acceptodds
Under review as a conference paper at ICLR 2027

Robust Federated Primal–Dual Learning via Dynamic Graph Topology Correction under Sparse Client Participation

Abstract

Federated primal–dual methods rely on client-specific dual variables to correct drift, yet sparse participation leaves most of these variables without fresh updates. Mean-based compensation addresses this missing information by assigning every inactive client the same correction, overlooking differences in their relationships with active clients. We propose RFL-DFDC, which reconstructs missing dual increments from client relationships and couples these reconstructions to the server update. A dynamic graph determines how observed increments contribute to each inactive client's virtual dual. The same weights determine the corresponding contributions to the global dual, preserving its equality to the average of virtual client duals. Our analysis bounds reconstruction error, accounts for missing graph support, and characterizes how these errors propagate to nonconvex optimization and returning clients under stated regularity conditions. Experiments on heterogeneous federated benchmarks show higher accuracy and fewer communication rounds to target accuracy than mean compensation. A comparison with a simple cosine-based alternative further supports the value of the proposed relationship weights.

Primary area:
optimization
Keywords:
federated learning, primal-dual methods, partial participation, client heterogeneity, stale dual states, graph signal diffusion

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rt4wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies RFL-DFDC, a federated primal–dual learning algorithm designed to address dual variable drift caused by partial client participation. Building on A-FedPD, which compensates for missing dual updates using the average update of participating clients, the proposed RFL-DFDC introduces client-specific compensation based on relationships among clients. Specifically, it constructs a dynamic graph using cosine similarities between client model updates and uses the resulting weights to reconstruct missing dual updates for inactive clients. This method also introduces a corresponding server-side correction to maintain consistency between the global dual variable and the average of the virtual client dual variables. Theoretical convergence analyses provide bounds on reconstruction errors on nonconvex optimization under certain assumptions. Experiments on heterogeneous federated learning benchmarks demonstrate improved prediction accuracy and communication efficiency compared with existing methods, including A-FedPD.

Critical strengths

S1. Simple and intuitive algorithmic extension.

The proposed method provides a simple and intuitive extension of A-FedPD by replacing uniform mean-based dual compensation with client-specific compensation based on relationships among clients. The overall algorithmic design is straightforward and easy to understand.

S2. Theoretical analysis of the proposed algorithm.

This paper provides theoretical analyses of the proposed algorithm, including bounds on dual reconstruction errors and their effects on nonconvex optimization.

S3. Empirical improvements over existing methods.

The proposed method demonstrates improvements in test accuracy and communication efficiency over existing FL methods, including A-FedPD and A-FedPDSAM, on the evaluated benchmarks.

Critical weaknesses

W1. Limited algorithmic novelty.

The problem of dual-variable drift under partial client participation has already been extensively investigated in federated primal–dual learning, particularly in A-FedPD. The main algorithmic contribution of this paper appears to be replacing the uniform mean-based dual compensation in A-FedPD with a weighted average based on cosine similarities between client model updates. Although the proposed extension is reasonable, I find its algorithmic novelty somewhat limited.

W2. Insufficient theoretical justification for the proposed dynamic graph-based reconstruction.

Although this paper provides a theoretical analysis of reconstruction errors and their effects on nonconvex optimization, I am not fully convinced that the analysis establishes the effectiveness of the proposed dynamic graph-based reconstruction mechanism. In particular, Theorem 1 provides a stationarity bound in terms of general reconstruction-error quantities, and the analysis appears to be applicable to various reconstruction schemes, including uniform mean-based compensation. Therefore, it remains unclear whether the theoretical results demonstrate that the proposed graph-based reconstruction achieves smaller reconstruction errors or improved convergence rates than existing methods.

Moreover, several key components of the proposed method, including cosine-similarity-based graph construction, EMA-based dynamic graph updates, and the limited observability of client relationships under sparse participation, do not appear to be explicitly characterized in the main convergence analysis. Although their effects may be implicitly captured by the general reconstruction-error terms, the theoretical implications of these specific design choices remain unclear.

