{"id":"65e5d449-b615-414d-9f19-8188306ff2b6","arxiv_id":"2412.12984","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"C3GNN improves class-imbalanced graph classification by clustering majority classes into balanced subclasses, applying Mixup inside each subclass, and learning with hierarchical supervised contrastive losses.","lead":"This paper introduces C3GNN, a training method for graph classification when some classes have far more examples than others. It splits large classes into smaller clusters, creates synthetic examples with Mixup, and uses contrastive learning to improve accuracy on imbalanced graph benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline empirical claim is not yet supported: no error bars or code, and gains may come from per-dataset tuning and a circular cluster-validity check.","rationale":"The reader's weakest_assumption targets cluster stability and semantic coherence. That is a real concern, but I see the empirical support as the primary load-bearing element: the paper provides no code, no error bars, and a per-dataset tuned δ, so the central \"outperforms all baselines\" claim cannot be independently verified from the manuscript. The random-label ablation would settle whether the clustering mechanism itself matters; the significance test would settle whether the accuracy differences are real. I keep the verdict CONDITIONAL rather than REJECT because the reported consistent gains across six datasets and the ablations are plausible, and the concerns are addressable with released code and additional analyses. If the requested tests fail, the verdict should move toward REJECT or UNVERDICTED; if they pass, the verdict could move to ACCEPT.","tokens_in":13769,"tokens_out":5411,"duration_ms":56420,"concrete_test":"Run an ablation in which subclass labels for each majority class are randomly permuted while preserving cluster sizes and all other hyperparameters, across 10 seeds. If accuracy with random labels is within run-to-run noise of the reported C3GNN accuracy, the semantic-clustering mechanism is not load-bearing and the superiority claim reduces to a tuned contrastive regularizer. Also release per-seed accuracies and a paired significance test (e.g., Wilcoxon signed-rank) for C3GNN versus the best baseline on each dataset; if those differences are not significant, the headline claim should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that C3GNN beats every baseline on all six datasets. For that claim to hold, the numbers in Table 1 must be reliable and the mechanism credited for the gains must actually be doing the work. Neither condition is currently established. Table 1 reports only \"average top-1 accuracy over 10 run times\"; no standard deviations, per-seed results, or significance tests are given. Several margins are small (e.g., 0.68 pp on Letter-high IF=25 and 1.25 pp on Synthie IF=15 vs. RAHNet), so without uncertainty quantification the superiority claim is not testable. The implementation section says the cluster-size control parameter δ is \"fine-tuned ... for each dataset individually,\" and Figure 2 shows accuracy swings of up to roughly 5 pp across cluster sizes, so the reported numbers may partly reflect per-dataset hyperparameter selection. On the mechanism, Figure 3 shows that intra-subclass distance is smaller than intra-class distance, but this is exactly the objective of Eqs. (6)-(8); it is a circular validation. Figure 4 is qualitative and cherry-picked. Because cluster labels are recomputed every T=10 epochs from a moving encoder, the encoder and clusters co-adapt; without a stability or semantic-purity measure, one cannot tell whether the gains come from discovering semantic substructure or from arbitrary partitions plus extra contrastive regularization.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes C3GNN, a method for class-imbalanced graph classification that combines adaptive k-means clustering of majority classes into balanced subclasses, subclass-level Mixup interpolation, and hierarchical supervised contrastive losses (intra-subclass and inter-subclass). The central claim is that C3GNN outperforms all considered baselines on all six benchmark datasets across various imbalance factors (IFs). The method is evaluated on synthetic and real-world graph classification datasets with GraphSAGE as the encoder, and ablations, hyperparameter sensitivity, feature-distance analyses, and a qualitative visualization of clusters are provided.","tokens_in":14055,"tokens_out":2839,"duration_ms":27088,"significance":"If the empirical claims are substantiated, the paper would make a useful contribution to the underexplored area of class-imbalanced graph classification. The central idea