{"id":"cc28f127-abfd-4662-ba26-4cab069533cf","arxiv_id":"2411.11020","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"LEGNN trains noise-resistant GNNs by ensembling high- and low-probability predicted labels from randomly masked neighbor views, improving accuracy while cutting complexity relative to similarity-based reliable labeling.","lead":"This paper introduces LEGNN, a graph neural network training method that keeps working when some node labels are wrong by gathering many candidate labels instead of trusting any single label. It reports better accuracy and lower computation than prior noise-robust GNN methods, including on graphs with over one hundred thousand nodes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The label-ensemble guarantee rests on a homophily assumption and a toy voting model that the actual algorithm does not implement; heterophilous graphs could put true labels into the negative set and reverse the reported gains.","rationale":"The reader's weakest assumption is exactly where my concern sits: the method's success depends on neighborhoods being informative about a node's true label. I agree with that identification, and I sharpen it by pointing out that the paper's only formal intuition, the majority-vote bound in Section IV-A, does not describe what LEGNN actually does. LEGNN does not vote over neighbor labels; it collects argmax and argmin classes from multiple masked GNN runs. Those runs are correlated, so the bound's independence assumption is unavailable. On heterophilous graphs the true label may be systematically excluded from Yp and included in Yn, where the bidirectional loss would penalize it. Since all experiments use homophilous citation networks or a KNN graph built from ImageNet features, this failure mode is never observed. The reader's verdict of CONDITIONAL is appropriate: the central claim is plausible and the experiments are consistent with it, but the support is conditional on homophily and on the candidate sets behaving as intended. I am not moving the verdict because the paper does not explicitly claim to handle heterophilous graphs, and the empirical evidence on the tested homophilous/KNN settings is fairly strong. The requested heterophilous benchmark would settle whether the scope limitation is real. I also note secondary reproducibility concerns, the undefined Eq. (11) denominator and the absent GNN Cleaner baseline, but they are less load-bearing than the untested homophily dependence for the central claim of robustness.","tokens_in":21562,"tokens_out":10503,"duration_ms":187048,"concrete_test":"Run LEGNN with the released code on heterophilous benchmarks (Chameleon, Squirrel, Actor, Texas) under Sym-20/40/50 and Pair-40 noise, with the same hyperparameter grid (mask rate 0.1-0.9, mask count 5-25). Report accuracy versus GCN, NRGNN, and RTGNN, and also measure per-node true-label containment in Yp and Yn. If the accuracy advantage disappears or the fraction of nodes whose true label lands in Yn is substantial, the homophily dependence and the negative-supervision risk are confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ensembling high- and low-probability labels from randomly masked neighbor contexts is more robust than reliable labeling and cheaper than O(N^2). For this to hold, the high-probability candidate set Yp must contain the true label for enough nodes. The only argument supplied is the majority-vote bound in Section IV-A, which assumes each neighbor's label is erroneous independently with probability alpha <= 0.5 and that one aggregates and votes over neighbor labels. LEGNN does neither: it runs the same GNN on Me masked graphs and collects every argmax softmax class into Yp (Eqs. 6-7) and every argmin class into Yn (Eqs. 8-9). The Me masked predictions are strongly correlated because they share the same model and mostly overlap; the independence assumption is not met, so no analogous bound applies. Under heterophily, masked contexts can be consistently wrong, so the true label can be absent from Yp and simultaneously appear in Yn; then Ln = Lp(O-Z,Yn) (Eq. 12) actively penalizes the correct class by treating 1-Z_j as a confidence weight. The method has no mechanism to detect or recover from this. All six datasets are homophilous citation graphs or a KNN graph built from ImageNet features, so this failure mode is untested. The 'superior performance' conclusion is therefore only established for homophilous/KNN graphs, and the mechanism that explains the gains is not robustly supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LEGNN, a