W3. Limited empirical evidence supporting the generality of the proposed method.

Although the proposed method outperforms the considered baselines, I have some concerns about the experimental evaluation scores. In particular, the absolute test accuracies appear relatively low. This may be partly attributed to the use of relatively small LeNet-style architectures for CIFAR-10/100, whereas ResNet-18 is evaluated only on Tiny-ImageNet.

Questions for authors

Q1. Theoretical characterization of the reconstruction-error bound.

Regarding W2, I have some concerns about Assumptions 1 and 2, which play central roles in the convergence analysis. In particular, Assumption 2 introduces a general upper bound on the graph reconstruction error, but it remains unclear how this bound can be theoretically characterized for the proposed method. If possible, could the authors provide a theoretical characterization of this bound? If such an analysis is difficult, could the authors empirically investigate the magnitude and evolution of the reconstruction-error bound during training, and compare it with that of uniform mean-based compensation?

2 comments

Thank you for recognizing the algorithm's clarity and empirical gains. The advance is client-specific evolution of persistent dual states, with a matching server update derived from the same reconstruction operator and an analysis of imperfect completion.

1. The contribution goes beyond choosing cosine weights (W1). A-FedPD supplies the mean-compensation framework. RFL-DFDC changes how each inactive client is represented within it. Rows of the reconstruction operator determine individual virtual-memory updates; column sums determine active clients' influence on the server dual. Consequently, the same inferred increment affects the returning client's next augmented local problem and the current global update. Equation (11) makes the need to coordinate these two paths explicit.

The existing interventions demonstrate the value of this construction:

  • Relationship-conditioned completion: at matched β=0.2\beta=0.2, Table 3 reports 51.81% versus 46.16% for mean completion, a 5.65-point gain.
  • Learned client assignments: last-update cosine gives 48.46%, random relationships 37.29%, and shuffled learned relationships 41.74%.
  • Coupled correction: Table 2 gives 48.25% with virtual-memory correction alone versus 51.81% with both memory and server corrections.
  • Continued state evolution: Table 4 gives 51.81% with correction throughout training versus 40.33% when it stops after ten rounds.

These controls support relationship-aware correction, coordination, and persistent application within the evaluated host.

2. The theory characterizes how reconstruction quality affects optimization (W2; Q1). The analysis establishes three complementary results. First, Eq. (9) preserves the global dual as the average of virtual client duals exactly, regardless of prediction accuracy. Second, Lemma 1 bounds the perturbation caused by incomplete reconstruction:

∥ξt∥2≤N−PtN2Eβt.\|\xi_t\|^2\leq\frac{N-P_t}{N^2}\mathcal E_\beta^t.

Third, Theorem 1 propagates this perturbation into stationarity under explicit host and graph budgets. Appendix B.1 additionally characterizes how one-step errors accumulate into a returning client's dual input under its local-transition conditions. This connects missing-update reconstruction to both current server optimization and subsequent client optimization.

Assumption 1 has a concrete sufficient characterization in Appendix A.2. Conditional uniform sampling, local-displacement budget ctc_t, and local-stationarity-residual budget rtr_t yield

vt=3(χt−2+L2)ct+3rt.v_t=3(\chi_t^{-2}+L^2)c_t+3r_t.

Finite-step local optimization, including SAM, enters through the residual budget. Assumption 2 controls the conditional expected scaled graph energy. Importantly, Eq. (12) includes (1−skt)∥zkt∥2(1-s_k^t)\|z_k^t\|^2, explicitly accounting for unsupported clients, and the analysis permits dependence between learned relationships and current updates.

The general kernel formulation is useful: mean completion is a special case, and Appendix A.4 specifies how a smaller matched-state energy budget improves the graph-error term with the host term fixed. Cosine, EMA, and co-participation history affect that term through the learned kernel. The theorem provides conditional error propagation, rather than a universal ranking of similarity metrics or an unconditional faster rate for cosine–EMA. Numerical energy trajectories are not reported; the rule's practical benefit is assessed through the controlled comparisons in Table 3.

