{"id":"51c2d12d-2527-475e-9265-aa3c48ee6783","arxiv_id":"2412.11550","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A graph clustering framework that uses semantic prototypes, cross-view assignment prediction, and Fused Gromov-Wasserstein optimal transport to improve cluster separability and beat prior methods on nine benchmarks.","lead":"This paper introduces THESAURUS, a graph clustering method that sorts nodes by comparing two altered views of a graph and matching them to learned 'prototype' clusters using optimal transport, then finishes with K-means on the resulting context-aware embeddings. It reports consistent gains over prior graph clustering methods on nine benchmark datasets, with the largest improvements on imbalanced classes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (12) updates the prototype marginal with the unnormalized count vector P^T 1_N, whose mass is N, contradicting the unit-mass constraint in Eq. (2); if the released code implements this literally, the FGW-OT target Q is infeasible, so the claimed structural assignment is not the one described.","rationale":"I read the paper as proposing a self-supervised target Q generated by FGW-OT; the central claim requires Q to be a feasible optimal coupling. Eq. (12) as written breaks feasibility: the prototype marginal is updated with a vector of mass N rather than 1, so no coupling with the node marginal μ exists. This is a more concrete and formally checkable flaw than the reader's stability concern about the self-referential prototype graph, though both point at the reliability of the FGW target. I am not alleging any misrepresentation; the released code may silently normalize ν, in which case the issue is a documentation/description error rather than an invalid method. Either way, the current text cannot be used to reproduce or even instantiate the claimed training objective, and this should be settled before the headline result is fully accepted. Because the reader already assigned CONDITIONAL and this concern reinforces, rather than redirects, the need for conditions, I keep the verdict unchanged.","tokens_in":22816,"tokens_out":8240,"duration_ms":82230,"concrete_test":"Read the released code at github.com/bwdeng20/THESAURUS and locate the implementation of Eq. (12)/Algorithm 1 line 15. Check whether ν is renormalized to unit sum (e.g., ν = ν / ν.sum(), or P^T 1_N is divided by N) before being passed as the marginal in FGW-OT. If renormalization is absent, rerun the Cora and Pubmed experiments with a properly normalized ν and compare the four metrics in Table 1; if renormalization is present, correct Eq. (12) in the paper and confirm that the stated Q solves Eq. (9).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (12) defines ν(t) = β2 ν(t−1) + (1−β2)(P^T 1_N). Since P in Eq. (3) is row-stochastic, P^T 1_N is a vector of soft class counts summing to N. With ν(0) = 1_S/S (sum 1), the momentum update is not mass-preserving: ν(t) converges to P^T 1_N, which has total mass N, not 1. But Eq. (2) and Definition 2 require ν to be a probability vector and require a coupling π with π 1_S = μ and π^T 1_N = ν, where μ also sums to 1. No such coupling exists when the two marginals have different total mass. Therefore the FGW-OT problem in Eqs. (8)-(10) is infeasible as stated, and the row-normalized assignment Q used in the cross-entropy loss Eq. (4) is not the optimal coupling claimed. If the released implementation normalizes ν after the update (or divides P^T 1_N by N) without saying so, the paper's central description is incomplete and the published numerics cannot be reproduced from the text; if it does not normalize, the training target is not a valid OT coupling and the mechanism behind the reported separability gains is not the one claimed. This is independent of the additional stability concern that B = P^T P is self-referential; it makes the OT target formally ill-posed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes THESAURUS, an unsupervised graph node clustering method that combines semantic prototypes, cross-view assignment prediction, and Fused Gromov-Wasserstein Optimal Transport (FGW-OT) to align a data graph with a prototype graph. The prototype graph and prototype marginal distribution are updated by momentum from the model's own soft assignments. The paper claims state-of-the-art clustering performance on nine benchmark datasets and argues that the method mitigates the Uniform Effect and Cluster Assimilation by improving cluster separability.","tokens_in":23217,"tokens_out":4480,"duration_ms":45337,"significance":"If correct, the method would be a strong general-purpose graph clustering pipeline: it unifies attribute and structural information in a single OT-based training objective, releases code, and reports large gains over Dink-Net (e.g., Pubmed F1 from 67.35 to 79.00, ARI from 30.44 to 48.25). The careful class-wise analysis of Uniform Effect and Cluster Assimilation on Cora and Pubmed is a useful diagnostic contribution. However, the central OT construction is formally inconsistent as written, and the structural target is primarily self-generated, so the claimed mechanism needs substantial clarification and additional validation before the empirical results