of decomposing majority classes into semantically coherent subclasses to balance class sizes and then applying hierarchical contrastive learning is simple and well motivated, and the paper combines several known components (clustering, Mixup, SupCon) in a novel configuration for the graph-classification setting. The experimental coverage includes six datasets and a range of baseline families (re-sampling, re-weighting, contrastive, and augmentation-based methods), which is a strength. However, the headline claim of universal superiority currently rests on point estimates without variance or significance testing, and the mechanism attribution is supported partly by a circular validation (the feature distances in Figure 3 are the direct objective of the losses in Eqs. 6-8). These gaps are load-bearing for the central claim and must be addressed before the results can be accepted at face value.","major_comments":[{"comment":"Table 1 reports only average top-1 accuracy over 10 runs, with no standard deviations, per-seed results, or significance tests. Several margins over the best baseline are small (e.g., 0.68 pp on Letter-high IF=25 and 1.25 pp on Synthie IF=15 relative to RAHNet). Without uncertainty quantification, the claim that C3GNN \"outperforms all baselines on all six datasets\" is not statistically supported. Please report mean +/- std over the 10 runs and conduct paired tests (e.g., paired t-test or Wilcoxon signed-rank test) against the strongest baseline for each configuration.","section":"Table 1 / Implementation details"},{"comment":"The cluster size control parameter δ (Eq. 4) is \"fine-tuned ... for each dataset individually,\" and Figure 2 shows accuracy variations of up to roughly 5 percentage points across cluster sizes (e.g., Synthie IF=15). If δ is selected using the test set or the same validation set used for early stopping, the reported gains may partly reflect per-dataset hyperparameter selection rather than the method's intrinsic advantage. Please clarify whether δ was chosen on a separate validation split, report the exact δ values used for each dataset and IF, and consider reporting results averaged over a range of δ or using a principled criterion (e.g., validation accuracy) to pick δ.","section":"Implementation details / Hyper-parameter Sensitivity (Figure 2)"},{"comment":"The construction of the Zipf-imbalanced training sets is not fully specified. The text says the training sets were \"transformed into class-imbalanced datasets with varying imbalance factors\" and Eq. (1) gives a general integral condition, but the exact sampling procedure is missing: how many classes are included, how the class frequencies are assigned under Zipf's law, whether the 6:2:2 split is performed before or after the imbalance transformation, and whether the test/validation sets are truly balanced with the same original class labels. Without this specification the reported IF values and the comparability of results across methods cannot be fully assessed.","section":"Datasets / Eq. (1)"},{"comment":"The validation of the clustering mechanism is circular. Figure 3 shows that intra-subclass distances are smaller than intra-class distances and that subclass distances are more uniform across many/medium/few regions; however, these properties are exactly what the losses in Eqs. (6)-(8) are designed to enforce, so the figure does not provide independent evidence that the clusters correspond to semantically meaningful substructures. Please add a non-circular evaluation, for example: (i) cluster stability across re-initializations or across repeated runs with different random seeds, (ii) comparison against random partitions of the same class into the same number of subclasses, or (iii) quantitative alignment with known semantic groupings if any exist in the datasets.","section":"Analysis of Feature Distribution / Eqs. (6)-(8)"},{"comment":"Figure 4 is a qualitative, hand-picked illustration, and the clustering is recomputed every T=10 epochs from a moving encoder, so the encoder and cluster assignments co-adapt. On its own this does not establish that the clusters capture stable semantic substructure; the gains could in principle come from arbitrary partitions combined with additional contrastive regularization. Please report a quantitative measure of cluster stability over training (e.g., adjusted Rand index between consecutive clusterings or agreement across different initializations) and, where possible, compare against a variant that uses fixed clusters computed once before training.","section":"Visualization of Clustering Results / Algorithm 1"}],"minor_comments":[{"comment":"The sentence \"which aims to classify the class labels of graphs and has