{"id":"451bc710-ce8c-4703-bce6-3acf162b2268","arxiv_id":"2412.17856","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A new framework, ECL-GSR, uses energy-based contrastive learning to refine noisy graph structure and reports state-of-the-art node classification accuracy on eight benchmarks.","lead":"This paper proposes a graph-learning method that learns which links between nodes are trustworthy using a mix of contrastive and energy-based training, then rewires the graph before classifying nodes. It reports top accuracy on eight benchmark graphs while using less training time and memory than several strong baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The generative EBM term as implemented in Eq. (16) is disconnected from the SGLD samples ν* of Algorithm 1, so the 'energy-based generative training' claim is not actually executed; this undermines the theoretical contribution, though the empirical comparisons may survive.","rationale":"The paper is best read as making two connected claims: (i) a methodological claim that ECL is an energy-based generative/discriminative objective, and (ii) an empirical claim that ECL-GSR beats thirteen baselines on eight datasets. The reader's CONDITIONAL verdict already targets the gap between Eqs. (13)-(16) and Algorithm 1, and I agree that this is the most load-bearing weakness. The manuscript itself provides the evidence: Algorithm 1 samples and Langevin-updates ν*, but Eq. (16) is written on z_n = fθ(ν_n), and no subsequent line uses ν* in the loss. A reader cannot reproduce the theoretical objective from the pseudocode; either Eq. (16) is mislabeled or the SGLD loop is dead code. I do not see a similarly concrete flaw in the empirical comparisons: the benchmark tables, t-tests, train-ratio experiments, and ablation study are extensive, and the gains on the small web graphs are within plausible ranges. The method might still work well as a contrastive structure-refinement pipeline; the concern is specifically that the 'energy-based generative training' explanation is not what was optimized. That does not force rejection of the empirical SOTA claim, but it does require the authors to release code and clarify the computation graph. I would keep the reader's CONDITIONAL verdict: acceptance should be conditioned on code release and on either a corrected derivation of Eq. (16) or an implementation that actually uses the SGLD samples.","tokens_in":20751,"tokens_out":7415,"duration_ms":65443,"concrete_test":"Instrument the published implementation (or a faithful reimplementation of Algorithm 1) so that, after the SGLD loop, the loss tensor for Eq. (16) is inspected: check whether the gradient ∂LE/∂ν*_K is nonzero. Because Eq. (16) is evaluated on fθ(ν_n) rather than fθ(ν*_n), this gradient will be exactly zero, which would prove that the SGLD loop is disconnected from the generative term. Then, as a follow-up, retrain with z_n in Eq. (16) replaced by fθ(ν*_n) and compare validation accuracy on Cora and Citeseer; a material change would confirm that the published loss is not the described EBM objective, while no change would indicate SGLD is irrelevant to the reported results.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 explicitly samples {ν*_n} from pd(ν), runs K SGLD updates via Eq. (3), and then says 'Calculate the generative term of Lb with Eq. (16).' But Eq. (16) is written in terms of z_n = fθ(ν_n), the encoder applied to the original positive-pair views, not to the SGLD-refined ν*_n. No term in Eq. (16) or in the surrounding text maps ν* back into the loss. Consequently, the negative phase of the EBM gradient in Eq. (13) — the Epθ[∇θEθ(ν)] term that is the entire reason for running SGLD — is absent from the implemented objective. The implemented 'generative' loss is instead a softmin (negative log-sum-exp) over the N positive-pair distances, a different functional whose gradient is a weighted average of positive-pair energy gradients. If this is the actual training loss, the method reduces to a SimCLR-style contrastive loss plus a positive-pair log-sum-exp term and the L2 regularizer Lr. The paper's central novelty claim ('first work to combine EBMs with CL', 'generative training') therefore rests on an approximation that is neither derived nor validated. The ablation in Fig. 3 ('without gen. term') only removes Eq. (16); it does not test whether SGLD sampling contributes anything. This is load-bearing because the theoretical contribution is the stated basis of the method, even if the empirical SOTA claim could still hold as a structure-refinement pipeline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ECL-GSR, a