The equations establish consistency and error propagation; the interventions test the relationship rule. Appendix A.5 distinguishes reconstruction error from neighbor dispersion, so energy serves as an analytical bound.

3. The gains extend beyond LeNet and across task types (W3). The evaluation includes LeNet-style CIFAR models, ResNet-18 with GroupNorm on Tiny-ImageNet, and a two-layer LSTM on naturally partitioned Shakespeare clients (Section 5.1; Appendix C.2).

On Tiny-ImageNet, RFL-DFDC achieves 39.27% versus 34.81% for A-FedPDSAM, a 4.46-point gain, and reaches 30% in 542 rather than 822 rounds. It also exceeds FedSpeed's 35.24%, the strongest baseline accuracy in that table. Thus the benefit is observed with ResNet-18, alongside the 1.48- and 3.97-point gains over A-FedPDSAM on CIFAR-10/100. Shakespeare further evaluates the method on a different architecture and naturally partitioned language data.

The relevant comparison holds the architecture and federated protocol fixed within each task. With 500 heterogeneous vision clients and nominal 2% participation, the results show improved accuracy and fewer communication rounds under the same evaluated setting. The empirical contribution is consistent gains across these sparse-participation tasks.

Thank you for the detailed responses. However, I must say that my main concerns remain unresolved.

Regarding W1 and W2, I understand that the proposed method jointly considers client-specific dual reconstruction and the corresponding server-side correction, and I acknowledge its empirical effectiveness. However, my main concern remains the limited methodological and theoretical novelty. The proposed method appears to be a natural extension of the uniform mean-based compensation in A-FedPD, while the server-side correction follows directly from the average-dual identity. Furthermore, although the theoretical analysis characterizes how reconstruction errors affect optimization, it does not establish why the proposed cosine-similarity-based dynamic graph should provide more accurate reconstruction or better convergence guarantees than existing approaches. Regarding W3, my concern is primarily about the relatively low test accuracies on CIFAR-10/100, which may be partly attributed to the use of relatively small LeNet-style architectures.

I will maintain my original rating at this stage.

Reviewer_c4dUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses stale dual states in federated primal–dual learning under sparse client participation. RFL-DFDC learns client relationships from updates observed during joint participation. It uses this graph to estimate inactive clients’ dual increments and applies a matching server correction. The paper analyzes how reconstruction errors affect optimization and returning clients under stated assumptions. It reports accuracy gains on image classification and Shakespeare benchmarks, and fewer rounds to target accuracy on the image tasks.

Critical strengths
  1. Client-specific reconstruction addresses a clear limitation of giving every inactive client the same correction.
  2. The coupled updates preserve the average-dual identity in Eq. (9). The error bound also accounts for unsupported clients.
  3. Table 3 includes useful mean, last-update cosine, and random-graph controls. Table 1 reports gains over A-FedPDSAM across all three image tasks.
Critical weaknesses
  1. The added value over earlier dual-compensation methods is unclear. RFL-NLCP [1] already uses client-update similarity for cluster-based dual amplification. FedCVC [2] instead compensates returning clients after local training using their own updates and absence duration. Neither is discussed or compared. These comparisons would help establish the benefit of reconstructing each inactive client’s dual state throughout its absence.
  2. The benefit of the graph-weight normalization for missing dual updates in Eq. (6) is not directly tested. For an inactive client, the weights over active clients are rescaled to sum to one when their total exceeds εA\varepsilon_A. Table 3 changes how graph weights are obtained, while Figure 3 varies a shared β\beta. Neither tests whether inactive clients with small total graph weights should receive smaller corrections.
  3. The accuracy gains over A-FedPDSAM at 0.5% and 1% client participation (Figure 5(b)) may partly come from skipping some dual corrections. Under Eqs. (6)–(7), an inactive client with no graph connection to any active client receives no correction. Mean compensation instead corrects every inactive client. The experiment does not separate this effect from better estimates of missing updates. Table 3 tests random connections with the same neighbor counts at 2% participation, but not at these lower rates.