emerged as one of the most interesting also popular topics\" contains a typo: \"also popular\" should be \"and popular\".","section":"Introduction"},{"comment":"The notation P(i)/Q(i) and A(i)/Q(i) is ambiguous; it should be written as P(i) \\setminus Q(i) and A(i) \\setminus Q(i) to denote set difference.","section":"Equation (7)"},{"comment":"The x-axis is labeled \"Cluster size,\" but Eq. (4) defines M as the threshold that indirectly controls cluster size via δ. Please clarify in the caption and text whether the x-axis is δ or the resulting number of subclasses per majority class, and explain how the displayed range corresponds to the tuning procedure.","section":"Figure 2"},{"comment":"The line \"Sample one augmentation from You et al. (2020)\" is underspecified; please state which of the four augmentations (node dropping, edge perturbation, attribute masking, subgraph) are used, and whether the augmentation is sampled per graph or per batch.","section":"Algorithm 1"},{"comment":"The abstract and introduction mention NCI as a motivating example, but NCI is not among the six datasets in Table 1. Adding it or removing the motivational reference would improve consistency.","section":"Experiments"},{"comment":"The complexity statement says the clustering complexity is O(I K N D) and \"simplified to O(N D)\" because I and K are constants; since K is the number of clusters, please clarify whether K is treated as bounded by the number of classes times a small constant, and state any assumptions about the number of majority classes.","section":"Computational Complexity Analysis"}],"recommendation":"major_revision","confidential_remarks":"This is an empirical paper, and the recommendation of major revision is driven by the need to strengthen the evidence for the central claim: the absence of variance/significance reporting, the per-dataset tuning of δ, and the partly circular validation of the clustering mechanism. These are fixable within the scope of the paper (e.g., by reporting std/significance tests, clarifying the hyperparameter selection protocol, and adding cluster-stability or random-partition baselines). I would not reject the paper outright; the idea is plausible and the experimental setup is reasonably broad. In revision, the authors should also consider releasing code to support reproducibility, as the implementation details are otherwise incomplete."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a reasonable engineering combination for class-imbalanced graph classification—clustering majority classes into balanced subclasses, mixup within subclasses, and hierarchical supervised contrastive learning—and it reports consistent gains across six benchmarks. But the empirical evidence has a real hole: no error bars, and the key hyperparameter is tuned per dataset. The central idea is a sensible adaptation of known techniques to graph-level tasks, not a conceptual breakthrough, but the combination is new and the paper is clearly written.\n\nWhat it does well: the method is simple and the ablation study shows each component contributes. The sensitivity analysis for cluster size is honest about the parameter's effect, and the authors try to characterize the learned representations with feature distances and cluster visualizations—more than many papers in this space do. On the numbers, C3GNN beats the baselines on all twelve settings (six datasets × two imbalance factors), and some margins are substantial (e.g., COIL-DEL IF=10: 47.54 vs. 43.04).\n\nThe soft spots are real. No standard deviations are reported despite \"10 run times,\" and some winning margins are under a point (Letter-high IF=25: 0.68 pp). The cluster-size control parameter δ is fine-tuned per dataset, and Figure 2 shows accuracy swings of several points as cluster size changes—so part of the reported advantage may reflect favorable tuning. The stress-test note's circularity point lands: showing that intra-subclass distance is smaller than intra-class distance (Figure 3) just verifies that the contrastive loss is doing what it was designed to do; it doesn't establish that the clusters capture genuine semantics. The qualitative visualization (Figure 4) is suggestive but cherry-picked. No code is provided, which makes reproduction hard.