can be taken as evidence for the proposed approach.","major_comments":[{"comment":"The prototype marginal update in Eq. (12) is {ν(t) = β2 ν(t−1) + (1−β2)(Pᵀ 1_N)}. Since P is row-stochastic, Pᵀ 1_N is a vector of soft class counts summing to N. With ν(0) = 1_S/S (mass 1), the momentum recurrence has total mass converging to N, not 1. However, Eq. (2) constrains π by π 1_S = μ and πᵀ 1_N = ν, and Definition 2 requires ν to be a probability measure. For μ of mass 1 and ν of mass N, no feasible coupling exists, so the FGW-OT problem in Eqs. (8)–(10) is infeasible as written, and the row-normalized Q in Eq. (10) is not an optimal OT coupling. The paper must state explicitly what normalization is applied after Eq. (12), or divide Pᵀ 1_N by N. If the released code already performs such normalization, the text is incomplete; if it does not, the training target is not a valid OT coupling and the reported numbers are not reproducible from the description.","section":"Eq. (12) and Algorithm 1"},{"comment":"The prototype graph is defined as B = Pᵀ P in Eq. (5) and updated in Eq. (11) as a momentum average of the model's own soft-assignment co-occurrence; similarly, ν in Eq. (12) is derived from P. Thus the structural target matched by GW-OT is not an independent source of cluster or structural information: it is a function of the current model's predictions. Cross-view swapping and momentum smoothing reduce but do not remove this self-reference. This weakens the claim in Section 3 that GW-OT 'exploits cluster information in the graph structure' (Limitation L3). Please provide a stability or collapse analysis, or an ablation that replaces B and ν with a fixed structure-derived target (e.g., a coarsened adjacency matrix or an isolated-graph target) and compare the resulting performance. A concrete check is to train with random or constant B and ν and measure the performance drop; without such a check, the structural-alignment mechanism cannot be distinguished from self-confirmation.","section":"Eqs. (5), (11), (12)"},{"comment":"The central empirical claim is consistent improvement over Dink-Net and other baselines across nine datasets, but the main results in Table 1 for Cora, Citeseer, Pubmed, A-Photo, and CoraFull are reported without variance or multiple runs, while Tables 4–5 report five-run statistics only for ACM, DBLP, UAT, and Wiki. In addition, Table 3 lists per-dataset hyperparameters (S, α, τ, pe, px, T, lr, wd) but no validation protocol is described for selecting them. Without multi-run statistics for all datasets and an explicit model-selection rule, it is difficult to assess whether the large reported gains (e.g., Pubmed F1 79.00 vs 67.35) are robust or reflect selection on the test set. Please report means and standard deviations for all nine datasets, or justify the single-run protocol, and specify the validation procedure used to choose the hyperparameters.","section":"Section 4 / Table 1 / Table 3"}],"minor_comments":[{"comment":"Line 28 of Algorithm 1 refers to 'Qτ_1' and 'Qτ_2', but the loss in Eq. (4) uses Q_1, Q_2, Pτ_1, and Pτ_2; the notation should be made consistent and the quantities defined.","section":"Algorithm 1"},{"comment":"Definition 2 writes the FGW objective as an infimum over a p-th root expression minus εH(π). The placement of the entropy term with respect to the power 1/p is nonstandard and ambiguous; please clarify the exact regularized objective solved in practice.","section":"Eq. (9)"},{"comment":"Lines 9–16 update B(t−1) and ν(t−1) after processing view 1 and before processing view 2, so the two views use asymmetric prototype-graph/marginal states. Please state whether this sequential asymmetry is intentional and how it affects the cross-view assignment prediction.","section":"Algorithm 1"},{"comment":"The figure legend includes a 'Sphere' operation that is not described in the main text; please specify whether node representations or prototypes are L2-normalized, and where this normalization occurs in the pipeline.","section":"Figure 2"},{"comment":"The method name is spelled inconsistently as 'Dink-Net' and 'DinkNet' across the text, tables, and figures; please standardize the spelling.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The formal issue with Eq. (12) is verifiable against the released code and is likely fixable by adding an explicit normalization step. The self-referential prototype graph is a deeper conceptual concern: the paper's central narrative is that GW-OT injects structural cluster information, but B is built from the model's own assignments, so the attribution of the gains to structure alignment is not yet established. The empirical claims also need multi-run statistics for the main datasets. These are load-bearing but addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: this is a real contribution to deep graph clustering, with a genuinely novel synthesis of semantic prototypes, cross-view assignment prediction, and Fused Gromov-Wasserstein OT. The empirical story is strong and mostly consistent with the claims. But the paper has a formal inconsistency in its central OT update that needs to be addressed before the theory matches the implementation — and the authors never tell us whether their code normalizes the marginal.