[1] Lai et al. “RFL-NLCP: Robust Federated Learning with non-IID Data and Limited Client Participation.” ICASSP, 2026.

[2] Lai et al. “FedCVC: Federated Primal-Dual Learning with Client-Driven Virtual Compensation for Mitigating Dual Drift.” CVPR Findings, 2026.

Questions for authors
  1. What benefit does online, client-specific reconstruction add over RFL-NLCP? Under matched training and participation settings, how does RFL-DFDC compare with RFL-NLCP and FedCVCSAM in accuracy, rounds to target, and server cost, including initialization?
  2. Would smaller corrections for clients with small total graph weights improve accuracy over Eq. (6)? How would this compare with simply reducing the shared β\beta?
  3. At 0.5% and 1% participation, does RFL-DFDC still outperform Table 3’s random-graph control? This control keeps the same clients receiving corrections and the same number of active neighbors for each.
2 comments

Thank you for identifying these distinctions. RFL-DFDC reconstructs individual inactive clients' persistent dual states from partial-participation observations, with a matching current server update. Existing controls establish the value of both relationship assignments and coupling.

1. How the method differs from RFL-NLCP and FedCVC (W1; Q1). The distinction is the granularity, timing, and information used for compensation:

  • RFL-NLCP: update-based clustering and cluster-level amplification.
  • FedCVC: retrospective compensation using a returning client's own current update and absence duration, after its local primal computation.
  • RFL-DFDC: online pairwise relationships from ordinary co-participation; client-specific dual reconstruction during absence; the accumulated state supplied before the next local primal computation; and a matching contribution to the current server dual.

Thus, DFDC changes both the state on which a returning client optimizes and the server's trajectory between visits. It starts from zero virtual duals and an identity graph, without an all-client graph-initialization round. The contribution is this relationship-conditioned evolution of persistent states, beyond cluster-level amplification or compensation performed after the returning client's primal solve.

The reported evidence supports these design choices within a shared framework. At matched β=0.2\beta=0.2, Table 3 gives 51.81% versus 46.16% for mean completion. Table 2 gives 48.25% for virtual-memory correction alone versus 51.81% for the coupled rule. Table 4 gives 51.81% for continued correction versus 40.33% when correction stops after ten rounds. These results establish the value of client-specific relationships, coordinated updates, and continuing compensation.

Against A-FedPDSAM, the closest evaluated SAM-based baseline, CIFAR-100 accuracy improves from 47.84% to 51.81%, while rounds to 40% fall from 627 to 547. Table 7 reports 5.05 versus 4.97 seconds per round and estimated time to target of 46.0 versus 51.9 minutes: a modest per-round increase accompanies fewer required rounds. Table 8 identifies the storage tradeoff explicitly: O(Nd)O(Nd) virtual memories dominate, with O(N2)O(N^2) graph storage. The method requires no all-client communication for graph initialization; the tabulated timings are per-round measurements. Direct accuracy/round/cost comparisons with RFL-NLCP and FedCVCSAM are not included, so the numerical claims concern the evaluated baselines.

2. Normalization separates neighbor relevance from correction strength (W2; Q2). Equation (6) assigns distinct roles to relative weights and total strength:

Aki=Wki/mk,z^k=β∑iAkizi.A_{ki}=W_{ki}/m_k,\qquad \hat z_k=\beta\sum_iA_{ki}z_i.

For supported rows, a common rescaling of the weights leaves the reconstruction unchanged. This is useful because mkm_k combines the number of available neighbors and their learned weights; it is not a calibrated confidence probability. Normalization prevents that raw accumulated mass from automatically setting the correction strength, while β\beta provides explicit control over that strength.

Figure 3 supports moderate compensation strength across both heterogeneity settings. Giving low-mass rows smaller corrections would introduce a different, client-dependent attenuation rule; reducing the shared β\beta scales all inactive corrections uniformly. These are mathematically distinct choices. The reported gains use this separation of relevance and strength; Figure 3 tests shared strength, not row-dependent attenuation.

