{"id":"c366a0f8-6fc1-489b-a60d-3dd129a08ceb","arxiv_id":"2411.17350","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CorGCN decomposes a multi-label graph into per-label views and propagates label correlations during message passing, achieving the best average rank on five benchmarks.","lead":"This paper proposes CorGCN, a graph neural network that creates a separate graph view for each label and then recombines them with learned label correlations for multi-label node classification. The authors report top average rank among ten models on five benchmarks, but on the largest dataset a baseline beats CorGCN on every metric.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4) is not a contrastive objective as written, so the label prototypes used to build the label-aware graphs in Eq. (9)-(11) lack the claimed discriminative training signal.","rationale":"The reader correctly notes in the rationale that Eq. (4) is not a proper contrastive objective because the denominator includes only positive labels, but the reader's stated weakest assumption is the Top-lambda structure decomposition. I agree that Top-lambda is a sensitive hyperparameter, but the more load-bearing issue is upstream: the label prototypes that define the projections used in the Top-lambda graphs are trained with an objective that, as written, cannot contrast positive labels against negative labels. If that equation is wrong, the central 'correlation-aware' mechanism is not realized as described, even though the final classifier might still work. The reader's Delve observation is also correct: in Table 2, CorGCN loses to LARN on all seven metrics on Delve, so the paper's claim of 'significant improvements over state-of-the-art methods on all experimental datasets' is false as stated. That overclaim independently supports a conditional verdict, but the Eq. (4) issue is more central to the proposed method's validity. A conditional acceptance is appropriate: the authors should correct or clarify Eq. (4), release working code, and rerun the main experiments and ablations with the corrected objective; they should also soften the 'all datasets' claim to reflect the Delve results.","tokens_in":21645,"tokens_out":9600,"duration_ms":103535,"concrete_test":"Analytically compute the gradient of Eq. (4) for a node with positive label set P: for any negative label r not in P, the partial derivative of the loss with respect to s_{i,r} is zero, while for p,q in P the loss pushes s_{i,p} and s_{i,q} toward each other. Then, if the authors' implementation actually uses the full label set in the denominator, rerun Table 2 with the corrected objective and check that the four reported Macro-AUC gains remain within one standard deviation. If the implementation literally follows Eq. (4), replace it with a proper InfoNCE-style loss that includes negative labels in the denominator and rerun; if the Macro-AUC gains shrink substantially, the published mechanism is not what produces the reported results.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central mechanism of CorGCN is the correlation-aware graph decomposition: label prototypes E^l are learned and then used in Eq. (7)-(8) to project node features and in Eq. (10)-(11) to rank neighbors for each label view. The claimed training signal for these prototypes is the contrastive mutual-information loss in Eq. (4). As written, that loss is not contrastive. Its denominator sums only over t in y_i, i.e., over the node's positive labels, so the term for a positive label p is log( exp(s_{i,p}) / sum_{q in Pos(y_i)} exp(s_{i,q}) ). Negative labels never appear. For a node with a single positive label the term is identically zero; for a node with several positives it only pushes the similarities to those positives toward equality. It cannot separate associated from unassociated labels. This matters because if the prototypes are not trained to distinguish labels, the per-label graphs G^k may simply reflect raw feature-similarity structure already present in X, rather than label-relevant structure. The reported gains could then be driven by extra model capacity and the hand-tuned Top-lambda construction rather than by the claimed label-correlation mechanism. The reader's ablation study shows performance drops when the decomposition modules are removed, but it does not isolate whether Eq. (4), the likelihood loss L_lm, or the final classification loss is actually responsible. Since the Github link in the paper is a placeholder, the intended denominator cannot be verified from released code, and the published equation must be taken at face value.