{"id":"881bb05d-8602-4c1f-b1c1-0586c0ff6bbb","arxiv_id":"2509.07198","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Fed-REACT first trains a shared encoder, then repeatedly clusters clients by smoothed task-model weights, improving federated learning accuracy on heterogeneous, non-stationary time series.","lead":"Fed-REACT combines self-supervised representation learning with evolutionary clustering so federated models keep working as client data shifts over time. If the reported gains hold up, it gives federated learning a practical way to handle both data spread across clients and drift within each client's data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SUMO's 1.3 RMSE is obtained by picking C=50 on the test set (Appendix H), i.e., full personalization; the headline superiority is a test-selected optimum, not a fair comparison.","rationale":"I agree with the reader's weakest assumption. The SUMO result is the largest and most visible margin in Table 1 and the only regression task; it is also the example where the cluster count is explicitly selected by test performance. Appendix H says the 'optimal performance' is achieved at C=50 (complete personalization), so the 1.3 RMSE in the main table is not a performance of the evolutionary clustering algorithm but of a test-set-selected hyperparameter. This makes the abstract's 'superior accuracy and robustness' claim dependent on selection bias. The unlabeled-data asymmetry compounds the issue, but the test-set C-selection is the single decisive problem because it directly invalidates the reported number as an unbiased estimate. Other concerns (missing error bars, theory for Phase 2) are secondary and would not by themselves change the verdict. The appropriate outcome remains conditional acceptance pending the re-analysis: the framework is coherent and the RTD/EEG results are suggestive, but the headline empirical claim needs a honest model-selection protocol. Thus I keep the reader's CONDITIONAL verdict unchanged.","tokens_in":21074,"tokens_out":3425,"duration_ms":41958,"concrete_test":"Rerun the SUMO experiment choosing C by a validation split or the Appendix E elbow/silhouette procedure (not by test-set RMSE), with at least 5 seeds, and report mean±std RMSE for all C∈{1,3,9,25,40,50}; also train a fully personalized baseline (same Phase-1 encoder, per-client linear head) under identical labeled-data access. If the validation-selected C gives RMSE within noise of the best supervised baseline, or the personalized baseline matches 1.3, the headline superiority claim fails as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is 'superior accuracy and robustness' across real-world datasets, with SUMO providing the largest margin (RMSE 1.3 vs 20.1 for the best supervised baseline). But Section 4.1 says C is chosen by searching for the value that gives the best performance, and Appendix H/Table 6 show the reported 1.3 occurs at C=50, which the appendix itself describes as 'complete personalization.' Thus the headline SUMO number is the maximum of a test-set sweep over C, i.e., an optimistic selection rather than an unbiased evaluation. The result is also not a clustered-FL result: with C=50 each vehicle is its own cluster, so evolutionary clustering is not exercised. Moreover, Fed-REACT's Phase 1 consumes all unlabeled data while the supervised baselines see only labeled samples, so part of the gap may be due to data access rather than the method. The manuscript is transparent about this in Appendix H, but Table 1 presents the selected value without this context. Since the strongest evidence of superiority rests on this single test-selected number, the empirical claim is not yet supportable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Fed-REACT, a two-stage federated learning framework for heterogeneous and evolving time-series data. Phase 1 learns a shared representation encoder via self-supervised contrastive learning on local unlabeled data; Phase 2 clusters clients based on task-model weights using evolutionary clustering (AFFECT) with an adaptive forgetting factor, and aggregates cluster-specific task models using either simple temporal averaging or weighted averaging with forgetting. The authors provide a regret-based convergence analysis of a simplified linear SSL objective and report experiments on RTD, EEG, and SUMO, claiming superior accuracy and robustness over supervised and clustered FL baselines.","tokens_in":21395,"tokens_out":8455,"duration_ms":85188,"significance":"If the claims hold, the framework addresses a real deployment gap: federated learning under both inter-client heterogeneity and intra-client distribution drift. The combination of self-supervised representation learning with evolutionary clustering is a plausible and timely idea, and the paper provides concrete algorithms, an extensive appendix, and a theoretical framing. However, the empirical evidence for the headline claim is weakened by a test-set-based selection of the cluster count and by an unaddressed data-access asymmetry with the baselines, while the theoretical proof contains an invalid limit step. The significance is therefore conditional on these issues being resolved.","major_comments":[{"comment":"The headline SUMO result (RMSE 1.3 vs 20.1 for the