3. The support-matched control addresses the skipping explanation (W3; Q3). At 2%, Table 3's random graph preserves each inactive row's active-neighbor count, including which rows receive zero correction, yet reaches 37.29% versus 51.81% for the learned graph. The same recipient pattern therefore does not reproduce the gain. Shuffling learned relationships gives 41.74%, further supporting the importance of assigning weights to the appropriate clients.

The fallback is also coupled consistently to the server: ∑i∈StDit=Pt+β∑k∉Stskt\sum_{i\in S_t}D_i^t=P_t+\beta\sum_{k\notin S_t}s_k^t. Thus server amplification reflects the virtual memories actually updated, as required by Eq. (9). Support handling and relationship weighting form one coherent completion rule.

Figure 5(b) demonstrates its end-to-end behavior at lower participation: 36.4% and 48.2% at 0.5% and 1%, versus 3.9% and 18.5% for A-FedPDSAM. These are complete-method comparisons; the support-matched random results are at 2%. Accordingly, Table 3 supplies the controlled evidence that recipient selection alone is insufficient in that setting, while Figure 5(b) establishes the practical value of the full rule under extremely sparse participation.

Thank you for the response and for clarifying the design choices. However, the response still lacks direct comparisons with RFL-NLCP and FedCVC, as well as the ablations. It is hard to tell how much this method improves over existing baselines. Since my main concerns remain, I will keep my original rating.

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

This paper proposes RFL-DFDC, a federated primal–dual method for sparse client participation. It uses a dynamic similarity graph to reconstruct inactive clients’ dual updates and applies consistent server-side corrections, preserving global–local dual consistency. Conditional theoretical analyses and experiments on heterogeneous image and language benchmarks demonstrate improved accuracy and communication efficiency over existing baselines.

Critical strengths
  1. RFL-DFDC coherently combines graph-based reconstruction of inactive clients’ dual increments with matched server-side corrections. This design exactly preserves the global dual as the average of all virtual client duals, independently of reconstruction accuracy.
  2. Beyond comparisons with multiple federated-learning baselines, Table 3 contrasts mean compensation, last-update cosine similarity, and the proposed dynamic graph under a fixed participation rate β=0.2\beta=0.2, providing strong evidence for the effectiveness of the graph-based design.
  3. The paper addresses clients without supported neighbors, clarifies the conditions under which reconstruction is applicable, and reports computation and storage costs, identifying virtual dual memories as the primary scalability bottleneck.
Critical weaknesses

1.The novelty beyond existing similarity-based compensation methods is not yet sufficiently established. [A-FedPD/A-FedPDSAM (NeurIPS 2024)] already includes virtual dual updates for inactive clients and server correction based on the population-average dual. SAFARI (2022) also maintains client similarity records and substitutes missing models with the models of the most similar active clients. The present method combines weighted multi-neighbor completion, cosine–EMA relationships, and persistent dual memories. However, its distinction from SAFARI and the additional value of these choices have not been directly validated. In particular, the dual increments satisfy zi=−αΔwiz_i=-\alpha\Delta w_i, establishing a direct correspondence between dual-increment completion and weighted model-update completion at the level of a single round’s update signals. The more substantive distinction should therefore concern how persistent states affect subsequent training. I would like to see a clearer account of these methodological differences and a nearest-neighbor increment-substitution control within the same primal–dual framework. This would directly affect my assessment of whether the contribution goes beyond a conventional extension of existing ideas.

2.Validation of the relationship metric and completion accuracy still relies primarily on final training outcomes. Table 3 supports the dynamic-graph method over mean compensation in the reported settings. However, it does not directly establish that the benefit comes from more accurate reconstruction of missing increments, or that cosine is preferable to alternative metrics. Cosine primarily captures directional agreement, whereas reconstructing full dual-increment vectors also involves magnitude differences. The paper does not report comparisons with kernel weights constructed from unnormalized L2 distances. Moreover, last-update cosine and the dynamic graph differ simultaneously in temporal alignment, historical smoothing, and available relationships. This comparison therefore does not isolate the effects of these factors.