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CorGCN, a Graph Convolutional Network for multi-label node classification. The method has two main modules: a Correlation-Aware Graph Decomposition that learns label prototypes, projects node features into per-label views via cosine-similarity coefficients, and sparsifies the original adjacency into per-label graphs using Top-λ selection; and a Correlation-Enhanced Graph Convolution that performs intra-label message passing on each label view followed by inter-label correlation propagation via attention over label prototypes. The model is trained with a sum of binary cross-entropy, a contrastive mutual-information loss (Eq. 4), and a focal-style likelihood loss (Eq. 6). Experiments on Humloc, PCG, Blogcatalog, PPI, and Delve report seven multi-label metrics against nine baselines, plus ablations, a backbone-generalization study, and parameter studies. The paper claims significant improvements over state-of-the-art methods on all five datasets.","tokens_in":21968,"tokens_out":3441,"duration_ms":36849,"significance":"If the central claims hold, the paper addresses a real gap: existing multi-label node classifiers use a unified message-passing process over ambiguous features and topology, and CorGCN's explicit per-label graph decomposition plus inter-label correlation propagation is a plausible remedy. The paper includes useful empirical studies: five datasets, seven metrics, ablations of the decomposition and convolution modules, generalization across five GNN backbones, and a complexity analysis. The method also attains the best average ranking over all baselines in Table 2. However, the headline claim of universal superiority is contradicted by the Delve results, and the key training signal for the label prototypes, Eq. (4), is degenerate as written. These issues must be resolved before the empirical and mechanistic claims can be accepted.","major_comments":[{"comment":"The statement that \"CorGCN can achieve significant improvements over state-of-the-art methods on all experimental datasets\" is contradicted by the Delve columns: LARN outperforms CorGCN on all seven metrics, e.g., LRAP 92.09 vs 89.20, Macro-AP 80.16 vs 72.80, and Hamming Loss 1.67 vs 2.48. The subsequent sentence listing Macro-AUC gains of 6.43%, 2.92%, 2.23%, and 4.45% correctly omits Delve, which shows the 'all datasets' claim is not supported by the table. The central claim should be revised to clearly state the four datasets on which CorGCN is the best, and the Delve result should be discussed rather than passed over.","section":"§5.2, Table 2"},{"comment":"The contrastive mutual-information loss as written only involves positive labels. For a node with exactly one positive label p, the term is log(exp(s_{i,p}) / exp(s_{i,p})) = 0; for a node with multiple positives, the denominator sums only over t ∈ Pos(y_i), so the loss merely pushes the similarities to co-occurring labels toward equality. Negative labels never appear in the denominator, so this is not a contrastive estimator and cannot provide the discriminative signal claimed for learning label prototypes. The authors should either correct the denominator to include negatives (e.g., all labels or sampled negatives) or explicitly identify L_lm and L_cls as the sources of prototype discrimination. Since the Github link in §5.1.3 is a placeholder, the intended formulation cannot be verified from the paper.","section":"§4.1.1, Eq. (4)"},{"comment":"The ablation study removes whole modules (w/o CFD, w/o CSD, w/o Intra, w/o Inter), but it does not isolate the contribution of the three loss terms in Eq. (20). Given the issue with Eq. (4), the reported performance drops from the decomposition modules could be attributed to extra model capacity or the hand-tuned Top-λ construction rather than the claimed contrastive label-correlation mechanism. Please add an ablation that removes L_cmi (or replaces it with a proper contrastive objective) while keeping the rest of the architecture unchanged, so the role of Eq. (4) can be assessed.","section":"§5.3, Figure 3"},{"comment":"The text states that \"the suitable value of λ for Humloc is larger than the value for PCG,\" but Table 5 reports Humloc λ=7 and PCG λ=19, which is the opposite. This inconsistency makes the parameter study and the exact experimental settings difficult to trust. Please correct the text or the table.","section":"§5.7.1 and Table 5"}],"minor_comments":[{"comment":"There are repeated typos where \"ambiguous\" is written as \"ambitious,\" for example in §5.2 (\"unified message passing with ambitious information\") and §5.4 (\"message passing paradigms in these backbones all remain ambitious\").","section":"Throughout"},{"comment":"The