\n\nWhat's new: the TSA module, which uses a learnable prototype graph updated by momentum as the structural target for FGW-OT, is not a trivial combination of SwAV and GW-OT. The context-aware representation derived from node-to-prototype similarities is a good idea, and the ablations in Table 2 support the design choices. The experiments span nine datasets, and the appendix includes 5-run means and standard deviations for four of them; the improvements over Dink-Net, especially on Pubmed and Cora, look real even allowing for variance.\n\nThe main soft spot is Eq. (12). The momentum update for the prototype marginal ν(t) = β2 ν(t−1) + (1−β2)(P^T 1_N) adds mass N, not 1, since each row of P sums to 1. Yet Eq. (2) and Definition 2 require ν to be a probability vector. So the FGW-OT problem is infeasible as stated; no coupling exists with marginals of different total mass. Either the code renormalizes ν after the update (likely) or the training target is not actually an OT coupling. The paper needs to state this explicitly. This is separate from the self-referential issue that B = P^T P depends on the model's own soft assignments; that is a known cost of self-supervised targets and is partly mitigated by momentum and cross-view swapping, but a collapse analysis would help.\n\nThe main table omits variance for five datasets, and the per-dataset hyperparameters are tuned without a described validation protocol. These are minor-to-moderate reproducibility concerns.\n\nWho this is for: anyone working on graph clustering or OT-based representation learning. It deserves a serious referee: the method is novel, the experiments are extensive, and the formal issue is fixable. I would accept it for review with the expectation of a major revision clarifying the marginal normalization and adding variance and robustness analysis.","headline":"A useful, well-engineered graph clustering method whose main innovation is undercut by an unnormalized marginal in the OT update; worth reviewing, but the formal issue needs fixing.","tokens_in":23720,"tokens_out":4213,"would_cite":true,"duration_ms":36386,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","62H30","90C08"],"pacs":[],"model":"deepseek-v4-flash","headline":"THESAURUS claims that swapping fused Gromov-Wasserstein couplings between two augmented views of a graph—using semantic prototypes and a momentum-updated prototype graph as the target—yields higher cluster separability than existing graph…","keywords":["graph clustering","optimal transport","Gromov-Wasserstein","contrastive learning","semantic prototypes","graph neural networks","unsupervised learning","cluster separability"],"falsifier":"Run the same training recipe but replace the momentum-updated prototype graph $B$ with a fixed, hand-specified structural target such as an isolated graph with $C$ components; if the reported accuracy and Macro-F1 gains over Dink-Net persist, then the prototype-graph mechanism is not what produces the cluster separability the paper attributes to it.","tokens_in":22635,"feed_emoji":"🕸️","tokens_out":6115,"duration_ms":51173,"temperature":0.7,"pith_summary":"Graph node clustering usually trains an encoder and then runs K-means, which inherits two failure modes when the representation space is not well separated: majority clusters leak into minority ones (Uniform Effect) and minority clusters get absorbed into majority ones (Cluster Assimilation). THESAURUS claims to fix this by never relying on raw embedding distance alone. It gives every node a context by measuring it against a set of learnable semantic prototypes, and it turns the swap-prediction idea from contrastive learning into a graph-aware assignment target: the target assignments are computed by Fused Gromov-Wasserstein Optimal Transport between the data graph and a momentum-updated prototype graph, so the supervision signal carries both attribute and structure information. The paper reports that this produces higher cluster separability than prior art across nine datasets, with the largest gains on the minority clusters that K-means-based methods usually lose.","feed_headline":"Swapped transport couplings lift graph cluster separability","feed_subtitle":"Semantic prototypes plus a momentum prototype graph help minority clusters survive K-means readout.","key_machinery":"The engine of the method is the Fused Gromov-Wasserstein Optimal Transport coupling between the data graph $G = (A, Z)$ and the prototype graph $G_B = (B, S)$, where $A$ is adjacency, $Z$ is the encoder output, $S$ are learnable semantic prototypes, and $B = P^{\\top}P$ is the momentum-smoothed prototype graph. The optimal coupling, row-normalized, gives the assignment $Q$ used as the cross-view prediction target; the Gromov term matches structural patterns between $A$ and $B$ while the fused term matches node-to-prototype attributes, and a single parameter $\\alpha$ balances the two. The prototype graph and the prototype marginal distribution $\\nu$ are updated by momentum ($\\beta_1 = 0.99$, $\\beta_2 = 0.999$), which the paper says adapts the target to the data while preventing it from being too noisy.","core_discovery":"THESAURUS claims that the two classic K-means failure modes in deep graph