graph structure refinement framework that combines energy-based models (EBMs) with contrastive learning. The method builds a joint distribution over augmented graph views using an energy function, decomposes the objective into discriminative and generative terms, and uses the learned representations to add or remove edges before training a GNN classifier. The authors report state-of-the-art node classification accuracy on eight benchmark datasets, faster training and lower memory use than leading baselines, plus ablations, robustness tests, statistical tests, and a graph-classification extension. The central theoretical claim is that the generative EBM term is trained via SGLD, making this the first EBM-plus-contrastive-learning method for graph structure refinement.","tokens_in":1520,"tokens_out":1781,"duration_ms":74956,"significance":"If the theoretical claim held, the paper would offer a simple and efficient structure-refinement module with broad applicability: a single unsupervised-style loss that improves GNN node classification across small and large graphs. The empirical package is a genuine strength: eight datasets, standard splits and train-ratio experiments, comparisons against 13 baselines, efficiency and scalability analysis, robustness under edge perturbations, component ablations, and statistical significance tests. The low-resource claims are also interesting and potentially useful. However, the core novelty rests on the generative EBM objective, and the implemented loss in Eq. (16) is not the SGLD-based objective derived in Eq. (13). The paper also does not release code and leaves the graph augmentation operator underspecified, which limits verifiability. The significance is therefore conditional: the empirical results may stand, but the stated theoretical contribution needs either repair or a substantially reframed presentation.","major_comments":[{"comment":"The implemented generative loss is disconnected from the SGLD sampling required by the theory. Algorithm 1 samples {ν*_n} from p_d(ν), runs K SGLD updates via Eq. (3), and then directs the reader to compute the generative term with Eq. (16). However, Eq. (16) is a log-sum-exp over exp(-||z_n - z'_n||^2/τ), where z_n = f_θ(ν_n) and z'_n = f_θ(ν'_n) are the original augmented views; the refined samples ν* never appear in the loss. Consequently, the negative phase E_{p_θ}[∇_θ E_θ(ν)] in Eq. (13) is absent from the optimized objective, and Eq. (13) is not the gradient of what is implemented. The implemented term is instead a softmin over the N positive-pair energies. Please either modify the algorithm so that ν* is actually used in the generative term, provide a proof that Eq. (16) is a valid surrogate for the SGLD-based objective, or explicitly reframe the contribution as contrastive graph structure refinement with a positive-pair log-sum-exp auxiliary loss.","section":"Methodology, Eqs. (13)-(16) and Algorithm 1"},{"comment":"The claim that ECL-GSR is an unsupervised method is contradicted by the training protocol. Eq. (19) minimizes L_E(θ) + µ L_C(θ), where L_C is the cross-entropy loss computed with the labeled nodes Y, and Algorithm 1 explicitly takes Y as input and updates both θ_E and θ_C. Since the encoder f_θ producing the refined structure is updated with L_C, the structure refinement in the reported experiments is supervised. The 'unsupervised' claim should be restricted to the ECL component, or the authors should report a variant trained without L_C to substantiate the abstract's wording.","section":"Abstract and Training Objective, Eq. (19) and Algorithm 1"},{"comment":"The data augmentation operator T is described only as 'a random Gaussian blur' in Section 4.1. For graph data it is not specified whether the blur applies to node features, to the adjacency matrix, or to both, nor how the two views t(g†) and t'(g†) differ. This makes the construction of positive and negative pairs, and the domain on which SGLD in Eq. (3) runs, underspecified. The ambiguity is compounded by the absence of released code. Please provide the exact augmentation recipe, including any parameters and whether it is topology-aware.","section":"Preprocessing and Implementation Details, §3.1 and §4.1"}],"minor_comments":[{"comment":"In the simplification preceding Eq. (16), the expression p_hat_theta(ν_n) = (1/N) Σ_{n=1}^N p_theta(ν_n, ν'_n) reuses the index n on both sides; a different summation index would make the averaging clear.","section":"Eq. (16) and surrounding text"},{"comment":"The text says there are N positive and 2(N-1) negative samples, but