{"id":"fc3edd89-56ff-4675-8a41-3e238ff47948","arxiv_id":"2505.10307","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"NML-GCL learns a negative metric network that down-weights false negatives during contrastive training, improving node classification and clustering on six standard graph benchmarks.","lead":"This paper proposes a graph contrastive learning method that learns to re-weight negative samples so that false negatives are pulled closer to the anchor node. The authors report consistent gains on six benchmark graphs and offer a theoretical analysis they claim proves a tighter mutual-information bound.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's proof assumes augmentations are generated from labels T plus independent noise, giving I(G_U;G_V)=I(T;T); the actual augmentations are generated from the original graph G, so this equality is false and the claimed collapse of false negatives does not follow.","rationale":"I read the paper as an engineering contribution with a claimed theoretical foundation. The experimental section is internally consistent: six benchmarks, ablations, hyperparameter analysis, and a consistent pattern of improvement over baselines. If the theoretical claims were removed, the method would stand as a plausible empirical contribution. However, the abstract explicitly advertises 'solid theoretical analysis,' and that analysis is not sound. The reader's identification of Appendix B.2's Markov-chain assumption is correct and is the most load-bearing flaw: it is not a minor regularity condition but contradicts the algorithmic description in Section 3.1, where both views are generated by perturbing G. On real graphs, G contains information beyond labels, so I(G_U;G_V)>I(T;T) in general, and the proof's conclusion that optimal embeddings collapse false negatives does not follow. I also note two secondary issues that reinforce the verdict: the variational bound in Theorem 1 (Appendix B.1) drops the term -E_Q[e^f]+1 and uses a critic that depends on the learned M, so the bound is at best informal; and the closed-form update in Equations (9)-(11) has exponent errors, with the numerator/denominator mixing θ(ui,vi) and θ(ui,vk), while the sigmoid in Equation (11) does not enforce the normalization in Equation (10). I do not rest the verdict on these secondary issues, but they support the conclusion that the theoretical analysis is not a reliable foundation for the central claim. The proposed concrete check, re-deriving Appendix B.2 with the actual augmentation chain, is decisive for the Theorem 2 concern: if the equality to I(T;T) fails, the collapse argument fails. Therefore the reader's REJECT verdict should remain unchanged.","tokens_in":19981,"tokens_out":13943,"duration_ms":142022,"concrete_test":"Re-derive the chain-rule expansion in Appendix B.2 using the actual augmentation model G_U=aug(G,e1) and G_V=aug(G,e2), i.e., replace the Markov chains by {G,e1}->G_U->U and {G,e2}->G_V->V. If the result is I(T;T) plus a nonzero term involving within-class structure or dependence of e1,e2 on G given T, then the equality I(G_U;G_V)=I(T;T) is false for real graphs and Theorem 2's collapse argument fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is that maximizing MI between views makes the optimal encoder rank false negatives closer than true negatives (Theorem 2). The proof in Appendix B.2 constructs Markov chains {T,e1}->G_U->U and {T,e2}->G_V->V and then derives I(G_U;G_V)=I(T;T). This requires that the original graph G carries no information about the views beyond the label variable T. But the method itself (Section 3.1, Algorithm 1) generates G_U and G_V by DropEdge and FeatureMasking applied to the observed graph G, not by sampling from T. Real graphs contain within-class attribute and structural variability that is not a deterministic function of Y, so I(G_U;G_V) generally exceeds I(T;T). Even granting the lossless-encoder step, equality of mutual informations I(U*;V*)=I(T;T) does not imply that every false-negative pair with Y(j)=Y(i) satisfies z_i=z_j; mutual information is an average statement and permits label-irrelevant variation in the embeddings. Consequently, Theorem 2's conclusion is not established, and the advertised guarantee that NML-GCL endows the encoder and negative metric network with false-negative discrimination is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NML-GCL, a graph contrastive learning method that augments InfoNCE with a learnable Negative Metric Network (NMN). The NMN assigns soft weights to negative samples, and the encoder and NMN are trained jointly through a bi-level optimization objective. The manuscript claims two theoretical results: Theorem 1 states that the NML loss is a tighter lower bound on view mutual information than InfoNCE, and Theorem 2 states that