best baseline) is obtained by selecting the number of clusters C on the test set. Section 4.1 states that 'we search over a range of values to select the one yielding the best performance,' and Appendix H shows that the reported optimum is C=50, which the appendix itself describes as 'complete personalization.' Thus Table 1 reports the maximum of a test-set sweep over C, not an unbiased estimate; moreover, with C=50 the evolutionary clustering component of Fed-REACT is not exercised at all. Since this is the largest margin claimed in the paper, the central empirical claim is not supportable as presented. Please select C on a validation set (or via the Appendix E estimators) and report the corresponding test performance, or present the full sweep as the main result with per-C uncertainty.","section":"Section 4.1 / Appendix H (Table 6)"},{"comment":"The comparison is confounded by data access. Fed-REACT's Phase 1 trains the encoder on all unlabeled local data, whereas the supervised baselines are trained only on labeled samples. With 90/10 train/test splits, the baselines presumably see a subset of the data that Fed-REACT also uses, plus labels; part of the reported improvement may therefore reflect additional unlabeled data rather than the method itself. Please include a semi-supervised or self-supervised baseline trained on the same unlabeled data, or restrict Phase 1 to the labeled training subset, so that the effect of representation learning is isolated.","section":"Section 4.1 / Table 1"},{"comment":"The proof does not establish the stated limit. After substitution, the bound contains the term (32βM/W)·(2−γ^w+γ^{w−1})/(1−γ). As γ→1−, W→w and the numerator tends to 2, so this term diverges as 1/(1−γ); it does not converge to 64βM/W as claimed. The line 'lim_{γ→1−} ||∇S||² ≤ (1/W)(64βM+2σ²)+5/8ε²' therefore does not follow from the preceding algebra. A correct derivation must keep (1−γ^w)/(1−γ) terms together so that the finite limit emerges (the theorem may be true with different constants, but the proof as written is invalid). In addition, the theorem is stated as a bound on the random quantity ||∇S||² without an expectation, while the proof uses conditional expectation; the statement should be in expectation or with high probability.","section":"Appendix C (Proof of Theorem 1)"},{"comment":"No error bars, standard deviations, or random seeds are reported for any experiment. The 'robustness' claim in the abstract and Section 4 cannot be assessed from single-point comparisons, especially when accuracies are reported to three decimals and differences between methods can be as small as 0.01. Please report means and standard errors over at least 3–5 seeds, and also report the variance of the Rand scores for the clustering experiments.","section":"Section 4 (Tables 1–3)"}],"minor_comments":[{"comment":"Notation is inconsistent: the problem setup uses n clients, while Algorithm 1 uses K for the number of clients and also overloads k as a client index. Please unify.","section":"Section 2 / Algorithm 1"},{"comment":"The text says 'Figure 11 demonstrates...' but the referenced figure appears to be Figure 1 (and the caption numbering is otherwise out of order). Please correct the cross-reference.","section":"Section 4.2"},{"comment":"The notation is confusing: ε_proj is defined as a vector, but the bound is written as ||ε_proj||₂ ≤ ε², which mixes a squared scalar with a norm. Please use, e.g., ||ε_proj|| ≤ ε or ||ε_proj||² ≤ ε² consistently.","section":"Assumption 3.1(c)"},{"comment":"The statement that C=50 denotes 'complete personalization' is important context for interpreting the main result, but it is relegated to the appendix. This context should be stated in Section 4.1 when the SUMO result is presented, or the main table should include the selected C value.","section":"Appendix H / Table 6"}],"recommendation":"major_revision","confidential_remarks":"The paper's contribution is potentially interesting, but the current evidence does not support the abstract's strong empirical claim. The test-set selection of C and the absence of uncertainty estimates are fixable with additional experiments, and the theory proof appears to have a local but consequential error that may be repairable. I would encourage the editor to solicit a revised version rather than reject, provided the authors can address the protocol issues and correct the proof."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper combines two known pieces — self-supervised representation learning and AFFECT-style evolutionary clustering — into a federated pipeline for non-stationary, non-IID time-series. The combination is new as far as the cited literature goes, and the RTD/EEG experiments do show Fed-REACT mostly ahead of clustered-FL baselines, with Rand-score curves that recover the ground-truth partitions. The theory is a modest piece: a regret bound for time-smoothed gradient descent on a linear feature model, basically following Aydore et al., with the projection error floor assumed in Assumption 3.1. It is honest about being a simplification.