\n\nThe central mechanism—k-means on GNN embeddings finding meaningful substructure—is plausible, but the paper doesn't quantify cluster stability or purity. None of this is fatal. The method is coherent, the experiments are extensive, and the issues are fixable: add error bars, run significance tests, release code, and include a cluster-stability metric. I'd send this to peer review with a request for major revision. It's a subfield advance, of interest to people working on imbalanced graph classification, and it deserves a serious referee—but the referee should demand the missing uncertainty quantification.","headline":"A plausible, clearly-written method for class-imbalanced graph classification that deserves a careful review, but the empirical case is undercut by missing error bars and per-dataset tuning.","tokens_in":14572,"tokens_out":3950,"would_cite":true,"duration_ms":34512,"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":"The paper claims that C3GNN—adaptive clustering of majority classes into balanced subclasses, subclass Mixup, and hierarchical supervised contrastive learning—outperforms re-sampling, re-weighting, and long-tailed graph baselines on all…","keywords":["class-imbalanced graph classification","graph neural networks","supervised contrastive learning","adaptive clustering","Mixup","long-tailed learning","subclass balancing","graph-level representation learning"],"falsifier":"Run C3GNN with the same cluster sizes but with each majority class randomly partitioned into subclasses, and compare accuracy on all six datasets: if random partitions achieve nearly the same results as the reported k-means assignments, the claim that discovered semantic substructure drives the gains is refuted. A complementary check is to measure cluster stability across the $T$-epoch updates with the adjusted Rand index; low stability without a corresponding accuracy drop would also weaken the mechanism.","tokens_in":13588,"feed_emoji":"📊","tokens_out":6344,"duration_ms":50190,"temperature":0.7,"pith_summary":"Class-imbalanced graph classification is hard because GNNs trained on skewed class distributions tilt toward the largest classes. The paper proposes C3GNN, which clusters each majority class into multiple subclasses whose sizes mimic the minority class, so that no single class dominates the learning budget. Inside each subclass, Mixup creates interpolated training graphs, and two supervised contrastive losses organize the representation space hierarchically: one pulls graphs in the same subclass together, the other pulls same-class but different-subclass graphs together. The paper reports that this combination outperforms all compared re-sampling, re-weighting, contrastive, and long-tailed graph baselines on all six benchmark datasets.","feed_headline":"Adaptive clustering of majority classes beats imbalance baselines","feed_subtitle":"Splitting big classes into balanced subclasses and contrasting them tops six graph benchmarks under class imbalance.","key_machinery":"The load-bearing mechanism is the adaptive subclass assignment produced by k-means on current GNN graph embeddings, with subclass size capped by $M = \\max(n_K, \\delta)$. This converts a skewed class distribution into a balanced set of pseudo-classes, which is what lets the two supervised contrastive losses (Eqs. 6 and 7) learn without letting majority classes dominate. Subclass Mixup (Eq. 5) guards against representation collapse in sparse subclasses, and cluster centers are refreshed every $T$ epochs so that the partition co-evolves with the encoder.","core_discovery":"On the paper's own terms, the discovery is that majority classes in imbalanced graph data should be treated as collections of semantic subclasses rather than as monolithic classes. C3GNN encodes graphs with a GNN, applies k-means to the embeddings of each majority class to form subclasses capped at size $M = \\max(n_K, \\delta)$, synthesizes additional samples inside each subclass by Mixup interpolation, and optimizes the joint loss $\\mathcal{L} = \\sum_i (\\mathcal{L}^{\\mathrm{intra}}_i + \\beta \\mathcal{L}^{\\mathrm{inter}}_i)$. The intra-subclass loss (Eq. 6) makes augmented views of graphs with the same subclass label attract each other, while the inter-subclass loss (Eq. 7) attracts graphs from different subclasses of the same class and repels graphs from other subclasses. In the reported experiments, this hierarchical structure yields the best top-1 accuracy on Synthie, ENZYMES, MNIST, Letter-high, Letter-low, and COIL-DEL across all tested imbalance factors.","pith_inferences":["A testable extension outside the paper is to replace k-means with a supervised or structural sublabel discovery method and check whether accuracy tracks cluster semantic purity; if it does not, the reported gains may come mainly from rebalancing rather than from recovered substructure.","The approach is stated for graph classification, but the same cluster-then-hierarchical-contrast recipe could transfer to other long-tailed structured inputs with latent subcategories, such as point clouds or code snippets, since nothing in Eqs. (5)–(7) is graph-specific beyond the encoder.","Because