{"id":"baef3f33-426d-42aa-b18e-b5a053e6e8db","arxiv_id":"2411.18905","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FedRGL combines dual-consistency noise filtering, contrastive pseudo-labeling, and entropy-based aggregation to make federated graph learning robust to label noise.","lead":"FedRGL is a new method for training graph models in federated settings when some client labels are wrong. It filters noisy nodes using the global model and graph structure, adds pseudo-labels via contrastive learning, and reweights client models by predictive entropy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Aggregation weights are computed on test-node entropy; the SOTA claim may be an artifact of transductive test-set access rather than robust noise filtering.","rationale":"The reader's weakest assumption about per-class dynamic thresholds is plausible and worth testing, but the empirical SOTA claim is more directly threatened by the test-node dependence of Eqs. (13)--(14). Even if the thresholds perfectly separate clean from noisy nodes, the comparison against baselines that never use test-node entropy is not apples-to-apples. Transductive GNN baselines already use graph structure that includes test nodes through message passing, but none uses test-node predictions to reweight aggregation; that is a FedRGL-specific extra degree of freedom. The paper's transductive footnote explains why features and structure are known during training, but it does not justify using test-set predictions for model selection. A clean ablation with validation-only entropy would disentangle the robust-filtering benefit from test-set adaptation. If the advantage persists, the method's claim is supported; if it collapses, the headline result is an artifact. The reader's conditional verdict is therefore appropriate, and the missing control experiment should be a stated condition for acceptance.","tokens_in":16400,"tokens_out":7999,"duration_ms":83843,"concrete_test":"Re-run the noisy-label experiments of Table 1 with FedRGL modified so that H_m in Eq. (13) is computed only on the validation split V^Va_m, leaving all hyperparameters and other training components identical. Also, if feasible, give the strongest baseline (e.g., CRGNN) the same validation-only entropy reweighting so the comparison is symmetric. If FedRGL no longer ranks first in a majority of the noisy cells, or if its margin over the best baseline shrinks substantially, the reported SOTA claim is not robust to the removal of test-node-dependent aggregation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (13) defines the client-quality signal H_m on V^U_m = V^Va_m ∪ V^Te_m, explicitly including test nodes, and Eq. (14) uses 1/H_m to reweight client models on the server. The 12 baselines are not given this mechanism, so FedRGL has a unique per-round access to test-node predictions when choosing aggregate weights, while the same test nodes are later used for the reported accuracies. This is not label leakage, but it is test-set-dependent model selection: the server can favor models that happen to be confident on the held-out nodes. The claimed large margins in Tables 1--3 (e.g., Cora uniform 78.75 vs CRGNN 61.67) could therefore be driven by the reweighting rule rather than by the class-aware dual-consistency filtering. The paper reports no control experiment without V^Te_m in H_m, and its ablation removes the reweighting component entirely rather than isolating the test-node dependence. Since the central claim is 'consistently outperforms existing methods,' this confound is load-bearing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FedRGL addresses label noise in subgraph federated graph learning. On each client, it filters noisy training nodes using an intersection of two views: a global-model softmax cross-entropy loss thresholded per class by mean plus phi_1 times the standard deviation (Eqs. 1-4), and a label-propagation-based soft-label loss thresholded by class mean plus phi_2 times the standard deviation (Eqs. 5-6). Filtered noisy nodes receive pseudo-labels from two graph augmentations under a confidence threshold gamma, and local training combines clean cross-entropy, contrastive, pseudo-label, and JS-consistency losses (Eq. 12). Clients upload predictive entropy of unlabeled nodes (Eq. 13), and the server reweights aggregation by inverse entropy (Eq. 14). Experiments on Cora, CiteSeer, PubMed, CS, Physics, Photo, and ogbn-arxiv compare against 12 baselines, with ablations and hyperparameter studies, and report