label-noise-resistant GNN training method that reframes noisy-label node classification as a partial label learning (PLL) task. Instead of constructing a single reliable label, LEGNN generates, for each node, a high-probability label set and a low-probability label set from predictions on several randomly edge-masked versions of the graph, and trains the GNN with a weighted bidirectional loss. The authors claim that this label-ensemble approach avoids the O(N^2) cost of similarity-based reliability assessment while achieving state-of-the-art accuracy on noisy node classification benchmarks. Experiments are reported on Cora, Citeseer, Pubmed, IGB, OGBN-Arxiv, and a KNN-constructed Clothing1m graph, under symmetric and pair noise, plus an analysis of label precision/recall and runtime.","tokens_in":21847,"tokens_out":4113,"duration_ms":46271,"significance":"If the claims hold, LEGNN is a meaningful contribution: it offers an alternative to reliable-labeling methods, is conceptually simple, and the reported speedups over NRGNN/RTGNN are substantial (e.g., Table III). The paper also provides a useful empirical analysis showing that higher recall of true labels can compensate for lower pseudo-label precision (Table VI and VII). The code is publicly available, and the experiments cover large-scale graphs (IGB, OGBN-Arxiv), which strengthens the scalability claim. The main weakness is that the theoretical motivation in Section IV-A is a toy model that does not match the actual algorithm, and the evaluation is restricted to homophilous graphs, so the generality of the 'superior performance' claim is not yet established.","major_comments":[{"comment":"The weighting function in Eq. (11) is printed as w(Zi, Yp_ij) = Zij Yp_ij / sum_k Zik Yp_ij. The denominator is independent of the summation index k, and when Yp_ij = 0 the denominator is zero, making the weight undefined. This is presumably a typo for sum_k Zik Yp_ik, which is the standard PLL weighting (e.g., Feng et al. [24]). Because this equation defines the core loss, it must be corrected and the surrounding text updated to match.","section":"Eq. (11)"},{"comment":"The theoretical justification for bootstrapped neighbor masking is a majority-vote bound that assumes each neighbor's label is erroneous independently with probability alpha <= 0.5. LEGNN does not perform such a vote: it takes argmax/argmin predictions from the same GNN on Me masked graphs, and those predictions are strongly correlated because they share the same model and largely overlapping neighborhoods. The independence assumption is therefore not satisfied, and the cited argument does not establish that the gathered high-probability set contains the true label. The paper should either provide an analysis that matches the actual algorithm or clearly label the discussion as an intuition only, without claiming a noise-reduction guarantee.","section":"Sec. IV-A"},{"comment":"LEGNN explicitly relies on the homophily assumption stated at the start of Section IV-A. Under heterophily, or when the constructed KNN graph is uninformative, the true label can be systematically absent from the high-probability set and present in the low-probability set; Eq. (12) would then actively penalize the correct class. All six benchmarks are homophilous citation networks or a KNN graph built from ImageNet features, so this failure mode is untested. The conclusion that LEGNN achieves 'superior performance' should be qualified to the homophilous setting, or the authors should add experiments on heterophilous benchmarks (e.g., Chameleon, Squirrel, Actor) and discuss how the method behaves when the assumption is violated.","section":"Sec. IV-A and Sec. V"},{"comment":"The claimed additional complexity O(Me(1+Ld)|E| + (MeLd^2+2C)N) omits the cost of computing argmax and argmin over C classes for every node and every masked graph, which is O(Me N C). This does not change the asymptotic advantage over O(N^2) when C is a constant, but the stated formula is incomplete and should be corrected. Please also clarify whether the 2CN term accounts for the loss computation only; as written, the argmax/argmin cost is missing.","section":"Sec. IV-E"}],"minor_comments":[{"comment":"The notation for the gathered label sets is inconsistent: Eq. (6) defines \\hat y_k^i, while Eqs. (7) and (9) use \\hat y_k^i and \\bar y_k^i with the indices in different positions. Please unify the superscript/subscript convention throughout.","section":"Eqs. (7) and (9)"},{"comment":"The subsection beginning 'Discuss: why NRGNN and