- 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