state-of-the-art accuracy under uniform and pair label noise.","tokens_in":16701,"tokens_out":5870,"duration_ms":55889,"significance":"If the empirical claims hold, FedRGL would be a valuable first dedicated method for robust federated graph learning under label noise, and the experimental breadth is a genuine strength: six datasets plus a large OGB graph, twelve baselines, ablations, and a hyperparameter sensitivity study. However, the central SOTA claim is weakened by a possible transductive test-set confound in the aggregation rule, by the limited statistical power of three seeds, and by the absence of direct validation of the core noise-filtering assumption. With a clean control experiment and some added diagnostics, the contribution would be substantial and publishable; in its current form the main claim is not fully supported.","major_comments":[{"comment":"The aggregation weights are computed on test-node predictive entropy: Eq. (13) defines H_m over V^U_m = V^Va_m union V^Te_m, and Eq. (14) uses 1/(H_m+epsilon) to reweight client models on the server. The reported accuracies in Tables 1-3 are evaluated on those same test nodes. This is not label leakage, but it is a test-set-dependent model-selection mechanism that is not available to any of the 12 baselines: at every round the server can favor models that happen to be confident on the held-out nodes. The large margins (e.g., Cora uniform 78.75 vs. CRGNN 61.67 in Table 1) could therefore be driven by this reweighting rule rather than by the class-aware dual-consistency filtering. The paper reports no control experiment in which V^Te_m is removed from H_m, and the ablation removes the reweighting component entirely rather than isolating the test-node dependence. This is load-bearing for the central claim of consistent SOTA performance, and the authors should either provide such a control or revise the claim accordingly.","section":"Methodology, Eqs. (13)-(14)"},{"comment":"The empirical claim 'consistently outperforms existing methods' rests on only 3 random seeds, with hyperparameters tuned by Optuna on each dataset. No significance tests are reported, and several margins in Tables 1-3 are within the reported standard deviations. Moreover, the manuscript does not state whether Optuna is optimizing validation accuracy or test accuracy; if test accuracy is used, this adds another layer of test-set selection. The authors should report more seeds, use paired or corrected significance tests, and clarify the Optuna objective.","section":"Experiment: Implementation Details and Tables 1-3"},{"comment":"The core filtering step assumes that, within each class, the loss distribution is sufficiently separable that a single mean-plus-phi-sigma threshold identifies noisy nodes. This assumption is not directly validated: the paper provides no diagnostic of the overlap between clean and noisy loss distributions, no precision/recall of the filter, and no report of how many nodes are filtered per client. The ablation in Fig. 4(b) shows that both views are needed for good accuracy, but it does not establish that the thresholds are identifying noise rather than discarding high-loss clean nodes or retaining confident noisy nodes. Given that the whole method degrades if this assumption fails, some empirical evidence about the filtering behavior is needed.","section":"Methodology, Eqs. (1)-(6)"},{"comment":"The manuscript repeatedly refers to an Appendix for the algorithm pseudo-code, detailed hyperparameter settings, baseline implementations, noise settings, and additional validation, but no Appendix is present in the submitted text. This missing material is necessary for reproducibility and for verifying claims such as the adaptation of baselines to the federated subgraph setting. The authors should supply the Appendix or move the omitted details into the main text.","section":"Appendix references throughout"}],"minor_comments":[{"comment":"The dataset name is spelled 'obgn-arxiv' in Table 3 and in the text, but the correct OGB name is 'ogbn-arxiv'; the table header also contains the typo 'Clinets' instead of 'Clients'.","section":"Table 3 and accompanying text"},{"comment":"In the Cora section, the FedProx row reports the identical value 47.98 +/- 0.42 for both Normal and Uniform noise, which appears to be a copy-paste error and should be corrected.","section":"Table 1, Cora row for FedProx"},{"comment":"The contrastive loss formula in Eq. (8) has corrupted notation: the denominator mixes indices j and i inconsistently, and the second