maximizing this bound makes the optimal encoder place false negatives closer to an anchor than true negatives. The authors report consistent improvements over several GCL baselines on six benchmark datasets, with ablations and hyper-parameter studies.","tokens_in":20235,"tokens_out":6769,"duration_ms":67653,"significance":"The empirical component is carefully presented: results are averaged over 10 runs, standard deviations and p-values are reported, ablation variants are compared, and hyper-parameter search spaces are documented. If the theoretical claims were sound, the work would offer a principled, prior-free treatment of false negatives in graph contrastive learning. However, the central theoretical results are not established: the proof of Theorem 1 contains an algebraic error in the variational lower-bound argument, and Theorem 2 relies on an augmentation model that does not match the actual algorithm. Because the advertised false-negative discrimination guarantee is the main conceptual contribution, the significance of the paper as submitted is substantially weakened.","major_comments":[{"comment":"The variational proof drops the penalty term. After substituting f(u_i,v_i)=log(N s_ii/(s_ii+(N-1) sum_j E_{P_U P_V}[m_ij s_ij])), the displayed chain contains the term '- E_{(u_i,v_i)~P_U P_V}[N s_ii/(s_ii+(N-1) sum_j E_{P_U P_V}[m_ij s_ij])] + 1'. In the next line this term is removed without justification; it is not identically 1. Without it, the inequality I(U;V) >= I_NML(U;V) does not follow from the variational formula. The theorem statement is also internally inconsistent with the proof: it defines I_NCE(U;V)=L_InfoNCE+C, while the proof uses I_NCE=-E[L_InfoNCE]+C. Since Theorem 1 is the stated basis for the bi-level objective, this is a load-bearing error.","section":"Appendix B.1, proof of Theorem 1"},{"comment":"The proof models the contrastive views as generated from the label variable T and independent perturbations e1,e2, leading to I(G_U;G_V)=I(T;T). This contradicts the actual method: Section 3.1 and Algorithm 1 generate G_U and G_V from the original graph G via DropEdge and FeatureMasking. In real graphs, attributes and topology contain information beyond labels, so I(G_U;G_V) generally exceeds I(T;T). Consequently, the claim that the optimal lossless encoder satisfies E[I(U*;V*)]=I(T;T) is not justified, and the conclusion z_i=z_j for all same-label pairs does not follow. Moreover, even if the mutual information equality held, it is an average statement and does not imply pointwise collapse of embeddings for every false-negative pair. Theorem 2's advertised guarantee that NML-GCL can distinguish false negatives from true negatives is therefore unsupported.","section":"Appendix B.2, proof of Theorem 2"},{"comment":"The stationary-condition derivation for the negative metric network is not correct. Differentiating L_NML^(i) with respect to m_ik gives (N-1)s_ik/(s_ii+(N-1) sum_j m_ij s_ij), not the expression in Eq. (9), which uses s_ii in every denominator term and omits the factor N-1. As a result, Eq. (11) is not a valid solution of the stated first-order condition. This matters because the inverse proportionality between m_ik and theta(u_i,v_k) is used to argue that the encoder and NMN reinforce each other in bi-level training.","section":"Section 3.3, Eqs. (9)-(11)"},{"comment":"The claim that I_NML(U;V) >= I_NCE(U;V) is essentially by construction: the uniform weighting m_ij=1/(N-1) with m_ii=0 is a feasible choice in Eq. (5), so the optimum over M is at least the InfoNCE objective. This only shows a comparison between two training objectives, not that the resulting value is a valid tighter lower bound on the true mutual information. Establishing the latter requires the missing penalty-term argument from Appendix B.1, which is not supplied.","section":"Section 4.1, Theorem 1"}],"minor_comments":[{"comment":"The phrase 'solid theoretical analysis' overstates the current state of the proofs, given the issues in Theorem 1 and Theorem 2.","section":"Abstract and Section 1"},{"comment":"Line 8 of Algorithm 1 refers to updating M according to Eq. (7), but Eq. (7) is the joint bi-level objective for both E and M; the intended update rule appears to be the inner minimization of Eq. (4)-(5). Please clarify the reference.","section":"Section 3.3, text near Eq. (7)"},{"comment":"The definition I_NCE(U;V)=L_InfoNCE+C conflicts with the proof's use of -E[L_InfoNCE]+C. This sign inconsistency should be corrected.","section":"Section 4.1, Theorem 1 statement"},{"comment":"The table reports hyper-parameter ranges but does not specify how alpha, T_M, and T_E were selected; for reproducibility, a short statement on the selection criterion would be helpful.","section":"Appendix E, Table 4"}],"recommendation":"reject","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the method is a plausible empirical contribution—learnable soft weighting of negative samples in graph contrastive learning, trained with alternating optimization. The experiments are decent: six benchmarks, consistent gains around 1–3 points, standard deviations, p-values, and ablations. The novelty is real but narrow: weighting negatives is established, but learning the weighting function with a small MLP is new.