\n\nThe problem is the SUMO headline. Section 4.1 says C is chosen by 'searching over a range of values to select the one yielding the best performance,' and Appendix H shows the reported RMSE of 1.3 is at C=50, complete personalization. So the best number in Table 1 is a test-set maximum, not an unbiased evaluation. At C=1 the same method gets 24.4, which is worse than the best supervised baseline (PatchTST at 20.1). And with C=50 the method is not doing any clustering. The main text hides this context behind the single 1.3. That is the paper's load-bearing empirical claim, so it cannot be taken at face value.\n\nTwo smaller issues compound it. No error bars or seeds anywhere in Section 4. And Fed-REACT's Phase 1 consumes all unlabeled data, while the supervised baselines train only on labeled samples; part of the gap may simply be additional data access. The paper should either match the baselines' data access or acknowledge the asymmetry.\n\nThe RTD and EEG results are the more credible part, though the margins over strong baselines are thin at times. The clustering evaluation in Table 2 is informative, and the adaptive-forgetting update is a reasonable adaptation of Xu et al. The appendix is transparent about the SUMO sweep — the flaw is in presentation: Table 1 states the optimum as if it were an honest result.\n\nWho gets value: anyone working on personalized FL or federated time-series, especially with drifting client distributions. The idea is worth testing. But the paper needs revision before publication: fix the cluster count selection (use a validation split), add seeds and variance, equalize data access, and reframe the SUMO claim. I would not desk reject it; I would send it to reviewers with clear instructions to check the experimental protocol.\n\nRegards.","headline":"Genuinely new combination of self-supervised learning and evolutionary clustering for federated time-series, but the SUMO headline is a test-set-selected optimum at full personalization, so the superiority claim needs an honest rework.","tokens_in":21821,"tokens_out":3478,"would_cite":true,"duration_ms":38242,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fed-REACT claims that separating a shared representation-learning phase from evolutionarily clustered task models handles both heterogeneous and drifting client data, reporting higher accuracy and lower error than supervised federated basel","keywords":["federated learning","self-supervised representation learning","non-IID data","concept drift","evolutionary clustering","time-series classification","client clustering","regret bound"],"falsifier":"Run the SUMO and RTD experiments again with C chosen by a validation set (elbow or silhouette) rather than by best test performance, and give the supervised baselines the same self-supervised encoder pretraining; if Fed-REACT's accuracy or RMSE margin shrinks to baseline levels, its empirical claim is selection- or data-access-driven.","tokens_in":20999,"feed_emoji":"🔄","tokens_out":5845,"duration_ms":68119,"temperature":0.7,"pith_summary":"The paper tries to show that federated learning can cope with two difficulties at once: client data distributions differ from each other, and each client's distribution drifts over time. Its proposed method, Fed-REACT, works in two stages: first, clients jointly learn a shared time-series encoder from unlabeled data using contrastive self-supervised learning; second, the server clusters clients by the similarity of their task-model weights, using evolutionary clustering that smooths cluster assignments across rounds so drift does not cause spurious regroupings. The paper also gives a regret bound for the representation-learning stage and reports better accuracy and lower error than supervised federated baselines on the RTD, EEG, and SUMO datasets. This matters because real deployments in healthcare, driving, and finance routinely face both non-IID and time-varying data.","feed_headline":"Federated learning that evolves client clusters beats static baselines","feed_subtitle":"A shared encoder plus adaptive client clustering handles both non-IID and time-varying local data.","key_machinery":"The load-bearing mechanism is the adaptive evolutionary clustering step, an AFFECT-style update with an adaptive forgetting factor: the server maintains a smoothed similarity matrix ψ^t = a_t ψ^{t-1} + (1 - a_t)W_t, where W_t is the current cosine similarity between clients' vectorized task-model weights and a_t is estimated from the variance and drift of those similarities. This smoothing is what lets cluster memberships persist through transient noise and only change when the underlying distribution genuinely shifts. The representation phase is carried by a contrastive loss on time-series sub-sequences, and its theory uses a time-smoothed gradient descent update whose regret is bounded by","core_discovery":"The paper's central claim is that two forms of heterogeneity should be handled separately: low-level temporal features are sufficiently shared across clients to be learned by one global encoder, whereas task-specific behavior is tied to local distributions and should be learned by per-cluster models whose membership evolves over time. Fed-REACT operationalizes this with a contrastive encoder in phase 1 and, in phase 2, an adaptive evolutionary clustering scheme that estimates a smoothed client-similarity matrix from task-model weights and groups clients with agglomerative hierarchical clustering. The