RTGNN perform worse than GCN in Pubmed and OGBN?' is written in an informal Q&A style and interrupts the flow of the experimental section. It should be integrated into the main text or moved to an appendix with a proper framing.","section":"Sec. V-C"},{"comment":"Algorithm 1, step 6, says 'if validation classification accuracy declined' without specifying the threshold, the frequency of checking, or how the noisy validation set is used. This condition is underspecified and should be described precisely, since it controls when the label gathering is repeated.","section":"Alg. 1"},{"comment":"The caption states 'The experimental details of (b) are inline with Section V-E', but Section V-E does not contain a description of Figure 1(b). Please either add the referenced details or remove the cross-reference.","section":"Fig. 1 caption"},{"comment":"There are several typos and minor wording issues, e.g., 'exibit imitations' (Sec. II-A), 'diagnal matrix' (Eq. 3), and 'PPL' vs 'PLL' (Sec. III-C). These should be cleaned up.","section":"General"},{"comment":"Reference [43] is cited for the complexity of an L-layer GCN, but the title of that reference is about molecular property prediction; it is unclear that this is the intended source. Please cite a standard GCN complexity reference instead.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution with a clear practical message, but the theoretical framing in Section IV-A does not match the actual algorithm, and the scope is limited to homophilous graphs. The Eq. (11) typo and the complexity omission are easily fixable. I recommend major revision rather than rejection because the core empirical claim is well supported on the tested benchmarks, but the manuscript needs to correct the technical errors and substantially qualify or expand the claims before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the central idea is genuinely new: reframing noisy-label node classification as partial label learning and gathering high- and low-probability label sets from randomly masked neighborhood contexts. That avoids the O(N^2) similarity costs of NRGNN/RTGNN, and the complexity analysis is honest. Second, the empirical work is stronger than the theory. LEGNN beats the compared baselines on most settings across six datasets, and the time comparison in Table III shows a real efficiency gain. The label-quality analysis in Table VI is a genuinely useful contribution; it shows the method trades precision for recall and wins on F1. That is the right way to think about pseudo-label construction.\n\nNow the soft spots. Biggest technical issue: Eq. (11) as printed is degenerate. The denominator sums Z_{ik} times Y^p_{ij}, which is independent of k, so the weight collapses to Z_{ij} (or zero). It has to be sum_k Z_{ik} Y^p_{ik}. As written, the core loss is undefined. That is likely a typo, but it must be fixed. Second, the theoretical motivation in Section IV-A assumes independent neighbor errors and a voting process, but the actual algorithm runs the same GNN on masked graphs and takes argmax/argmin of softmax outputs. Those predictions are correlated, so the majority-vote bound doesn't directly transfer. Not fatal — this is an empirical method paper, not a guarantee paper — but the justification is weaker than it looks. Third, the method leans on homophily. Under heterophily, the true label can end up in the low-probability set, and Eq. (12) would penalize it. All six datasets are homophilous or KNN-built, so that failure mode is untested. Also, GNN Cleaner is cited but not included in the comparisons; it should be, since it is a direct relative.\n\nFor a subfield that needs cheaper noise-robust methods, this is a solid contribution. It deserves a serious referee, but the formula typo and missing baseline need to be handled before publication. I would send it to review with a request for a clean revision.","headline":"A well-executed empirical method paper worth refereeing, with a misprinted core equation and an untested homophily assumption.","tokens_in":22373,"tokens_out":2532,"would_cite":true,"duration_ms":27396,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that graph node classification under label noise can be made both more accurate and cheaper by ensembling many candidate labels from randomly masked neighbor views instead of certifying one reliable label per node.","keywords":["graph neural networks","label noise","node classification","partial label learning","label ensemble","semi-supervised learning","noisy labels","graph scalability"],"falsifier":"On a