\n\nWhat does not hold up is the theory the paper leans on. Theorem 2's proof assumes augmented views are generated from the label variable plus independent noise, so the original graph structure and features contribute nothing beyond the labels. That is false for the actual augmentations, which are DropEdge and feature masking applied to the observed graph. Even if the mutual information equality were true, it is an average statement and does not imply every false-negative pair gets identical embeddings. So the claimed guarantee that maximizing MI collapses false negatives to the anchor is not established. Theorem 1 is more defensible but close to trivial: since the learned weights include uniform as a special case, the bound is automatically at least as tight as InfoNCE. Calling that a theoretical result is overclaiming. And Equations (9) and (11), the closed-form update for the negative metric network, have exponent errors and a misspecified denominator; as printed they do not follow from the loss.\n\nThe empirical section is the real strength. The method consistently beats the baselines, and the ablations (w/o M, replace with cosine similarity) show the learned metric matters. But there is no code release, which hurts reproducibility, and the hyperparameter search for baselines is not fully detailed. These are minor relative to the theory problem, but they matter.\n\nWho is this for? Someone working on false negatives in graph contrastive learning might get ideas from the loss design and the alternating training. It is not a paper to cite for theoretical guarantees. My recommendation: this deserves a serious referee, but not acceptance in the current form. Ask the authors to either fix the proofs with real assumptions or, preferably, resubmit as a purely empirical method paper with the theory section cut down to a simple inequality. The method is worth having; the packaging is not.","headline":"A learnable negative-weighting module gives consistent but modest gains in graph contrastive learning, yet the advertised theory (especially Theorem 2) does not hold up; treat it as an empirical paper with a broken theoretical veneer.","tokens_in":20760,"tokens_out":4676,"would_cite":false,"duration_ms":48344,"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 a learnable negative metric network tightens the mutual-information lower bound for graph contrastive learning and consistently outperforms prior methods on six benchmark graphs.","keywords":["graph contrastive learning","false negatives","negative metric learning","mutual information","InfoNCE","graph neural networks","bi-level optimization","self-supervised learning"],"falsifier":"Run NML-GCL on a heterophilous graph where connected nodes usually have different labels, using the same augmentation recipe, and measure the median distance ratio between false negatives and true negatives as in the paper's Table 3. If the ratio no longer falls below the InfoNCE baseline on such graphs, the label-only Markov-chain assumption behind Theorem 2 is the cause; the paper's current evidence covers only homophilous citation and co-purchase networks.","tokens_in":19707,"feed_emoji":"🕸️","tokens_out":6666,"duration_ms":62651,"temperature":0.7,"pith_summary":"Graph contrastive learning treats every non-anchor node as a negative sample, so same-label nodes are often wrongly pushed apart. This paper proposes replacing the fixed negative weights in the InfoNCE loss with a learnable Negative Metric Network that outputs a soft, normalized weight for every pair, indicating how likely one node is a true negative of another. Because these weights are learned from the contrastive views themselves through a bi-level objective, the method needs no hand-picked similarity thresholds, cluster counts, or neighborhood rules. The paper proves that optimizing this weighted loss maximizes a lower bound on mutual information that is at least as tight as InfoNCE, and that an optimal encoder keeps false negatives closer to the anchor than true negatives. On six graph benchmarks, the resulting embeddings improve downstream node classification and clustering over nine baselines.","feed_headline":"A learnable negative metric beats InfoNCE on graph contrastive learning","feed_subtitle":"False negatives no longer need hand-picked thresholds: the network learns which negatives to trust, improving