the denominator of Eq. (15) appears to include the positive pair among the N views ν'_m; please clarify the counting so the denominator matches the SimCLR-style softmax.","section":"Eq. (14)-(15)"},{"comment":"The efficiency and scalability figure is described as 'nearer to the upper left corner signifies superior overall performance,' but the axes and plotted quantities are not labeled; please define the axes and units.","section":"Figure 2"},{"comment":"The pairwise t-test description does not state whether the compared results are paired or independent across runs, and several P-values are extremely small despite modest accuracy differences; please report effect sizes or confidence intervals.","section":"Table 3"},{"comment":"In the training-stability subsection, the text states α and β are set to 0.1 and 0.001, whereas the main text reports α=0.1, β=0.01; please reconcile these values.","section":"Appendix, Training Stability"}],"recommendation":"major_revision","confidential_remarks":"The empirical contribution could be publishable if the paper is reframed as an efficient contrastive graph structure refinement method with a positive-pair log-sum-exp auxiliary loss. As written, the theoretical novelty claim about SGLD-based energy-based generative training is not supported by the implementation. I would also encourage the editor to require either code release or a fully specified graph augmentation operator, since the reproducibility of the eight-dataset results depends on details that are currently missing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTwo things to know. First, the paper's empirical claim is real: a simple contrastive-learning module for graph structure refinement beats thirteen baselines on eight node-classification datasets, including OGB-Arxiv, and the efficiency gains look genuine. Second, the theoretical framing does not match the implementation. Algorithm 1 samples ν* via SGLD, but Eq. (16) — the 'generative term' it points to — is computed on the original views ν_n, not on ν*. The SGLD samples never enter the loss. So the negative phase of the EBM gradient, the whole reason to run Langevin dynamics, is absent. What is implemented is SimCLR plus a log-sum-exp over positive-pair distances and an L2 regularizer.\n\nWhat is genuinely new: applying energy-based contrastive learning, adapted from Kim and Ye's image work, to graph structure refinement. That combination appears new, and the edge-prediction step through cosine similarity is clean. The experiments are extensive: standard splits, low-label ratios, robustness to edge perturbation, and a graph-classification appendix. The performance gains over second-best are modest (0.15 to 1.61 pp) but consistent. The ablation shows both terms help, though the ablation only removes Eq. (16); it doesn't isolate whether SGLD sampling contributes anything.\n\nThe soft spots are proportionate to the claim. The central novelty claim — 'first work to combine EBMs with CL for GSR' — rests on the generative term being what the theory says it is. As written, it isn't. This needs a fix, not a footnote. Also missing: a precise graph augmentation operator ('random Gaussian blur' on a graph is unclear), and code or data to reproduce the numbers. The t-test appendix is fine but not a substitute for released artifacts.\n\nWho gains from this? People working on graph structure learning will find the empirical recipe useful as a strong baseline. The EBM theory part should be read with suspicion until clarified.\n\nMy recommendation: send it to peer review, but with a request for major revision. The authors should either align the implementation with the derivation, or drop the EBM claim and present the method as a contrastive structure-refinement pipeline. The empirical work deserves referee time; the theory as presented does not.","headline":"Strong empirical structure-refinement results undercut by a load-bearing mismatch: the implemented generative loss never uses the SGLD samples the theory requires.","tokens_in":21659,"tokens_out":3765,"would_cite":true,"duration_ms":34202,"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":"ECL-GSR claims state-of-the-art node classification on eight benchmarks by repairing graph structure with an energy-based contrastive objective, and reports faster training with fewer samples and less memory.","keywords":["graph structure refinement","energy-based models","contrastive learning","graph neural networks","node classification","unsupervised representation learning","stochastic gradient Langevin dynamics","edge prediction"],"falsifier":"Train