Critical strengths
  1. The paper combines bid-distribution estimation from historical payment thresholds with hierarchical response anticipation to address cloud–edge coordination under an unknown bid distribution. The motivation and overall design are practically relevant. Within the proxy model, the finite-sample estimation bounds and sublinear-regret derivation form a coherent analytical chain under explicit conditions, providing concrete results on how sample size affects proxy decision error.
  2. The experiments cover multiple datasets, system scales, and random seeds, with joint accuracy, payment, and regret evaluation on a subset of the datasets to characterize their trade-offs. Additional drift and exact-versus-proxy comparisons examine key design choices and their limits.
Critical weaknesses
  1. Definition 2 lets each ES choose its own candidate count, whereas A.2 first assigns a common count to all ESs, which is then optimized in Eqs. (32)–(33) of A.3 and line 10 of Algorithm 1. I am concerned that optimizing this common parameter changes the symmetric configuration without establishing a unilateral best response with the other ESs' strategies fixed. Even a symmetric equilibrium must rule out unilateral deviations. Whether the constructed response corresponds to the original multi-follower game therefore remains a material concern.
  2. The normalization in A.2 preserves the expected winner count, but the utilities in Eq. (8) are further constructed proxies. No general control is provided over how their approximation error affects action rankings or utility loss under the original objective. I am concerned that proxy decisions may depart from the original objective even when the bid distribution is accurately estimated.
  3. Theorems 2–3 explicitly concern independent proxy feedback, and Theorem 3 uses an exploration phase for every action, whereas Algorithm 1 observes actual payment thresholds and supplements samples of the current action. The analysis therefore does not establish a regret guarantee for the implemented algorithm against the original objective in Eq. (7). How do these differences affect the implemented algorithm's decision quality and cumulative regret under the original objective?
Questions for authors
  1. Please see Weaknesses 1–3. Could the authors prioritize addressing the specific concerns raised there?
  2. I remain unsure how q(ζ)q(\zeta) is obtained before client selection, given that A.7.1 defines it using test accuracy. Are the data used to construct these valuations also used for the final reported test evaluation?
Reviewer_b1RfOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DataFed, an online auction mechanism for hierarchical client selection in federated edge learning. It integrates bid-distribution estimation, Stackelberg coordination, and online decision-making. The problem is relevant, and the experimental results demonstrate promising performance improvements.

Critical strengths

None

Critical weaknesses
  1. The paper exhibits significant traces of AI-generated text; certain sentences lack coherence, leaving the reader confused. Although the authors acknowledge using AI to refine the writing, the actual quality remains suboptimal, and further manual review is recommended.

For instance, lines 41–43 state: "This setting naturally motivates the adoption of online federated edge learning (OFEL) (Xin et al., 2025; Chen et al., 2020), where hierarchical client selection becomes an online decision-making problem."

Furthermore, the paper relies heavily on excessive conjunctions and "defensive writing"—hallmarks of AI-generated text. While these elements do not constitute technical errors per se, they often make the text confusing or awkward for the reader.

  1. The analysis of the research motivation (i.e., the limitations of existing work) lacks depth; it merely reiterates limitations already identified in previous studies, and the most recent references cited date back only to 2024 (Hu et al., 2024; Jin et al., 2024; Liu et al., 2024). Personally, I find the authors' contribution regarding the analysis of research motivation to be negligible.

  2. The paper heavily emphasizes the concepts of "Online" and "Dynamic" in the title, abstract, and introduction, describing the model as a "Dynamic Bayesian Stackelberg Game." However, the theoretical analysis relies on assumptions such as "independent sampling of client bids," "fixed bid distributions during analysis," and "independent agent feedback." These assumptions diverge significantly from the realities of edge computing environments, which are characterized by "dynamic client resources" and "non-stationary cost fluctuations." I suggest the authors re-evaluate the focus of their research.

  3. The paper appears to lack a traditional ablation study; it presents only experiments on hierarchical response prediction but fails to conduct ablation analyses on key functional modules such as "historical bid distribution learning" and "decoupled cloud-edge coordination."

  4. The theoretical regret guarantee is established for an independent normalized proxy-feedback learner rather than the actual proposed algorithm. The normalization lacks general approximation guarantees, and the relationship between the exploration procedure and the assumptions of the regret theorem is unclear.

