{"id":"2c39a7e4-3351-4b6c-888e-34974db4675e","arxiv_id":"2411.17236","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A class-conditioned graph diffusion model trained with a discriminative ELBO loss and permutation majority voting achieves competitive graph classification accuracy.","lead":"This paper turns a graph diffusion model into a classifier by training it with a discriminative objective based on the model's own likelihood lower bound, then voting over random node permutations at test time. On three small graph benchmarks the method matches or slightly beats standard graph neural network classifiers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (15) states the ELBO inequality in the wrong direction; as written, LCLF and Equation (19) do not approximate class-conditional likelihood, so the reported SOTA is not tied to the stated objective.","rationale":"The reader's weakest_assumption identifies the ELBO bound in Equation (15) as load-bearing and notes that the inequality appears to have the wrong direction. My reading agrees: this is the single most important technical flaw because every subsequent construction (LCLF in Equation (16), the claimed bound in Equation (18), and approximate inference in Equation (19)) inherits the sign error. If the sign is wrong, the paper does not actually show that the discriminative training objective optimizes classification accuracy, and the state-of-the-art claim is detached from the formal derivation. I do not escalate to REJECT because the empirical results in Table 2 could still be valid if the implementation uses the corrected sign; the paper is a workshop-level empirical contribution, and a small typo or a missing minus sign would resolve the inconsistency. The appropriate response is to keep the CONDITIONAL verdict, conditional on code release, sign clarification, and a matched-protocol comparison against baselines, exactly as the reader proposed. The concrete test above is cheap and decisive: it either confirms the typo or shows that the published equations do not describe the implemented method.","tokens_in":15614,"tokens_out":5345,"duration_ms":52622,"concrete_test":"Re-derive Equation (15) from the ELBO in Equation (2) for the Gaussian diffusion used in Section 3.3, showing that log p(A|y) >= -LDEN(A,y,theta) + C rather than LDEN(A,y,theta) <= ln p(A|y). Then inspect the released code (or, if code is unavailable, the training loop description) to determine whether the softmax in LCLF receives LDEN or -LDEN. If the code uses -LDEN, Equation (16) is a typo and the concern is resolved. If it uses LDEN, run a diagnostic on IMDB-BINARY validation graphs: compute per-class LDEN and per-class ODE-based log-likelihood from Equation (20), and compare their rankings; if LDEN and log-likelihood are negatively correlated, Equation (19) cannot be the source of the reported accuracy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on treating LDEN(A,y,theta) in Equation (15) as an approximation to ln p_theta(A|y) and on LCLF in Equation (16) as a discriminative training objective. For any diffusion model, the standard ELBO is log p(x0) >= E_q[log p_theta(x0:T)/q(x1:T|x0)], and the denoising MSE in Equation (3)/(12) is, up to constants, the negative of that ELBO. Hence the correct inequality is log p(A|y) >= -LDEN(A,y,theta) + C, not LDEN(A,y,theta) <= ln p_theta(A|y) as printed. Equation (18) therefore does not establish that LCLF is a lower bound on the training cross-entropy, and Equation (19) replaces ln p(A|y) with -LCLF = ln Softmax(LDEN)_y, which is a posterior probability, not a class-conditional likelihood. Read literally, minimizing LCLF pushes the true class's denoising MSE upward relative to other classes, which should worsen likelihood-based classification rather than improve it. Table 2 shows LCLF performing best, so either the implementation feeds -LDEN into the softmax (making Equation (16) a sign typo) or the reported results come from a different objective than the one derived. Either way, the formal justification of the central claim is internally inconsistent, and the empirical headline is not connected to the stated method without clarification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper adapts the SwinGNN score-based graph diffusion model to graph classification. It introduces a class-conditional diffusion model trained with three objectives (the pure denoising loss LDEN, a discriminative loss LCLF, and their sum LSUM) and combines them with two inference schemes (approximate inference using the variational loss and ODE-based likelihood computation), together with random permutation augmentation at both training and test time and majority voting over permutations. Experiments on K-Regular, IMDB-BINARY, IMDB-BINARY with features, and PROTEINS report that LCLF trained with approximate inference and permutation voting achieves the best accuracy on all four dataset configurations and is state of the art compared with GNN baselines from Errica et al. (2019).","tokens_in":15883,"tokens_out":4968,"duration_ms":45264,"significance":"If the results hold, the paper offers a useful demonstration that class-conditional graph diffusion