summation uses Z^2_j while the original definition of L_cl is unclear. Please rewrite the equation with consistent index names and explicit positive and negative pairs.","section":"Eq. (8)"},{"comment":"Several figure captions and labels contain rendering artifacts (e.g., 'C_Cm' in Fig. 2 and the garbled axis labels in Fig. 4), making the graphs hard to read; the figures should be regenerated with clean fonts.","section":"Figures 2, 4, and 5"},{"comment":"The OOM entries for CRGNN at 20 clients should be explained; if CRGNN runs out of memory on ogbn-arxiv, it is unclear how the reported average improvements are computed, and the absence of this baseline should be noted explicitly.","section":"Table 3, CRGNN row"}],"recommendation":"major_revision","confidential_remarks":"The pivotal issue is the transductive use of test nodes in the entropy-based reweighting. If the authors can show that the gains persist when H_m excludes V^Te_m, or that a strong baseline benefits equally from the same reweighting scheme, I would view the paper favorably after minor-to-moderate revisions. The three-seed evaluation and Optuna tuning are secondary but should also be addressed. I would not recommend rejecting on the filtering-assumption concern alone, since an ablation and diagnostic study can mitigate it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, two things you should know about this paper. First, it is the first method I have seen that directly targets label noise in federated subgraph learning, and the combination of ideas is genuinely new. Second, the headline result has a comparison-fairness problem that needs a control experiment before I would trust the reported margins.\n\nWhat the paper does well: it assembles known components into a plausible pipeline. Client-side, it filters noisy nodes by intersecting two views: a global-model loss thresholded per class, and a label-propagation soft-label loss thresholded per class. It then uses graph contrastive learning to assign pseudo-labels to filtered noisy nodes, with a Jensen-Shannon consistency term. Server-side, it reweights client models by inverse predictive entropy. The experiments are broad: six datasets plus ogbn-arxiv, 12 baselines, ablations, and hyperparameter sensitivity. That is solid empirical labor, and the gains under noise are large, e.g., Cora uniform 78.75 vs. CRGNN 61.67.\n\nThe soft spot: Eq. (13) defines the client-quality signal H_m on unlabeled nodes V^U_m = V^Va_m ∪ V^Te_m, explicitly including test nodes, and Eq. (14) uses that to reweight client models. The 12 baselines do not have this mechanism. This is not label leakage, but it is test-set-dependent model selection: the server can favor models that happen to be confident on the held-out nodes that are later used to measure accuracy. The reported margins could come from this reweighting rule rather than from the noise-filtering machinery. The ablation removes the reweighting component entirely, but that does not isolate the test-node dependence. I would want a control where H_m is computed only on validation or training nodes, with test nodes excluded. If the margin shrinks or vanishes, the claim needs to be reframed. Also minor: only 3 seeds, per-dataset Optuna tuning, no code.\n\nThe central argument is plausible and the dual-consistency filtering is sensible. There is no circularity in the derivation; it is an empirical method with a clean design. But the load-bearing comparison has a confound. I would send this to a serious referee, and ask for that control experiment plus a fairer baseline setup. If the authors provide it and the results hold, this becomes a solid contribution. I would bring it to a reading group to discuss the transductive-reweighting question.","headline":"A genuinely new method for label noise in federated subgraph learning, but the SOTA claim is weakened by test-node-entropy reweighting that baselines don't get.","tokens_in":17123,"tokens_out":2428,"would_cite":false,"duration_ms":24261,"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":"FedRGL makes federated graph learning robust to label noise by fusing two complementary noisy-node views with contrastive pseudo-labeling, and reports consistent wins over 12 baselines across seven graph datasets.","keywords":["federated graph learning","label noise","noise filtering","graph contrastive learning","pseudo-labeling","predictive entropy","robust aggregation","subgraph federated learning"],"falsifier":"Run FedRGL on a client subgraph where the ground-truth noisy-node mask is