theoretical analysis shows that time-smoothed gradient descent on the encoder's simplified o","pith_inferences":["An implicit consequence of the two-phase split is that part of the accuracy gain could come from the contrastive encoder seeing all unlabeled data, while supervised baselines see only labeled samples; ablating the encoder under identical labeled budgets would separate these two contributions.","Because Appendix H shows that SUMO's best RMSE occurs at C=50, full personalization, the paper's regression claim is effectively about representation quality rather than cluster coordination; similar patterns may appear in other highly heterogeneous tasks.","The adaptive forgetting factor estimates cluster stability from pairwise weight similarities, so in very large client populations the O(K^2) similarity matrix could become a bottleneck, suggesting a batched or sampled similarity estimator as a natural extension."],"forward_implications":["With distribution drift, snapshot clustering misassigns clients; Fed-REACT's smoothed similarity matrix recovers ground-truth clusters within a few rounds and keeps them stable, as measured by Rand score.","The shared encoder is trained once on unlabeled data and reused for lightweight cluster-specific task models, reducing communication: the server sends each client one model per cluster rather than all cluster models.","The regret bound for time-smoothed gradient descent says that with step size 1/β and enough smoothing, the average gradient norm of the smoothed objective is controlled mainly by projection error, not by the drift itself.","The framework extends to partial client participation by reusing the most recent task-model weights for absent clients when computing the similarity matrix.","If the number of clusters is unknown, the paper's experiments show that an elbow method on within-cluster sum of squares can select the correct cluster count."],"supporting_citations":[{"why":"Defines federated averaging and the communication-round setup that all methods build on.","marker":"[McMahan et al., 2017]"},{"why":"Supplies the contrastive time-series loss and the causal dilated CNN encoder used in phase 1.","marker":"[Franceschi et al., 2019]"},{"why":"Contributes AFFECT, the adaptive evolutionary clustering algorithm central to phase 2.","marker":"[Xu et al., 2014]"},{"why":"Gives the contrastive loss form used for self-supervised representation learning.","marker":"[Chen et al., 2020]"},{"why":"Provides the time-smoothed gradient descent regret framework that Theorem 1 extends.","marker":"[Aydore et al., 2019]"},{"why":"Is the clustered federated learning baseline (IFCA) that Fed-REACT must outperform in phase 2 comparisons.","marker":"[Ghosh et al., 2020]"},{"why":"Supports the simplification of the SSL objective to a covariance-matching problem used in the theory.","marker":"[Wang et al., 2022]"}],"fun_headline_variants":["Fed-REACT separates shared encoder from evolving client clusters","Adaptive client clustering tackles time-varying and non-IID data in FL","Two-stage FL: global encoder for features, evolving clusters for tasks","Adaptive clustering in federated learning beats static approaches"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The reported gains depend on an honest comparison: cluster count C must not be tuned on test labels (the paper reports doing so), and the unlabeled encoder data available to Fed-REACT must not be the true source of the advantage.","fun_headline_variants_meta":{"raw":{"variants":["Fed-REACT separates shared encoder from evolving client clusters","Adaptive client clustering tackles time-varying and non-IID data in FL","Two-stage FL: global encoder for features, evolving clusters for tasks","Adaptive clustering in federated learning beats static approaches"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000547,"raw_usage":{"total_tokens":2421,"prompt_tokens":685,"completion_tokens":1736,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":429,"completion_tokens_details":{"reasoning_tokens":1665}},"tokens_in":429,"tokens_out":1736,"duration_ms":15047,"temperature":1.0,"reasoning_tokens":1665,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T22:38:53.054219+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the SUMO and RTD experiments again with C chosen by a validation set (elbow or silhouette) rather than by best test performance, and give the supervised baselines the same self-supervised encoder pretraining; if Fed-REACT's accuracy or RMSE margin shrinks to baseline levels, its empirical claim is selection- or data-access-driven.","supporting_citations":[{"cited_title":"Unsupervised scalable representation learning for multivariate time series","cited_arxiv_id":null,"evidence_quote":"Supplies the contrastive time-series loss and the causal dilated CNN encoder used in phase 1."},{"cited_title":"Dynamic local regret for non-convex online forecasting","cited_arxiv_id":null,"evidence_quote":"Provides the time-smoothed gradient descent regret framework that Theorem 1 extends."},{"cited_title":"An efficient framework for clustered federated learning","cited_arxiv_id":null,"evidence_quote":"Is the clustered federated learning baseline (IFCA) that Fed-REACT must outperform in phase 2 comparisons."}],"review_version":1}