C3GNN's contrastive objectives shape representations but do not alter the final classifier, stacking it with logit adjustment or class-balanced losses at the classification head is a natural combination the paper does not explore."],"forward_implications":["On each of the six benchmark datasets, C3GNN reports the highest average top-1 accuracy among the compared methods, with the gap over baselines generally widening as the imbalance factor increases.","The ablation study reports that removing hierarchical subclass contrastive learning, adaptive clustering, or subclass Mixup each lowers accuracy on all six datasets, so the three modules are presented as jointly necessary for the result.","Feature-distance analysis on MNIST with imbalance factor 50 shows smaller intra-subclass distances than intra-class distances and more uniform subclass volumes across many, medium, and few-shot regions, indicating a subclass-balanced representation space.","The sensitivity analysis shows an optimal cluster size that grows when the minority class is larger, implying the method's preferred granularity is tied to minority-class size."],"supporting_citations":[{"why":"Supplies the supervised contrastive learning formulation that Eqs. (6) and (7) adapt to subclass and class labels.","marker":"Khosla et al. 2020"},{"why":"Supplies the graph augmentation distributions used to build positive pairs and the GraphCL baseline.","marker":"You et al. 2020"},{"why":"Supplies the Mixup interpolation mechanism used in Eq. (5) to synthesize samples inside each subclass.","marker":"Zhang et al. 2017"},{"why":"Supplies the adaptive clustering and subclass-balancing idea for long-tailed recognition that C3GNN extends to graphs.","marker":"Hou et al. 2023"},{"why":"Provides the RAHNet baseline, a retrieval-augmented long-tailed graph classification method that C3GNN must beat.","marker":"Mao et al. 2023"},{"why":"Provides the G2GNN baseline, graph-of-graph neural networks for imbalanced graph classification.","marker":"Wang et al. 2022"},{"why":"Provides the class-balanced loss baseline representing re-weighting strategies.","marker":"Cui et al. 2019"},{"why":"Provides the logit-adjusted cross-entropy (LACE) baseline for long-tail learning.","marker":"Menon et al. 2020"}],"fun_headline_variants":["Clustering majority classes improves imbalanced graph classification","Split big classes into subclasses for better graph imbalance","C3GNN: cluster-then-contrast beats imbalance on graphs","Majority classes as subclasses: key to imbalanced graphs","Subclass clustering lifts imbalanced graph classification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that k-means on the current GNN embeddings of each majority class separates that class into semantically coherent subclasses, so that balancing those subclasses against the minority class is meaningful rather than arbitrary; if the clusters are unstable or do not match real substructure, the two contrastive losses would reinforce accidental partitions.","fun_headline_variants_meta":{"raw":{"variants":["Clustering majority classes improves imbalanced graph classification","Split big classes into subclasses for better graph imbalance","C3GNN: cluster-then-contrast beats imbalance on graphs","Majority classes as subclasses: key to imbalanced graphs","Subclass clustering lifts imbalanced graph classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000167,"raw_usage":{"total_tokens":1274,"prompt_tokens":976,"completion_tokens":298,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":218}},"tokens_in":592,"tokens_out":298,"duration_ms":3414,"temperature":1.0,"reasoning_tokens":218,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:30:35.171226+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run C3GNN with the same cluster sizes but with each majority class randomly partitioned into subclasses, and compare accuracy on all six datasets: if random partitions achieve nearly the same results as the reported k-means assignments, the claim that discovered semantic substructure drives the gains is refuted. A complementary check is to measure cluster stability across the $T$-epoch updates with the adjusted Rand index; low stability without a corresponding accuracy drop would also weaken the mechanism.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the graph augmentation distributions used to build positive pairs and the GraphCL baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the RAHNet baseline, a retrieval-augmented long-tailed graph classification method that C3GNN must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the G2GNN baseline, graph-of-graph neural networks for imbalanced graph classification."}],"review_version":1}