acceptodds
Withdrawn from ICLR 2027

Online Missing-Update Correction under Informative Client Participation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-device federated learning relies on partial participation, yet client availability can depend on prior activity, staleness, and optimization state. The returned updates can therefore be an informative rather than representative subset, causing standard averaging to favor repeatedly available clients. Existing methods typically correct participation probabilities or reconstruct missing updates; either correction can become misspecified as availability and client trajectories drift. We introduce FedDORA, a server-side framework that treats participation as sequential missingness and combines leakage-free online propensity estimation with vector-valued update prediction. A population-normalized orthogonal aggregate couples these two nuisance models: under history-conditional ignorability and overlap, either a correct propensity or update model removes conditional aggregation bias, while joint misspecification enters through a product remainder. We establish an exact bias identity, an explicit variance formula under conditional client independence, and a general non-convex convergence bound. Across CIFAR-10, CIFAR-100, and OfficeHome, controlled observable and latent participation mechanisms, residual adaptation, and oracle interventions, FedDORA achieves competitive endpoint performance while providing strong tail and mechanism-level evidence for informative-participation correction. Ablations and hidden-driver tests further delineate its identification boundary.

Primary area:
optimization
Keywords:
federated learning, informative client participation, client availability, missing updates, doubly robust estimation, orthogonal estimation, partial participation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dbYdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FedDORA treats partial client participation as sequentially missing vector updates. Before observing the current round's availability, the server estimates participation probabilities and predicts client updates from two-update memories and sketch-based trajectory routing. A population-normalized augmented inverse-probability aggregate combines these predictions with weighted observed residuals. Under history-conditional ignorability and overlap, the paper derives a product-form bias remainder, a conditional variance expression, and a non-convex optimization bound. The evaluation uses controlled participation, primarily frozen visual representations, five seeds, and nuisance/oracle controls.

Critical strengths

A careful estimand and chronological information boundary.

The evaluation is more informative than endpoint accuracy alone.

Critical weaknesses

a) The paper does not yet convincingly establish a new estimation principle, a subtantive closed-loop learning guarantee, or a sufficiently large practical advance to make the overall contribution compelling at this threshold (Sect. 2 and Appx. B)

b) Strongest empirical claim but lack of correction-mechanism validation: Appx. C specifies that outcome error is proxied by residuals over observed clients and gradient bias by a participant vs. memory discrepancy. These are not the conditional population quantities in the bias identity and can contain selection effects and local-updated noise. The primary participation mechanism also uses the same logistic history-feature structure as the learner. Oracle propensities test one branch, but do not isolate the correct-outcome/misspecified-propensity branch. The endpoint gains do not compensate for these gaps, ie. on CIFAR-100, mean improvement over FedAvg is 0.08 points with a 0.18-point tail deficit; on OfficeHome, the mean improvement over the strongest baseline is 0.59 points. I am not requiring universal SOTA, but rather stronger evidence for the contribution the paper actually emphasizes.

Questions for authors

a) Is a small fixed-state simulation, can you estimate the full-client conditional mean by repeated local-update sampling and report the actual aggregation bias and variance for a 2x2 correct/misspecified propensity and outcome design? This would directly test both claimed robustness branches without requiring a new large benchmark.

b) Which result goes beyond applying conditional AIPW to a drifting vector outcome plus a general biased-descent bound? Can the implemented nuisance learners be connected to a nontrivial error or tracking guarantee, rather than only assuming their rates?

c) Does the benefit persist when participation depends on fully observed history through interactions or nonlinearities outside the logistic model class, while conditional ignorability is retained? This is distinct from the existing hidden-driver violation, where identification is intentionally lost

Reviewer_P5RCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces FedDORA, a federated learning algorithm where the client's availability depends on observable history. The proposed scheme uses a probability weighting estimator to weigh the client's data importance and therefore to achieve the true optimal solution.

Critical strengths

The setting of informative client participation is interesting.

The usage of AIPW weighting is a good idea and makes intuitive sense.

Critical weaknesses
  1. The paper is not clearly written and full of jargon. What is propensity or AIPW? Different fields use these terms differently and they need to be carefully explained.

  2. Math were not rigorous. Are Δ\Delta and μ\mu the same thing? Corollary 2 assumes a bound on ϵ\epsilon, but how do you guarantee these bounds can be established? They were used as upper bounds in Theorem 2. With clipping and constant stepsize, these errors could be constant as t approaches infinity.

  3. In the simulation, the ground truth propensity error is known, which is not a reasonable assumption in practice.

  4. Numerical experiment suggests little or no improvement. It would also be good to report on memory usage. The algorithm seems to require Nd storage, which would not be very scalable.

Questions for authors

See weakness.

Reviewer_fvk3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FedDORA, which addresses biased client participation in federated learning. The server predicts each client's participation probability and missing update, then corrects the predictions using received updates. Under the stated assumptions, the aggregate is unbiased if either predictor is correct. The paper gives bias, variance, and convergence bounds and tests simulated participation on some image datasets.

Critical strengths
  1. The idea proposed in this paper makes sense. The bias and variance derivations appear correct under the stated assumptions. The paper clearly separates the mean full-client update from a noisy realized update, and makes predictions before observing the current round.

  2. The evaluation includes five paired seeds, oracle-probability controls, router ablations, and a hidden-driver stress test.

Critical weaknesses
  1. The novelty is limited. The score, product-bias identity, and orthogonality follow standard doubly robust estimation. AAggFF-D already uses this idea in FL, although for a different target and known sampling probabilities. The main new part is the online vector-update implementation and its predictor.

  2. The convergence result does not establish that the proposed predictors become accurate. Corollary 2 assumes their error rates, and Eq. (15) assumes that stored update directions identify clients with similar current updates. Neither property is proved for the implemented logistic learner and router as the global model changes. In addition, Assumption 2 is particularly strong since the local-SGD drift is uniformly bounded. Hence, the theoretical guarantees are also very weak.

  3. The main participation simulator uses the same logistic form and history features as the propensity learner (Eq. (18)). This is a favorable test. The controls do not demonstrate the other half of double robustness, which is an accurate update predictor protecting against an incorrect propensity model.

  4. The interpretation of the hidden-driver test needs justification. Its driver is a fixed statistic of each client's label mixture. With fixed client data and independent availability and minibatch draws, hiding this statistic from the learner need not violate Assumption 1. It may instead test propensity-model misspecification.

  5. The central mechanism evidence is indirect. Figure 4 uses residuals on participating clients as a proxy for prediction error. These also contain update noise and selection effects. Its bias reference uses stored updates. These quantities do not directly measure the conditional aggregation bias or prediction error in the theorem.

  6. The practical gains are small and mixed. CIFAR-100 improves mean accuracy over FedAvg by only 0.08 points and loses 0.18 points in tail accuracy. OfficeHome is more encouraging, but its mean gain over the strongest baseline is 0.59 points. Stronger mechanism evidence is therefore important.

Questions for authors
  1. What randomness makes the hidden label-mixture driver violate conditional independence, rather than just make the logistic model incorrect? Please clarify whether the analysis conditions on the sampled client datasets.

  2. Can you estimate true aggregation bias using repeated all-client updates at fixed models, and independently vary the two prediction errors? In particular, test accurate update predictions with deliberately wrong propensities.

  3. Do the gains hold for larger networks trained from pixels, longer runs, and equal tuning budgets for each method? Please also report server memory and time at those model sizes.

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

Related papers

Discussion (0)

Sign in to comment.