notation \"t ∈ y_i\" is undefined for a multi-hot vector; it should be \"t ∈ Pos(y_i)\" or the set of positive label indices should be defined explicitly for the summation.","section":"§4.1.1, Eq. (4)"},{"comment":"The Bonferroni-Dunn test is mentioned but the corresponding critical-difference diagram or post-hoc p-values are not shown; please provide the actual post-hoc results or state the comparison procedure used.","section":"§5.2"},{"comment":"The sentence \"the end-to-end label correlation modeling strategy employed by CorGCN yields an efficiency that is second to that of the original GCN\" is awkward and should be rephrased for clarity.","section":"§5.6, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a camera-ready or post-acceptance version with a KDD '25 DOI and ACM copyright notice. If this is under review for a journal as a novel submission, the reproducibility and claim-consistency issues are still the main concerns. The Delve contradiction in Table 2 is objective and fixable, and the degenerate Eq. (4) can be addressed with a corrected formulation or a clarified training-signal attribution. I do not see grounds for outright rejection, but the current version overstates its empirical contribution and leaves its core mechanism insufficiently specified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: CorGCN brings a genuinely new mechanism to multi-label node classification—building a separate graph view for each label and letting correlated views exchange messages during propagation. On four of the five datasets the numbers are credible and often large (e.g., +4.45 Macro-AUC on PPI over the best baseline). But the paper's central claim that it wins on all five datasets is false: on Delve, LARN beats CorGCN on all seven metrics, and by a wide margin on some (Ma-AP 80.16 vs 72.80). The opening of Section 5.2 needs to be rewritten.\n\nThe second soft spot is the contrastive loss in Eq. (4). As written, the denominator is the sum over t in y_i, i.e., only the node's positive labels. So for a single-positive node the term is identically zero, and for multi-positive nodes it only pushes the similarities to positives toward equality. It never compares against negative labels, so the label prototypes are not trained to distinguish associated from unassociated labels. The surrounding text claims positives and negatives are used, but the equation doesn't match that. This matters because the prototype quality is what makes the per-label graphs label-aware rather than just feature-similarity-aware. The overall model still gets a classification loss, so it may not be fatal, but the stated mechanism is different from what's implemented.\n\nThere is also the reproducibility issue: the code link is a placeholder, so the intended denominator cannot be checked. And the ablations (Figure 3) show that removing each component hurts, but they do not isolate whether Eq. (4) or the likelihood loss is doing the work. Those are minor compared to the Delve overclaim and the Eq. (4) problem.\n\nWhat is good: the per-label graph decomposition is not in prior MLNC work (ML-GCN, LANC, LARN, SMLG all use unified message passing). The correlation-enhanced convolution with inter-label attention is a clean extension. The experimental section is thorough: nine baselines, seven metrics, five datasets, plus ablations, backbone generalization, efficiency, and parameter sensitivity. The average rank (1.37) supports the overall picture. The large-label-space extension via macro prototypes is a sensible addition.\n\nWho should read this: researchers working on graph neural networks and multi-label learning. It is a solid empirical contribution that deserves to be in the literature, once the claims are matched to the actual numbers and the contrastive loss is either fixed or properly described. I would send it to review; the flaws are correctable.","headline":"New per-label graph decomposition that mostly works, but the paper overclaims on Delve and Eq. (4) is not a contrastive loss as written.","tokens_in":22528,"tokens_out":3623,"would_cite":true,"duration_ms":30225,"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":"Per-label graph decomposition with correlation-aware convolution lifts multi-label node classification accuracy on all five tested datasets, with Macro-AUC gains up to 6.43%.","keywords":["multi-label node classification","graph convolutional networks","graph structure learning","label correlation","graph decomposition","message passing","contrastive learning","semi-supervised learning"],"falsifier":"Replace the cosine-similarity edge selection in Eq. (10) with random