models can serve as generative classifiers for graph-level tasks, complementing analogous results in the image domain. The permutation-based augmentation and the use of the variational loss for efficient checkpoint selection are practical ideas that could benefit future work. However, the central theoretical derivation contains a sign error in the ELBO that breaks the formal connection between the proposed objective and likelihood-based classification, and the empirical protocol deviates from the GNN baselines in several ways. The paper does not release code or provide statistical significance tests, so the state-of-the-art claim is not fully supported.","major_comments":[{"comment":"The ELBO inequality is given in the wrong direction. For a denoising diffusion model, the evidence lower bound has the form log p(x) >= -E[||D_theta(x_tilde, t) - x||^2] + C, so the denoising loss LDEN is an upper bound on the negative log-likelihood, not a lower bound on the log-likelihood as written in Eq. (15). Consequently, placing LDEN directly inside the softmax in Eq. (16) makes minimizing LCLF push the true class's denoising loss upward relative to other classes, which is the opposite of likelihood-based classification. Eq. (18) therefore does not establish that LCLF is a lower bound on the training cross-entropy, and Eq. (19) is also inconsistent because -LCLF is a posterior log-probability, not a class-conditional log-likelihood. The empirical success of LCLF in Table 2 suggests that the implementation actually feeds -LDEN into the softmax; if so, the equations must be corrected and the corrected objective must be stated explicitly. As written, the formal justification of the central claim is internally inconsistent.","section":"Section 4, Eqs. (15)-(18), and Section 5, Eq. (19)"},{"comment":"The evaluation protocol is not directly comparable to the baselines. Hyperparameters are selected on a single fold and then fixed for all folds, only one training run per fold is performed instead of three as in the Errica et al. protocol, and graphs above a size cutoff are removed from training and then counted as incorrect at test time. Each of these deviations can change accuracy by more than the small margins over the best baselines in Table 1 (roughly 0.6-1.7 points). To support the state-of-the-art claim, the authors should either follow the reference protocol exactly or quantify the effect of each deviation, for example by running per-fold hyperparameter selection and multiple seeds on at least one dataset.","section":"Section 6 and Appendix A.1.2"},{"comment":"The reported improvements over the best GNN baselines are within one standard deviation for IMDB-BINARY (70.5 +/- 5.7 vs. 69.9 +/- 4.6) and PROTEINS (75.4 +/- 3.4 vs. 73.7 +/- 3.5), and no significance tests or per-fold win/loss counts are reported. Calling these results state-of-the-art overstates the statistical evidence. I recommend reporting paired per-fold accuracies with a significance test and tempering the abstract's claim accordingly.","section":"Section 6.1, Table 1"}],"minor_comments":[{"comment":"The sentence 'For non-graph i.i.d. data, it is well known that a classification model can be derived from a generative model (?)' contains a missing citation placeholder; please insert the intended reference.","section":"Section 2"},{"comment":"The surrounding text says 'for all i' but the formula uses the index j; please unify the notation for class labels.","section":"Section 4, Eq. (14)"},{"comment":"Exact inference is described as using t=4 rather than integrating to the maximum noise level; given this truncation, calling the ODE-based procedure 'exact inference' in Section 5 and Table 2 is misleading. Please either rename it or justify why the truncated likelihood is sufficient.","section":"Appendix A.1.2"},{"comment":"There are numerous typos and formatting issues, including 'repamatrize', 'ususally', 'trianing', 'permutaiton', the uncited '?' in Section 2, and the garbled 'ORGNN' in Figure 2; these should be corrected.","section":"Throughout"},{"comment":"Figure 4 shows accuracy curves without error bars or per-fold variability, although Section 6.2 notes that the effects depend on the data split; adding confidence bands would make the trend more informative.","section":"Section 6.2, Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Eqs. (15)-(18) is the most serious issue: before the revision can be evaluated, the authors must confirm which objective was actually optimized in the experiments. Without a code release, the empirical results alone cannot resolve whether the implementation matches the corrected or the printed equations. If the implementation uses -LDEN in the softmax, the paper is likely salvageable with a corrected derivation; if it matches the printed equations, the empirical claims do not support the proposed method. The paper's prior appearance at an ICML 2024 workshop may also be relevant to the journal's prior-publication policy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this paper's stated objective doesn't do what it claims, because Eq. (15) has the ELBO inequality backward. For any diffusion model, the denoising loss LDEN is (up to constants) the negative ELBO, so the correct bound is log p(A|y) >= -LDEN + C, not LDEN <= log p(A|y). As written, LDEN is an upper bound on negative log-likelihood, not a lower bound on log-likelihood. That makes Eq. (18) false and Eq. (19) wrong—LCLF is a softmax over denoising losses, not a likelihood, and -LCLF is a log posterior, not a class-conditional log-likelihood. If the implementation actually feeds -LDEN into the softmax, the equations are a sign typo; if not, the reported SOTA comes from a different objective than the one derived. Either way, the central justification needs a fix.