3.The choice of cosine similarity is insufficiently justified. Cosine similarity captures directional agreement but ignores magnitude differences, which may be important when reconstructing full dual-increment vectors. Comparisons with alternatives such as kernels based on unnormalized ℓ2\ell_2 distances would help establish whether cosine similarity is the appropriate metric. Moreover, the last-update cosine and dynamic-graph variants differ simultaneously in temporal alignment, smoothing, and relationship coverage, so the current ablation does not isolate these factors.

4.The convergence analysis relies on a graph-reconstruction error assumption that is not verified for the proposed cosine–EMA graph during training. In addition, the full-state stability result assumes fixed graph support and a minimum neighbor mass, excluding support changes caused by hard thresholding and the discontinuous zero-update fallback.

5.Graph coverage may be inadequate under sparse participation because pairwise relationships are updated only when two clients participate together. Rarely co-participating clients may therefore remain unsupported and repeatedly receive zero-update compensation. The paper would benefit from reporting graph coverage, unsupported-client frequency, and neighbor-mass statistics across participation rates.

6.Most experimental results are based on only three random seeds, limiting confidence in the reported variance under stochastic client sampling and sparse participation. Additional runs or statistical significance tests would make the empirical conclusions more reliable.

7.Figures 1 and 2 contain numerous missing or corrupted symbols, making the algorithm and information flow difficult to follow. The notation should be corrected in the final manuscript.

Questions for authors
  1. Beyond the compensation target and weighting rule, what additional benefit arises from maintaining persistent virtual dual states? Within the same local optimizer and primal–dual framework, how does RFL-DFDC compare with a nearest-neighbor baseline that substitutes each inactive client’s increment with that of its most similar active client?

  2. Have cosine-based weights been compared with weights based on unnormalized ℓ2\ell_2 distances while holding co-participation observations, EMA smoothing, normalization, and β\beta fixed? If not, what evidence supports cosine similarity as particularly suitable for this vector-completion task? Please distinguish unnormalized from normalized ℓ2\ell_2 distance.

  3. At selected checkpoints, could the authors compute counterfactual one-step increments for a subset of inactive clients and compare the reconstruction errors of the dynamic graph, mean compensation, last-update cosine, and nearest-neighbor substitution? It would also be helpful to report errors separately for well-supported and unsupported clients.

  4. Can the authors report the graph reconstruction energy and error budget throughout training? Does the proposed cosine–EMA graph empirically satisfy the bounded-error condition required by Theorem 1?

  5. What fraction of client pairs co-participate at least once, and how often do inactive clients receive the zero-row fallback at different participation rates? How does performance vary with graph coverage and client support?

  6. How sensitive is RFL-DFDC to clients crossing the active-neighbor-mass threshold in Eq. (6)? Could the authors evaluate a soft fallback or extend the stability analysis to cover changes in graph support?

  7. Could the authors report results over more random seeds or provide statistical significance tests, particularly under low participation rates where client sampling may introduce substantial variance?

  8. Could the authors correct the missing symbols in Figures 1 and 2 and provide high-resolution versions so that the reconstruction procedure and server-correction equations are unambiguous?

2 comments

Thank you for the detailed review. RFL-DFDC generalizes uniform dual compensation to client-specific reconstruction that coordinates the current server update and a returning client's next local optimization. The existing ablations directly support this mechanism.

1. The contribution concerns state evolution, not renaming the update signal (W1; Q1). Equation (3), zit=−αΔwitz_i^t=-\alpha\Delta w_i^t, identifies equivalent observable signals. What distinguishes the algorithms is how that signal changes subsequent optimization. SAFARI substitutes missing models for aggregation. RFL-DFDC uses each reconstruction to evolve an inactive client's dual input to its next augmented local problem, while incorporating the same change into the current server dual.

Building on A-FedPD's virtual-dual framework, our contribution is to replace its common surrogate with a relationship-conditioned operator. Its rows determine individual memory updates, and its column sums give Dit=1+β∑k∉StAkitD_i^t=1+\beta\sum_{k\notin S_t}A_{ki}^t. Equation (11) explains why changing the memory rule requires matching server coefficients. This turns client relationships into a consistent update of the persistent primal–dual process.