node classification and…","key_machinery":"The load-bearing objects are the Negative Metric Network $M$ and the bi-level optimization that couples it with the encoder $E$. $M$ maps an embedding pair $(u_i, v_j)$ through an MLP and normalizes over $j$ so that each row of the negative metric matrix $M\\in[0,1]^{N\\times N}$ sums to one; cell $m_{ij}$ is the soft label for whether $j$ is a negative of $i$. The proof machinery is the variational mutual-information lower bound: representing the KL divergence as a maximization over test functions shows that minimizing $\\mathcal{L}_{\\mathrm{NML}}^{(i)}$ over $M$ maximizes a lower bound on $I(U;V)$ that dominates InfoNCE. Equation (11) is the feedback mechanism: the optimal metric weight $m_{ik}^{(t+1)}$ is inversely proportional to the encoder's current similarity $\\theta(u_i^{(t)}, v_k^{(t)})$, so the encoder's push-out of $v_k$ makes the metric mark $v_k$ as a stronger negative, which then instructs the next encoder update to push it even farther, forming a self-training loop. Theorem 2 rests on a Markov-chain argument in which the augmented views are generated by the label variable plus independent random perturbations.","core_discovery":"The central claim is that the false-negative problem in graph contrastive learning can be addressed by learning the geometry of negativity itself rather than assuming it. NML-GCL augments a GCN encoder with a Negative Metric Network, an MLP followed by a softmax over nodes, that assigns each candidate negative pair a weight $m_{ij}$, interpretable as the probability that node $j$ is a true negative of anchor $i$. Training alternates between updating the metric network with the encoder frozen and updating the encoder with the metric network frozen, in a bi-level objective; the self-supervised signal that the two views of the same node form a positive pair implicitly supervises the metric network, since the hinge-form loss forces large weights onto dissimilar candidates. The theoretical part proves $I(U;V) \\ge I_{\\mathrm{NML}}(U;V) \\ge I_{\\mathrm{NCE}}(U;V)$, so maximizing the proposed loss maximizes a tighter MI lower bound than InfoNCE, and proves that an MI-optimal encoder satisfies $\\mathbb{E}_{j\\in S_i^+}[d(z_i,z_j)] \\le \\mathbb{E}_{j\\in S_i^-}[d(z_i,z_j)]$, meaning it ranks negatives the way the oracle label function would. The experiments report consistent accuracy and clustering gains over baselines that use fixed, prior-driven negative weights.","pith_inferences":["If the label-generation model behind Theorem 2 fails on graphs where structure carries information beyond class labels, NML-GCL's advantage on heterophilous graphs may shrink; testing distance ratios on heterophilous benchmarks would reveal how much of the gain is label collapse versus structure preservation.","The learned negative metric matrix could be exported as a soft edge-weight assignment for tasks beyond contrastive pretraining, such as label propagation or graph clustering; the paper does not explore this direction.","The method's $\\mathcal{O}(N^2)$ per-iteration cost and full-batch softmax over all nodes limit it to moderate-size graphs, so a sampled or blockwise version of the metric network is a natural testable extension.","Because the hinge form resembles margin-based self-training, NML-GCL could inherit known instabilities such as confirmation bias, which the KL regularization only partially addresses; this is an inference, not a claim of the paper."],"forward_implications":["Because the learned loss is a provably tighter MI lower bound than InfoNCE, NML-GCL can replace fixed negative weighting without sacrificing the theoretical grounding of contrastive learning.","At convergence the metric network assigns smaller weights to false negatives and larger weights to true negatives, so the encoder's loss no longer pushes same-label nodes apart; the reported weight curves in the appendix show this on all six datasets.","The bi-level training is self-reinforcing: a better encoder gives cleaner similarity supervision for the metric network, and a better metric network gives sharper push-away instructions for the next encoder update.","Downstream, the method reports higher node classification accuracy and clustering purity than hard-weight, soft-weight, and no-false-negative baselines on Cora, CiteSeer, PubMed, Photo, Computers, and Wiki-CS.","The method needs no human priors such as similarity thresholds or cluster counts to identify false negatives, which is the specific limitation the paper targets."],"supporting_citations":[{"why":"Defines InfoNCE as a variational lower bound on mutual information, the baseline bound the paper claims to tighten.","marker":"[Poole et al., 2019]"},{"why":"Supplies the definition of mutual information and the data-processing inequality