\n\nThat's the load-bearing flaw. The idea itself is worth taking seriously. To my knowledge it's the first attempt to turn a score-based graph diffusion model into a generative classifier, and the permutation-based training/test augmentation is a sensible response to SwinGNN's lack of invariance. The empirical results are modest but real-looking: LCLF with approximate inference wins on all four dataset configurations, albeit by 0.6 to 1.7 points over GNN baselines.\n\nThe protocol has a few wrinkles. They use one training run per fold instead of three (Errica's protocol), pick hyperparameters on one fold rather than per fold, and omit large graphs from training. But they count omitted graphs as incorrect at test time, which is conservative, not favorable. So the \"SOTA\" claim is weaker than it looks, not stronger.\n\nThe paper needs code release and a corrected derivation before the results can be trusted. If the sign is just a typo, the method is plausible; if not, the experimental numbers might be real but unexplained.\n\nMy take: send it to review anyway—the idea is new and the fix is tractable—but the authors need to clean up the math and match the baseline protocol. I wouldn't cite it in its current form.\n\nBest,","headline":"A clear sign error in the central ELBO inequality guts the stated objective's formal justification, but the graph-diffusion-classifier idea is novel and worth a corrected resubmission.","tokens_in":16447,"tokens_out":3880,"would_cite":false,"duration_ms":32400,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A score-based graph diffusion model, trained with a softmax over per-class ELBOs and combined with permutation majority voting, achieves state-of-the-art graph classification accuracy on the evaluated benchmarks.","keywords":["graph classification","diffusion models","generative classifiers","score-based diffusion","ELBO","permutation invariance","graph neural networks"],"falsifier":"For a set of held-out graphs, compute the exact class-conditional log-likelihoods via the ODE solver (Eq. 20) and the per-class ELBOs $L_{DEN}$; if the two produce different class rankings on a non-negligible fraction of graphs, the central approximation is falsified. An even simpler check: the same model should fail to beat a degree-histogram or graphlet-count baseline on a dataset where those features separate classes.","tokens_in":15388,"feed_emoji":"🔀","tokens_out":4958,"duration_ms":40118,"temperature":0.7,"pith_summary":"The paper asks whether generative graph diffusion models can classify graphs as well as discriminative graph neural networks. It finds that a purely generative training objective does not work for classification, so it introduces a discriminative objective: train the diffusion model so that the per-class ELBOs, placed through a softmax, predict the class. With this objective, random permutation augmentation at training and majority voting over permutations at inference, the model beats the GNN baselines on IMDB-BINARY, IMDB-BINARY with node features, and PROTEINS, and reaches perfect accuracy on a synthetic K-regular graph task. The result matters because it shows a generative model can match or exceed discriminative classifiers in the graph domain, opening a new route for graph classification.","feed_headline":"Graph diffusion model beats GNNs when trained discriminatively","feed_subtitle":"Softmax over class-conditional diffusion ELBOs plus permutation voting sets new accuracy marks on IMDB-BINARY and PROTEINS.","key_machinery":"The load-bearing object is the class-conditional denoising ELBO $L_{DEN}(A,y,\\theta)$, a variational lower bound on the log-likelihood of an adjacency matrix under a class-conditioned score-based diffusion model. The paper inserts the vector of per-class ELBOs into a softmax (Eq. 16) to form $L_{CLF}$, so training pushes the relative ELBO values to rank the true class first, and the same vector is used for approximate inference. Around this, random permutation sampling replaces one adjacency matrix by samples from its isomorphism class, with predictions aggregated by majority vote, compensating for the non-permutation-invariant SwinGNN backbone.","core_discovery":"The central claim is that a score-based graph diffusion model can act as a generative classifier that outperforms discriminative GNN baselines, provided classification is driven by the lower-bound objective $L_{CLF}$ rather than by the generative ELBO. The paper proposes approximating the true class-conditional log-likelihood $\\ln p(A|y)$ with the denoising ELBO $L_{DEN}(A,y,\\theta)$ from the diffusion model, then defining the training loss as the softmax cross-entropy over the vector of per-class ELBOs. At inference it estimates $\\ln p(A|y) \\approx -L_{CLF}(A,y,\\theta)$ (approximate inference) or computes the exact likelihood via ODE integration. On the evaluated datasets, the combination of $L_{CLF}$ training with approximate inference and permutation majority voting achieves the best accuracy among the six design choices and exceeds the GNN baselines from Errica et al. (2019), for example 75.4% on PROTEINS and 72.6% on IMDB-BINARY with features. The paper also reports that purely generatively trained diffusion models, matched with exact likelihood inference, perform poorly as zero-shot classifiers.","pith_inferences":["If the ELBO ranks classes correctly on larger benchmark suites, the same recipe could make diffusion-based generative classifiers competitive on molecular or social graph datasets without task-specific GNN design.","A direct test of the ranking property would compare per-class ELBOs against exact ODE log-likelihoods on held-out graphs; disagreement would break the approximation.","The permutation majority voting suggests a calibration strategy: variance across permutations could serve as a proxy for classification uncertainty, which the paper does not explore.","Fine-tuning a pre-trained generative diffusion model with $L_{CLF}$, which the paper names as future work, would test whether generation quality and classification accuracy can be obtained jointly in one model."],"forward_implications":["A generative graph diffusion model trained with $L_{CLF}$ can serve as a standalone classifier that beats the GNN baselines on IMDB-BINARY, IMDB-BINARY with features, and PROTEINS.","Approximate inference with the $L_{CLF}$ objective is over 2000 times faster for checkpoint selection than exact ODE-based likelihood, making model selection feasible in practice.","Increasing the number of inference-time permutations from 1 to 5 yields 2 to 10 percentage points of accuracy gain, with further gains saturating around 100 permutations.","Purely generative training with the $L_{DEN}$ objective and exact likelihood gives weak classification accuracy, so the discriminative objective is necessary for classification."],"supporting_citations":[{"why":"It supplies the SwinGNN backbone architecture that the method adapts for graph classification.","marker":"(Yan et al., 2023)"},{"why":"It establishes the generative-classifier paradigm of using diffusion likelihoods for classification and provides the ODE likelihood formula the paper builds on.","marker":"(Zimmermann et al., 2021)"},{"why":"It provides the score-based SDE formulation and the likelihood computation via probability flow ODE used for exact inference.","marker":"(Song et al., 2021)"},{"why":"It defines the evaluation protocol, dataset splits, and GNN baselines that the paper compares against.","marker":"(Errica et al., 2019)"},{"why":"It supplies the preconditioning and noise schedule used for training the denoising network.","marker":"(Karras et al., 2022)"}],"fun_headline_variants":["Diffusion model beats GNNs on graph classification","Graph diffusion classifier: discriminative loss wins","Diffusion ELBO loss sets new graph accuracy bar","Score-based diffusion for graph classification beats GNNs","Discriminative diffusion training lifts graph classification accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The per-class ELBO is a tight enough lower bound on the true class-conditional log-likelihood that sorting the ELBOs across classes reproduces the true posterior ranking; if the bound is loose or orders classes differently, the $L_{CLF}$ objective stops tracking classification accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model beats GNNs on graph classification","Graph diffusion classifier: discriminative loss wins","Diffusion ELBO loss sets new graph accuracy bar","Score-based diffusion for graph classification beats GNNs","Discriminative diffusion training lifts graph classification accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000413,"raw_usage":{"total_tokens":2119,"prompt_tokens":914,"completion_tokens":1205,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":1132}},"tokens_in":530,"tokens_out":1205,"duration_ms":8697,"temperature":1.0,"reasoning_tokens":1132,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:21:35.000668+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a set of held-out graphs, compute the exact class-conditional log-likelihoods via the ODE solver (Eq. 20) and the per-class ELBOs $L_{DEN}$; if the two produce different class rankings on a non-negligible fraction of graphs, the central approximation is falsified. An even simpler check: the same model should fail to beat a degree-histogram or graphlet-count baseline on a dataset where those features separate classes.","supporting_citations":[{"cited_title":"Swingnn: Rethinking permutation invariance in diffusion models for graph generation, 2023","cited_arxiv_id":null,"evidence_quote":"It supplies the SwinGNN backbone architecture that the method adapts for graph classification."}],"review_version":1}