the same ECL-GSR pipeline twice on Cora, once with the implemented generative loss of Eq. (16) and once with the SGLD-based gradient of Eq. (13) using the sampled views $\\nu^*$ from Algorithm 1; if the two versions reach nearly identical accuracy and refined graphs, then the implemented objective does not actually depend on the energy-model sampler, and the paper's generative-training claim is not doing the work attributed to it.","tokens_in":20601,"feed_emoji":"🕸️","tokens_out":9622,"duration_ms":72885,"temperature":0.7,"pith_summary":"Real-world graphs arrive with missing and spurious edges, and message-passing GNNs amplify that noise. This paper proposes ECL-GSR, an unsupervised structure-refinement module that learns node representations with an energy-based contrastive objective and then rewires the graph by adding or deleting edges according to representation similarity. The paper claims that on eight node-classification benchmarks this refined graph yields state-of-the-art accuracy, with faster convergence and lower memory than leading structure-learning baselines, and that the method is especially strong with very few training labels and under random edge perturbations. It also argues that the standard contrastive loss is a special case of the proposed energy-based objective when the generative term is switched off. The combination matters because structure refinement is a bottleneck for deploying GNNs on noisy or incompletely observed networks.","feed_headline":"Energy-based graph repair beats state of the art on eight benchmarks","feed_subtitle":"An unsupervised hybrid cleans noisy edges, needs fewer labels, and scales to large graphs.","key_machinery":"The engine is the ECL loss built on the energy $E_\\theta(\\nu,\\nu') = \\lVert z - z' \\rVert^2 / \\tau$ over pairs of augmented views of a dual-attribute graph, whose node features concatenate raw attributes with DeepWalk structural embeddings. Bayes' rule decomposes the objective into the discriminative conditional likelihood and the generative marginal likelihood; the former is implemented as the contrastive loss of Eq. (15), the latter as the positive-pair log-sum-exp of Eq. (16), and SGLD is the sampler behind Eq. (13). A normalized-cosine edge predictor plus relaxed Bernoulli sampling turns learned representation similarities into a refined adjacency matrix, and the full objective adds an L2 energy regularizer and a cross-entropy classification term.","core_discovery":"ECL-GSR models the joint distribution of two augmented views of a graph as an energy-based model, with energy given by the squared distance between their representations divided by a temperature. Using Bayes' rule, the negative log-likelihood separates into a conditional term, which becomes a SimCLR-style contrastive loss, and a marginal term, which acts as a generative loss shaped by stochastic-gradient Langevin dynamics; the paper proves that the marginal distribution of a view is itself an energy-based model. After this representation learning, an edge predictor computes normalized cosine similarities between node embeddings and binarizes them with relaxed Bernoulli sampling to produce a refined adjacency matrix, and a three-layer GNN classifies nodes on the refined graph. The paper's empirical claim is that this two-step pipeline outperforms thirteen structure-learning baselines on eight datasets, and its theoretical claim is that the discriminative contrastive loss is a special case of the ECL objective when the generative term is removed.","pith_inferences":["Inference beyond the paper: because Eq. (16), the implemented generative loss, contains no dependence on the views $\\nu^*$ produced by Algorithm 1's SGLD loop, the practical method may be equivalent to a SimCLR-style contrastive loss plus an L2 regularizer on positive-pair energies; an ablation that replaces Eq. (16) with the true SGLD gradient would settle whether the energy-based sampler contrib","Inference beyond the paper: the similarity-based edge-prediction rule is task-agnostic, so the framework should transfer to link prediction and graph classification; the paper's appendix results on three graph-classification datasets are consistent with, but do not prove, that transfer.","Inference beyond the paper: on the three heterophilic webpage graphs, where raw homophily is below 0.2, the method still gains, suggesting the dual-attribute features learn semantic similarity that overrides noisy topology; a direct test would compare the method's added edges against oracle within-class edges to identify whether gains come from adding intra-class links