used in the proofs of Theorems 1 and 2.","marker":"[Alajaji et al., 2018]"},{"why":"Provides the variational KL-divergence representation used to construct the MI lower bound in the proof of Theorem 1.","marker":"[Nguyen et al., 2010]"},{"why":"Supplies the theoretical basis for self-training, which the bi-level update of encoder and metric network is said to instantiate.","marker":"[Wei et al., 2020]"},{"why":"Defines the GCN architecture used as the encoder in all experiments.","marker":"[Kipf and Welling, 2017]"},{"why":"Provides the GRACE baseline and the augmentation recipe of edge dropping and feature masking that NML-GCL adopts.","marker":"[Zhu et al., 2020]"},{"why":"Provides ProGCL, a soft-weight false-negative baseline that NML-GCL is compared against.","marker":"[Xia et al., 2022]"},{"why":"Provides GRAPE, a soft-weight baseline using subspace-preserving hard negative mining that NML-GCL must outperform.","marker":"[Hao et al., 2024]"},{"why":"Provides LOCAL-GCL, a hard-weight false-negative baseline based on neighborhood distances.","marker":"[Zhang et al., 2022]"},{"why":"Provides GRACE+, a soft-weight false-negative baseline based on node similarity.","marker":"[Chi and Ma, 2024]"}],"fun_headline_variants":["Learn the geometry of negatives, not the thresholds","NML-GCL: false negatives need no hand-picked priors","Graph contrastive learning that ranks negatives like an oracle","A trainable negative metric tightens the MI bound for GCL","GCL learns which negatives to trust, no fixed weights"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theory's guarantee that a trained encoder ranks false negatives closer than true negatives assumes the two augmented views are generated entirely from the node's class label plus independent random noise, so the original graph structure contributes nothing beyond the label; real graphs carry structure beyond labels, so that premise is typically false.","fun_headline_variants_meta":{"raw":{"variants":["Learn the geometry of negatives, not the thresholds","NML-GCL: false negatives need no hand-picked priors","Graph contrastive learning that ranks negatives like an oracle","A trainable negative metric tightens the MI bound for GCL","GCL learns which negatives to trust, no fixed weights"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000187,"raw_usage":{"total_tokens":1337,"prompt_tokens":963,"completion_tokens":374,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":291}},"tokens_in":579,"tokens_out":374,"duration_ms":4203,"temperature":1.0,"reasoning_tokens":291,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:12:27.332640+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run NML-GCL on a heterophilous graph where connected nodes usually have different labels, using the same augmentation recipe, and measure the median distance ratio between false negatives and true negatives as in the paper's Table 3. If the ratio no longer falls below the InfoNCE baseline on such graphs, the label-only Markov-chain assumption behind Theorem 2 is the cause; the paper's current evidence covers only homophilous citation and co-purchase networks.","supporting_citations":[{"cited_title":"On variational bounds of mutual information","cited_arxiv_id":null,"evidence_quote":"Defines InfoNCE as a variational lower bound on mutual information, the baseline bound the paper claims to tighten."},{"cited_title":"An introduction to single-user information theory","cited_arxiv_id":null,"evidence_quote":"Supplies the definition of mutual information and the data-processing inequality used in the proofs of Theorems 1 and 2."},{"cited_title":"Estimating divergence functionals and the likelihood ratio by convex risk min- imization","cited_arxiv_id":null,"evidence_quote":"Provides the variational KL-divergence representation used to construct the MI lower bound in the proof of Theorem 1."},{"cited_title":"Deep graph contrastive representation learning","cited_arxiv_id":null,"evidence_quote":"Provides the GRACE baseline and the augmentation recipe of edge dropping and feature masking that NML-GCL adopts."},{"cited_title":"Progcl: Rethinking hard nega- tive mining in graph contrastive learning","cited_arxiv_id":null,"evidence_quote":"Provides ProGCL, a soft-weight false-negative baseline that NML-GCL is compared against."},{"cited_title":"Towards expansive and adaptive hard negative mining: Graph con- trastive learning via subspace preserving","cited_arxiv_id":null,"evidence_quote":"Provides GRAPE, a soft-weight baseline using subspace-preserving hard negative mining that NML-GCL must outperform."},{"cited_title":"Enhancing contrastive learning on graphs with node similarity","cited_arxiv_id":null,"evidence_quote":"Provides GRACE+, a soft-weight false-negative baseline based on node similarity."}],"review_version":1}