graph whose edges predominantly connect nodes of different classes (high heterophily), directly measure the fraction of nodes whose true label actually appears in the high-probability label set under 50% symmetric label noise; if that fraction is near chance while accuracy still improves, the homophily mechanism is not the source of the gain.","tokens_in":21370,"feed_emoji":"🏷️","tokens_out":6527,"duration_ms":67998,"temperature":0.7,"pith_summary":"LEGNN claims that noisy-label node classification on graphs is better framed as collecting a set of plausible labels per node than as deciding one reliable label. Instead of computing pairwise node similarities to filter noise, it randomly masks each node's neighbors several times, lets the current GNN label each masked view, and unions the top predictions into a high-probability label set plus a symmetric low-probability set. A partial-label-learning weighted loss trains the GNN from these sets, so a single mistaken pseudo-label cannot by itself steer training. The paper reports that this label-ensemble view beats reliable-labeling methods such as NRGNN and RTGNN on six datasets, with the largest gains at high noise rates, while adding only near-linear-to-subquadratic cost rather than $O(N^2)$.","feed_headline":"Label ensembles beat pairwise similarity for noisy-graph GNNs","feed_subtitle":"LEGNN gathers labels from masked neighbor views and trains with partial-label loss, beating reliable-labeling GNNs.","key_machinery":"The load-bearing object is the bootstrapped neighbor-context label ensemble, built on top of partial label learning, a setting in which each training example is given a set of candidate labels known to contain the true label. A random edge mask with rate $K$ is applied $M_e$ times to create diverse graph views; the current GNN labels each view, and the ensemble keeps every label that ever appears as an arg-max (high-probability set) and every label that ever appears as an arg-min (low-probability set). The second piece is the weighted bidirectional loss, which trains on both sets at once and makes the method, in the authors' account, shift from precision-oriented single-label correction to a precision-recall-balanced ensemble.","core_discovery":"Central claim: the reason previous noise-robust GNNs degrade at high noise is that they sacrifice recall for precision when they assign one reliable pseudo-label, and LEGNN can have both by ensembling labels across bootstrapped neighbor contexts. In LEGNN, the model first predicts each node under $M_e$ random edge masks; for each node the arg-max predictions form the high-probability multi-label set and the arg-min predictions form the low-probability multi-label set. Training then minimizes a weighted bidirectional loss, $L(Z,Y^p,Y^n)=L_p(Z,Y^p)+L_p(O-Z,Y^n)$, where the partial-label-learning weighting down-weights low-confidence labels and the low-probability term supplies negative supervision. The paper shows on Cora, Citeseer, Pubmed, IGB, OGBN-Arxiv, and a Clothing1m KNN graph that this yields accuracy gains that grow with noise, e.g. 14.55 points on Citeseer and 8.47 points on OGBN-Arxiv at 50% symmetric noise, and that the additional complexity is $O(M_e(1+Ld)|E|+(M_eLd^2+2C)N)$, avoiding the $O(N^2)$ of similarity-based reliable labeling.","pith_inferences":["A natural stress test beyond the paper is heterophily: because the bootstrapping argument assumes neighbors are informative about the true label, LEGNN should be evaluated on graphs where edges mostly connect different classes; the homophily premise predicts a sharp drop.","The random-mask ensemble is essentially a cheap way to create diversity; replacing the fixed arg-max/arg-min rule with sampled labels or temperature-scaled probabilities could interpolate between precision and recall and possibly improve the reported trade-off.","The label-gathering step is only triggered when validation accuracy drops, so the amortized cost is lower than the worst-case complexity suggests; a systematic study of gathering frequency could sharpen the efficiency claim.","If the precision-recall diagnosis is correct, other pseudo-label methods could adopt the same symmetric low-probability negative supervision without any graph-specific machinery, transferring the idea to image or text noisy-label learning."],"forward_implications":["Noisy-label node classification can be handled without expensive pairwise-similarity reliability modules, so robustness scales to graphs with hundreds of thousands of nodes and millions of edges.","Performance degrades more