or deleting inter-class nois"],"forward_implications":["If the central claim holds, noisy or incomplete graphs can be repaired without labels, and the repaired graph alone is enough to lift GNN node-classification accuracy above previous structure-learning methods on eight benchmarks.","The reported label-efficiency results mean that in semi-supervised settings with 1% training data, ECL-GSR outperforms baselines, so the method is a candidate for graphs with very few labels.","Because training uses mini-batch subgraphs and only three SGLD steps, the approach scales to large graphs like Pubmed and OGB-Arxiv, where several competing methods run out of memory.","The ablation studies imply that the value is in the combination: disabling either the generative or the discriminative term measurably hurts accuracy.","The robustness experiments suggest that the refined structure survives random edge additions and removals up to 80% intensity, and the learned graph keeps inter-class edges sparser than intra-class edges."],"supporting_citations":[{"why":"Supplies the energy-based contrastive learning formulation and the empirical-mean estimator used in Eq. (14).","marker":"Kim and Ye 2022"},{"why":"Provides the SimCLR contrastive loss that Eq. (15) uses for the discriminative term.","marker":"Chen et al. 2020"},{"why":"Defines the stochastic-gradient Langevin-dynamics sampler in Eq. (3), the generative sampling step of ECL.","marker":"Welling and Teh 2011"},{"why":"Gives the relaxed Bernoulli sampling used to binarize the predicted edge-probability matrix.","marker":"Zhao et al. 2021"},{"why":"Frames graph structure refinement as the problem ECL-GSR solves.","marker":"Zhu et al. 2021"},{"why":"Supplies the LDS bilevel structure-learning baseline that ECL-GSR must beat.","marker":"Franceschi et al. 2019"},{"why":"SUBLIME is the unsupervised contrastive structure-learning baseline that defines the comparison setting.","marker":"Liu et al. 2022b"},{"why":"Provides the GSLB benchmark reimplementation and experimental protocol used for all comparisons.","marker":"Li et al. 2023"},{"why":"GCN is the backbone encoder of ECL-GSR and the primary plain-GNN baseline.","marker":"Kipf and Welling 2016"}],"fun_headline_variants":["Energy-based contrastive learning repairs noisy graphs","ECL-GSR: Unsupervised graph repair with fewer labels","Graph refinement via energy-based contrastive learning beats baselines","Energy-based learning cleans graphs for better node classification","Unsupervised energy-based graph refinement outperforms on 8 datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the approximated generative term in Eq. (16) delivers the same benefit as actually sampling from the energy model in Eq. (13), even though the implemented loss never uses the sampled views that Algorithm 1 produces.","fun_headline_variants_meta":{"raw":{"variants":["Energy-based contrastive learning repairs noisy graphs","ECL-GSR: Unsupervised graph repair with fewer labels","Graph refinement via energy-based contrastive learning beats baselines","Energy-based learning cleans graphs for better node classification","Unsupervised energy-based graph refinement outperforms on 8 datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000606,"raw_usage":{"total_tokens":2832,"prompt_tokens":956,"completion_tokens":1876,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":1796}},"tokens_in":572,"tokens_out":1876,"duration_ms":14958,"temperature":1.0,"reasoning_tokens":1796,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:19:45.814777+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same ECL-GSR pipeline twice on Cora, once with the implemented generative loss of Eq. (16) and once with the SGLD-based gradient of Eq. (13) using the sampled views $\\nu^*$ from Algorithm 1; if the two versions reach nearly identical accuracy and refined graphs, then the implemented objective does not actually depend on the energy-model sampler, and the paper's generative-training claim is not doing the work attributed to it.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the energy-based contrastive learning formulation and the empirical-mean estimator used in Eq. (14)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the stochastic-gradient Langevin-dynamics sampler in Eq. (3), the generative sampling step of ECL."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the relaxed Bernoulli sampling used to binarize the predicted edge-probability matrix."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LDS bilevel structure-learning baseline that ECL-GSR must beat."}],"review_version":1}