selection of the same number λ of neighbors per node, retrain on Humloc, and compare Macro-AUC: if the score stays near the reported 77.31, the label-aware graph structure is not the source of the improvement; if it drops toward the 72.64 of the best baseline, the decomposition mechanism is confirmed as load-bearing.","tokens_in":21427,"feed_emoji":"🕸️","tokens_out":6560,"duration_ms":55835,"temperature":0.7,"pith_summary":"Multi-label node classification fails when a graph convolution treats every edge and feature as if it carried one clear label: real nodes belong to many categories, so neighborhoods and features are ambiguous, and messages get mixed. The paper argues that the fix is to decompose the graph per label — build a separate label-aware graph for each class — while keeping related labels connected, then let messages flow within each label view and between views. It proposes CorGCN, which does exactly this, and reports that it beats nine baselines on five datasets, with Macro-AUC gains of 6.43%, 2.92%, 2.23%, and 4.45% over the best baseline on Humloc, PCG, Blogcatalog, and PPI. A sympathetic reader would care because the approach replaces the one-graph-fits-all assumption that underlies standard GCNs in multi-label settings.","feed_headline":"Per-label graphs lift multi-label node classification by up to 6.4%","feed_subtitle":"Learning one graph per label and passing label correlations between views beats nine baselines on five datasets.","key_machinery":"The load-bearing object is the Correlation-Aware Graph Decomposition: for each label k, node features are projected onto the k-th label prototype, neighborhood-aggregated, and an edge is kept only if the cosine similarity of the aggregated projected features is among the node's Top-λ scores, yielding K label-aware graphs plus the original multi-label graph. This decomposition converts the ambiguous shared topology into label-specific message-passing paths. Carrying the second half is the Correlation-Enhanced Graph Convolution, whose inter-label propagation computes a per-node K×K correlation matrix from label-prototype and message similarities and uses it to mix messages across label views. The named components are the Correlation-Aware Graph Decomposition (CDG) and the Correlation-Enhanced Graph Convolution.","core_discovery":"The paper's central claim is that multi-label node classification is held back by two ambiguities — features and topology — that standard GCN message passing compounds, and that both can be reduced by learning one graph per label. CorGCN first projects node features toward label prototypes learned with a contrastive mutual-information objective and a focal likelihood decoder, then, for each label k, keeps only the edges whose cosine similarity in projected space is in the Top-λ for each node, producing label-aware graphs. Convolution runs twice: intra-label message passing on each label view, then inter-label correlation propagation that reweights each node's label messages by a softmax over prototype correlations. The paper reports that this beats the best baseline by 2.23% to 6.43% Macro-AUC on four datasets, and also improves over the backbone across GCN, GAT, SAGE, GIN, and SGC message-passing schemes.","pith_inferences":["Editorial inference: the per-label decomposition is a generic front-end — any message-passing backbone could consume the K label-aware graphs, so the same idea may transfer to link prediction or anomaly detection on multi-label graphs without retraining the decomposition.","Editorial inference: because the quality of the label-aware graphs hinges on the Top-λ ranking, a testable extension is to replace the hard Top-λ mask with a soft, learned neighbor selector; the paper's own w/o CSD ablation suggests the mask is doing real work, so a soft version might preserve gains with fewer hyperparameters.","Editorial inference: the macro-prototype extension implies the method's complexity scales with the number of meta-labels rather than raw labels, which could open very-large-label-space applications beyond the 121-label PPI setting tested.","Editorial inference: if the mechanism is truly about reducing ambiguity, the gains should be largest on graphs with low multi-label homophily (edge-connected nodes sharing few labels); that is a directly testable prediction the paper does not run."],"forward_implications":["On Humloc, PCG, Blogcatalog, and PPI, CorGCN raises Macro-AUC by 6.43%, 2.92%, 2.23%, and 4.45% over the best baseline respectively.","The decomposition is not tied to GCN: equipping GAT, SAGE, GIN, and SGC with CorGCN improves Micro-AUC by averages of 3.50% and 3.15% on Humloc and