The evidence addresses three distinct aspects:

  • Client assignments matter: Table 3 gives 51.81% for the dynamic graph, versus 46.16% for mean completion at the same β=0.2\beta=0.2, 37.29% for a support-matched random graph, and 41.74% after shuffling learned relationships.
  • Both update paths matter: Table 2 gives 48.25% for memory correction alone versus 51.81% for the coupled rule.
  • Continuing correction matters: Table 4 gives 51.81% with correction throughout training versus 40.33% when it stops after ten rounds.

Together, these controls support the relationship-conditioned, coupled mechanism. The reported controls evaluate the coupled rule, not a same-host nearest-neighbor variant; the contribution concerns the full process.

2. Cosine estimates relevance; reconstruction retains magnitude (W2–W3; Q2–Q3). Equation (7) mixes full, unnormalized increments. Cosine is used to estimate directional relationships, not to replace update vectors by unit vectors. Co-participation supplies same-round observations, and EMA accumulates those observations over time. This separates relationship estimation from correction strength, controlled by β\beta.

Table 3 evaluates the resulting complete rule: dynamic relationships reach 51.81%, versus 48.46% for last-update cosine and 46.16% for matched-strength mean completion. The appropriate supported claim is the effectiveness of this complete observation-and-weighting rule. Raw-L2 and counterfactual-MSE comparisons are outside the reported evaluation; we do not infer cosine optimality or an isolated EMA effect from Table 3.

3. The analysis links imperfect completion to optimization (W4; Q4/Q6). Equation (9) establishes exact consistency independently of reconstruction accuracy. Lemma 1 bounds the induced global perturbation pathwise, including unsupported rows, and Theorem 1 propagates it through explicit host and graph-error budgets. The kernel formulation includes mean completion and allows graph/update dependence. Appendix A.4 states the graph/mean comparison under its energy condition; Appendix A.5 separates prediction error from neighbor dispersion.

These results characterize the consequences of imperfect completion; the selected rule's training benefit is established by the controls above. Numerical error-budget trajectories are not reported. The support-margin restriction applies specifically to Appendix B.2's full-state stability result. Equation (9) and Lemma 1 remain valid for changing kernels and zero rows; B.2's result is conditional on a common support regime. Soft-fallback performance is outside the reported evaluation.

4. Sparse coverage, seeds, and notation (W5–W7; Q5/Q7/Q8). Reconstruction requires a connection to the current cohort, not observation of every pair. Appendix C.4's ideal Bernoulli reference, q=p2/[1−(1−p)N]q=p^2/[1-(1-p)^N], gives pair coverage 1−(1−q)T1-(1-q)^T: approximately 4.26%, 14.88%, and 47.28% after 1,600 rounds at 0.5%, 1%, and 2% participation. These are analytical pair-coverage values; actual row-support and neighbor-mass statistics are not reported. For unsupported rows, Eq. (12) explicitly charges the missing increment, so fallback does not remove its error from the analysis.

The three-run mean/std summaries are descriptive, as specified in Appendix C.3; we make no independence-based significance claim. The empirical case combines cross-task results with the controls above.

For Figures 1–2, the mathematical flow is zit→z^kt=β∑iAkitzit→Ditz_i^t\rightarrow\hat z_k^t=\beta\sum_iA_{ki}^tz_i^t\rightarrow D_i^t. Equations (3), (6)–(9) specify the reconstruction and matching server update unambiguously; clearer figure labels would improve their presentation.

Thank you for your detailed response, which clarifies several aspects of the proposed method and its theoretical analysis. However, my main concerns regarding the novelty of the method, the direct verification of the reconstruction accuracy, and the practical applicability of the theoretical assumptions have not yet been fully addressed. After carefully considering the author's response as well as the comments of the other two reviewers, I recognize the encouraging empirical results of the paper, but have not found sufficient evidence to change my assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.