known, then compare the selected clean set against that mask at the reported threshold settings. If clean and noisy nodes have heavily overlapping per-class loss histograms, the intersection filter will show low precision or recall on noisy-node detection, and the method's accuracy advantage should shrink accordingly. A direct test: on Cora with 10 clients under pair noise at rate 0.3, compute the per-class loss distributions separately for true clean and true noisy nodes and check whether the reported $\\varphi_1,\\varphi_2$ thresholds actually separate them.","tokens_in":1766,"feed_emoji":"🧠","tokens_out":5791,"duration_ms":69071,"temperature":0.7,"pith_summary":"The paper tries to establish that label noise in federated graph learning can be handled by a method that filters noisy nodes using two complementary signals: the global model's per-class training loss and a structure-aware label-propagation loss. It then reuses the noisy nodes by giving them high-confidence pseudo-labels through graph contrastive learning, and reweights client contributions by predictive entropy on unlabeled nodes. The authors report that this method, FedRGL, improves test accuracy over 12 baseline methods under uniform and pair label noise, across different noise rates, client numbers, and proportions of noisy clients. If correct, this would make federated graph learning practical in settings where client-side labels are unreliable, which existing federated label-noise methods built for computer vision do not handle well.","feed_headline":"Method beats 12 baselines on noisy federated graphs","feed_subtitle":"A dual-perspective filter and contrastive pseudo-labels keep GNN accuracy high under label noise.","key_machinery":"The key machinery is the class-aware dual-consistency filter: per-class dynamic thresholds $\\rho^c_m = t^c_m + \\varphi_1\\sigma^c_m$ on global-model cross-entropy losses and $\\mu^c_m = t^c_m + \\varphi_2\\sigma^c_m$ on losses computed against structure-propagated soft labels, where $t^c_m$ and $\\sigma^c_m$ are the per-class mean and standard deviation of losses in client $m$. The clean training set is the intersection $V^C_m = V^{C1}_m \\cap V^{C2}_m$ of the two views. This intersection, combined with high-confidence pseudo-labels from graph contrastive views and server-side aggregation reweighting by predictive entropy, carries the argument: it is what lets the method identify noisy nodes without a clean-label prior and prevents noisy clients from dominating the global model.","core_discovery":"FedRGL's central claim is that label noise in subgraph federated learning can be tamed by combining two independently computed noisy-node filters: one that uses the global model's per-class cross-entropy losses with class-aware dynamic thresholds, and one that uses soft labels obtained from masked label propagation on the local subgraph, also thresholded per class. The final clean set is the intersection of the two filtered sets. Noisy nodes are not discarded; instead, graph contrastive augmentations generate confident pseudo-labels for training and a Jensen-Shannon consistency term stabilizes them. On the server side, client model quality is estimated by predictive entropy on unlabeled nodes, and aggregation weights are inversely proportional to this entropy. The paper reports that this design consistently outperforms 12 baseline methods on Cora, CiteSeer, PubMed, CS, Photo, Physics, and ogbn-arxiv, and that it matches or surpasses existing subgraph FGL methods even when labels are clean.","pith_inferences":["Editorial inference: the intersection of the two filters likely trades recall for precision, so the method may be conservative when clean and noisy loss distributions overlap; a probabilistic or learned combination could recover some discarded clean nodes.","Editorial inference: the predictive-entropy reweighting might also detect Byzantine or adversarially corrupted clients, not only noisy-label clients, since both tend to produce high-uncertainty predictions on local unlabeled nodes.","Editorial inference: because the entropy signal relies on transductive unlabeled nodes, applying FedRGL to inductive or graph-level federated tasks would require an alternative model-quality estimator, such as entropy on held-out clean anchors.","Editorial inference: the same dual-view filtering signals could be used to estimate each client's local noise rate, enabling personalized or noise-rate-aware local training schedules rather than a shared global threshold schedule."],"forward_implications":["If