Questions for authors

None

Reviewer_2jLgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on client selection under federated edge learning and proposes DataFed, a bid-agnostic auction mechanism that integrates dynamic Bayesian bid distribution estimation with Stackelberg-based client selection. The authors validate the effectiveness of the proposed method under given settings.

Critical strengths
  1. Generally, this paper is well-structured and easy to follow.

  2. The combination of bid distribution estimation and hierarchical Stackelberg decision-making provides a potentially useful approach to coordinating cloud and edge servers.

Critical weaknesses
  1. The main concerns for me are regarding the compatibility of the selected baselines with the proposed reverse auction setting. Several baselines, particularly RLB-AFL, are designed for fundamentally different auction settings (forward auction), strategic roles, and optimization objectives, raising concerns about the fairness and validity of the reported comparisons.

  2. It seems that the paper does not establish a theoretical bound on the approximation error between the original multi-winner reverse auction and the normalized single-winner proxy used for equilibrium and regret analysis.

  3. The motivation of this paper and the issues raised by the paper seems to be common in federated edge learning and have been widely studied in prior work. To my understanding, the motivation of this paper is less compelling and the novelty of the proposed method is insufficiently clear.

  4. Several grammatical errors and imprecise expressions, such as "a natural objective for evaluate long-term utility," reduce the clarity and technical precision of the manuscript.

  5. Table 2 reports identical SVHN results across different edge-server configurations, while the implementation and adaptation details of the baselines remain insufficient.

Questions for authors

Please refer to weaknesses.

Reviewer_pDfyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DataFed, a bid-agnostic auction framework for hierarchical client selection in online federated edge learning. It integrates a k-reverse Vickrey auction, Bayesian bid distribution estimation, and Stackelberg-based cloud-edge coordination. The authors establish equilibrium existence, estimation error bounds, and sublinear cumulative regret under a normalized proxy model. Experiments on four real-world datasets report a 23.0% average reduction in cumulative regret and a 7.9% improvement in cloud utility.

Critical strengths
  1. The paper addresses two-sided incomplete information and conflicting cloud-edge objectives in online federated learning. Figure 1 illustrates the hierarchical selection process, extending conventional single-level client selection to a multi-client, cloud-edge setting.

  2. DataFed combines payment-threshold-based bid distribution estimation with Stackelberg response anticipation. The formulation in Eqs. 8–12 provides a structured framework for hierarchical decision-making under incomplete bid information.

  3. Table 2 reports a 23.0% average reduction in cumulative regret across the evaluated configurations. Furthermore, Table 6 demonstrates an 11.0% regret reduction and a 7.9% improvement in cloud utility through edge-server response anticipation, supporting the effectiveness of hierarchical coordination.

Critical weaknesses
  1. Theorem 3 establishes sublinear regret under independent normalized proxy feedback, whereas Algorithm 1 uses realized multi-winner auction payment thresholds. As acknowledged in Section 5.3, the theoretical guarantee does not directly apply to the implemented algorithm, limiting the strength of the main theoretical contribution.

  2. The analysis relies on i.i.d. bids and symmetric edge-server strategies. These assumptions may limit the applicability of the theoretical results to heterogeneous federated edge environments, where client cost distributions and edge-server strategies can differ substantially.

  3. able 2 shows that DataFed does not achieve the lowest mean regret in several configurations, including FEMNIST (M=10, N_m=25) and CIFAR-10 (M=10, N_m=50). Moreover, Table 7 reports a 7.68%–9.82% utility gap between proxy-selected and exact optimal actions in the evaluated synthetic settings, highlighting the approximation error introduced by the normalized proxy model.

Questions for authors
  1. Can the authors extend the regret analysis in Theorem 3 to the actual multi-winner payment feedback used in Algorithm 1, or provide additional empirical evidence quantifying the gap between proxy and realized regret?

  2. How does DataFed perform under heterogeneous, non-i.i.d. bid distributions and asymmetric edge-server strategies? Could the authors evaluate these settings to further assess the generality of the proposed framework?

  3. How is the cloud valuation q=10xAcc estimated before client selection in the online experiments?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.