slowly as noise increases; the reported margin over NRGNN and RTGNN is largest at 50% symmetric and 40% pair-flip noise.","The same training wrapper works with GCN, GAT, and GraphSAGE backbones, so it can be dropped onto an existing GNN rather than requiring a bespoke architecture.","Gathered label sets cover more true labels than confidence-threshold selection, which explains why training on noisier candidate sets can still improve accuracy.","The method's additional cost stays below $O(N^2)$ except on nearly complete graphs, since it depends on edge count and label-ensemble size rather than all-pairs similarity."],"supporting_citations":[{"why":"Supplies the GCN backbone whose message-passing and semi-supervised setup LEGNN inherits and extends.","marker":"[8]"},{"why":"NRGNN is the main reliable-labeling baseline and the source of the $O(N^2)$ pairwise-similarity complexity LEGNN avoids.","marker":"[16]"},{"why":"RTGNN supplies the noise-generation protocol, label-rate settings, and the strongest reliable-labeling comparison on most benchmarks.","marker":"[18]"},{"why":"Provides the loss-weighted partial-label-learning strategy that LEGNN adapts into its weighted bidirectional loss.","marker":"[24]"},{"why":"GNN Cleaner is the baseline for the real-world Clothing1m-graph and motivates the KNN graph construction.","marker":"[19]"},{"why":"Defines symmetric label-flip noise and the Forward loss-correction baseline used in the experiments.","marker":"[26]"},{"why":"Defines pair-flip noise and the Co-teaching baseline used in the experiments.","marker":"[27]"},{"why":"Supplies Cora, Citeseer, and Pubmed benchmark graphs used for the synthetic-noise evaluations.","marker":"[44]"},{"why":"Supplies the OGBN-Arxiv large-scale graph used for scalability and high-noise evaluation.","marker":"[46]"}],"fun_headline_variants":["GNNs resist label noise by ensembling neighbor predictions","Avoid O(N^2) with label ensembling for noisy graphs","Bootstrapped label sets boost GNN accuracy under noise","Partial-label training from masked views toughens GNNs","LEGNN: multi-label ensemble for noise-robust GNNs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"LEGNN's advantage depends on homophily: a node's true label must usually be among the labels of its neighbors, so that the randomly masked neighbor views and the model's bootstrapped predictions keep the correct label inside the collected high-probability set.","fun_headline_variants_meta":{"raw":{"variants":["GNNs resist label noise by ensembling neighbor predictions","Avoid O(N^2) with label ensembling for noisy graphs","Bootstrapped label sets boost GNN accuracy under noise","Partial-label training from masked views toughens GNNs","LEGNN: multi-label ensemble for noise-robust GNNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000324,"raw_usage":{"total_tokens":1894,"prompt_tokens":1095,"completion_tokens":799,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":711,"completion_tokens_details":{"reasoning_tokens":710}},"tokens_in":711,"tokens_out":799,"duration_ms":11650,"temperature":1.0,"reasoning_tokens":710,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:00:55.659834+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a graph whose edges predominantly connect nodes of different classes (high heterophily), directly measure the fraction of nodes whose true label actually appears in the high-probability label set under 50% symmetric label noise; if that fraction is near chance while accuracy still improves, the homophily mechanism is not the source of the gain.","supporting_citations":[{"cited_title":"Nrgnn: Learning a label noise resistant graph neural network on sparsely and noisily labeled graphs,","cited_arxiv_id":null,"evidence_quote":"NRGNN is the main reliable-labeling baseline and the source of the $O(N^2)$ pairwise-similarity complexity LEGNN avoids."},{"cited_title":"Provably consistent partial- label learning,","cited_arxiv_id":null,"evidence_quote":"Provides the loss-weighted partial-label-learning strategy that LEGNN adapts into its weighted bidirectional loss."},{"cited_title":"Gnn cleaner: Label cleaner for graph structured data,","cited_arxiv_id":null,"evidence_quote":"GNN Cleaner is the baseline for the real-world Clothing1m-graph and motivates the KNN graph construction."},{"cited_title":"Making deep neural networks robust to label noise: A loss correction approach,","cited_arxiv_id":null,"evidence_quote":"Defines symmetric label-flip noise and the Forward loss-correction baseline used in the experiments."}],"review_version":1}