FedRGL is correct, federated graph learning can tolerate heterogeneous client-side label noise without requiring a clean public dataset at the server.","The class-aware threshold mechanism provides a template for handling class imbalance during noise filtering in non-IID subgraphs, where a single global threshold fails.","Predictive entropy on local unlabeled nodes can serve as a data-free proxy for model quality in aggregation, reducing the influence of badly trained clients.","Contrastive pseudo-labels let noisy nodes contribute to training without overwriting original labels, which the paper argues avoids error accumulation across rounds.","The reported ogbn-arxiv results suggest the method scales to large graphs where existing federated label-noise approaches show little or no gain."],"supporting_citations":[{"why":"Supplies FedCorr, the baseline the paper compares against and the source of the uniform noise-rate sampling assumption.","marker":"Xu et al. 2022"},{"why":"Supplies RHFL, a loss-level federated label-noise baseline the paper must outperform.","marker":"Fang and Ye 2022"},{"why":"Supplies FedNoro, a federated label-noise baseline using class imbalance and distance-aware aggregation.","marker":"Wu et al. 2023"},{"why":"Supplies FedNed, a negative-distillation baseline that the paper positions against.","marker":"Lu et al. 2024"},{"why":"Supplies CRGNN, the graph label-noise method whose uniform and pair noise settings are adopted and which is a strong baseline in the experiments.","marker":"Li et al. 2024b"},{"why":"Supplies the subgraph masking technique and the Earse baseline for structure-aware label noise handling.","marker":"Chen et al. 2023"},{"why":"Supplies FedTAD, the subgraph FGL method whose Louvain partitioning, architecture, and hyperparameter settings are followed.","marker":"Zhu et al. 2024"},{"why":"Supplies the GRACE edge-drop and feature-masking augmentations used to build the two graph contrastive views.","marker":"Zhu et al. 2020b"},{"why":"Supplies RTGNN, a centralized graph noise-resistance baseline included in the comparison.","marker":"Qian et al. 2023"}],"fun_headline_variants":["Dual-perspective filtering beats 12 baselines on noisy graphs","Intersection-based noise filter lifts federated graph learning","Contrastive pseudo-labels tame label noise in federated graphs","FedRGL: dual-filter strategy outperforms on noisy graph data","Robust federated learning on graphs: two filters, 12 baselines beaten"],"cache_read_input_tokens":19328,"weakest_assumption_plain":"The load-bearing premise is that, for each class in each client, the loss distribution of cleanly labeled nodes can be separated from that of noisy nodes by a threshold at the class mean plus a fixed multiple of the class-loss standard deviation; if the distributions overlap, the filter either keeps noisy nodes or discards clean ones, and the whole method degrades.","fun_headline_variants_meta":{"raw":{"variants":["Dual-perspective filtering beats 12 baselines on noisy graphs","Intersection-based noise filter lifts federated graph learning","Contrastive pseudo-labels tame label noise in federated graphs","FedRGL: dual-filter strategy outperforms on noisy graph data","Robust federated learning on graphs: two filters, 12 baselines beaten"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000848,"raw_usage":{"total_tokens":3672,"prompt_tokens":909,"completion_tokens":2763,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":2671}},"tokens_in":525,"tokens_out":2763,"duration_ms":17642,"temperature":1.0,"reasoning_tokens":2671,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:45:22.587340+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FedRGL on a client subgraph where the ground-truth noisy-node mask is known, then compare the selected clean set against that mask at the reported threshold settings. If clean and noisy nodes have heavily overlapping per-class loss histograms, the intersection filter will show low precision or recall on noisy-node detection, and the method's accuracy advantage should shrink accordingly. A direct test: on Cora with 10 clients under pair noise at rate 0.3, compute the per-class loss distributions separately for true clean and true noisy nodes and check whether the reported $\\varphi_1,\\varphi_2$ thresholds actually separate them.","supporting_citations":[{"cited_title":"Z.; and Wu, J","cited_arxiv_id":null,"evidence_quote":"Supplies RTGNN, a centralized graph noise-resistance baseline included in the comparison."}],"review_version":1}