PCG.","Ablations show that removing the decomposed graph, the feature decomposition, the structure decomposition, the intra-label passing, or the inter-label propagation each hurts performance, so both the decomposition and the correlation propagation contribute to the gain.","On large-label-space datasets (PPI with 121 labels and Delve), clustering prototypes into macro labels (20 and 10 respectively) keeps performance close to the full label count while cutting per-label computation.","CorGCN trains faster than the top baselines on Humloc (0.76s versus 15.79s per epoch for ML-GCN) but is slower at inference (0.57s versus 0.49s), a trade-off the paper reports alongside the accuracy gains."],"supporting_citations":[{"why":"Supplies the Humloc and PCG benchmark datasets and the semi-supervised multi-label node classification setting the paper builds on.","marker":"[47]"},{"why":"Contributes the Blogcatalog dataset and the LANC baseline that the paper's per-label message passing must beat.","marker":"[49]"},{"why":"Contributes the Delve dataset and the LARN baseline; Delve is used to test the large-label-space macro-prototype extension.","marker":"[38]"},{"why":"Provides the GCN backbone whose unified message passing is the paper's starting point and baseline.","marker":"[18]"},{"why":"Supplies the contrastive mutual-information estimator with Gaussian mixture variational autoencoder that the paper adapts for node-label correlation modeling.","marker":"[2]"},{"why":"Provides the contrastive graph structure learning baseline SUBLIME that the paper compares against.","marker":"[23]"},{"why":"Supplies the PPI dataset with 121 labels used to test the macro label prototype extension.","marker":"[44]"},{"why":"Provides the ML-GCN baseline that models label correlations in embedding space and is the strongest comparison on several datasets.","marker":"[11]"}],"fun_headline_variants":["CorGCN: per-label graphs up multi-label node classification by 6.4%","One graph per label lifts multi-label node classification by 6.4%","Correlation-aware GCN improves multi-label node classification by 6.4%","Per-label graphs cut ambiguity in multi-label node classification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The per-label graphs are chosen by keeping, for each node, only the Top-λ neighbors by cosine similarity of aggregated projected features, and if that ranking does not reflect true label-relevant adjacency, the decomposed graphs propagate the wrong messages and the reported gains disappear, with λ itself hand-tuned per dataset.","fun_headline_variants_meta":{"raw":{"variants":["CorGCN: per-label graphs up multi-label node classification by 6.4%","One graph per label lifts multi-label node classification by 6.4%","Correlation-aware GCN improves multi-label node classification by 6.4%","Per-label graphs cut ambiguity in multi-label node classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000703,"raw_usage":{"total_tokens":3175,"prompt_tokens":950,"completion_tokens":2225,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":2143}},"tokens_in":566,"tokens_out":2225,"duration_ms":13847,"temperature":1.0,"reasoning_tokens":2143,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:12:53.693405+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the cosine-similarity edge selection in Eq. (10) with random selection of the same number λ of neighbors per node, retrain on Humloc, and compare Macro-AUC: if the score stays near the reported 77.31, the label-aware graph structure is not the source of the improvement; if it drops toward the 72.64 of the best baseline, the decomposition mechanism is confirmed as load-bearing.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Humloc and PCG benchmark datasets and the semi-supervised multi-label node classification setting the paper builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the Blogcatalog dataset and the LANC baseline that the paper's per-label message passing must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the Delve dataset and the LARN baseline; Delve is used to test the large-label-space macro-prototype extension."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the contrastive mutual-information estimator with Gaussian mixture variational autoencoder that the paper adapts for node-label correlation modeling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the PPI dataset with 121 labels used to test the macro label prototype extension."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ML-GCN baseline that models label correlations in embedding space and is the strongest comparison on several datasets."}],"review_version":1}