clustering—the Uniform Effect and Cluster Assimilation—can be mitigated by making the clustering objective itself structure-aware instead of treating node embeddings as points in a metric space. The method learns soft assignments $Q$ by solving a Fused Gromov-Wasserstein problem between the augmented data graph $(A, Z)$ and a prototype graph $(B, S)$, where $S$ are semantic prototypes and $B = P^{\\top}P$ is a graph built from the model's own assignment logits, smoothed by momentum. These assignments are used as swapped cross-view prediction targets: view 1's representations must predict view 2's OT-derived assignment and vice versa. Because the target carries cluster information drawn from both node attributes and graph topology, the encoder is pushed toward a representation space with large inter-cluster and small intra-cluster distances, which the paper argues is precisely what prevents majority clusters from spilling into minority ones and minority clusters from being absorbed. On nine attributed graphs, the reported accuracy, NMI, ARI, and Macro-F1 all improve over the previous state of the art, Dink-Net, with Macro-F1 gains up to 11.65 points on Pubmed.","pith_inferences":["The self-referential target ($B = P^{\\top}P$) resembles an expectation-maximization loop; one testable extension is to study whether stability depends on the momentum constant $\\beta_1$, with a predicted phase transition where too-small $\\beta_1$ makes the prototype graph chase the current assignment noise.","The paper's argument that contextual prototypes distinguish 'synonymous' nodes suggests a direct comparison: on datasets with planted adjacent cross-class edges, THESAURUS should retain accuracy where distance-only baselines degrade; this is a testable hypothesis the paper does not run.","The per-node entropy of the optimal FGW coupling could be reused as an unsupervised confidence measure, flagging boundary or ambiguous nodes that the model itself is unsure about.","The momentum prototype marginal $\\nu$ could act as an online estimator of cluster-size imbalance, which might be exported as a lightweight diagnostic for other contrastive clustering methods."],"forward_implications":["If the central claim holds, graph clustering no longer needs a separate K-means finetune stage: the encoder is trained from the start toward cluster separability via the FGW-derived assignment targets, so the final K-means call is merely a readout.","The swapped-prediction target should generalize across datasets to imbalanced cluster sizes, since the momentum prototype marginal $\\nu$ tracks the true cluster size distribution instead of assuming uniform clusters.","Because the Gromov term compares adjacency structure to prototype structure, the method should keep working in graphs where node attributes are weak but topology is informative, and vice versa, as governed by $\\alpha$.","The same swapped FGW framework transfers to other unsupervised tasks that can be cast as assigning points to prototypes under a structure prior, such as community detection or multi-view clustering.","The method's reported robustness on minority clusters implies that adding more prototypes than the number of ground-truth classes does not hurt separability, which is a direct corollary of the momentum prototype graph adapting to the data."],"supporting_citations":[{"why":"Supplies the swapped cross-view assignment prediction idea (SwAV) that THESAURUS adapts to graphs.","marker":"Caron et al. 2020"},{"why":"Defines Fused Gromov-Wasserstein OT, the core transportation problem used to derive the assignments.","marker":"Titouan et al. 2019"},{"why":"Defines Gromov-Wasserstein distance between metric-measure spaces, the structural matching cost in the assignment.","marker":"Mémoli 2011"},{"why":"Provides the Sinkhorn algorithm that makes the entropy-regularized OT problem scalable.","marker":"Cuturi 2013"},{"why":"The GCN encoder whose output $Z$ is the attribute part of the fused OT problem.","marker":"Kipf and Welling 2017"},{"why":"The previous state-of-the-art baseline and the failure-mode analysis (Uniform Effect, Cluster Assimilation) that the paper diagnoses and claims to fix.","marker":"Liu et al. 2023a"},{"why":"Introduces the isolated-graph structural target that the paper replaces with the complete prototype graph $B$.","marker":"Xu et al. 2019"}],"fun_headline_variants":["Swapped couplings fix K-means flaws in graph clustering","Structure-aware clustering defeats uniform effect and assimilation","Semantic prototypes + swapped couplings = better graph clusters","Optimal transport swaps enhance graph cluster separability","Momentum prototype graph lifts graph clustering quality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The target that guides training is the model's own soft assignments: the prototype graph $B = P^{\\top}P$ and the marginal $\\nu$ are built from the encoder's current predictions, and the paper does not analyze what happens if this self-referential target drifts or collapses instead of providing independent structural information.","fun_headline_variants_meta":{"raw":{"variants":["Swapped couplings fix K-means flaws in graph clustering","Structure-aware clustering defeats uniform effect and assimilation","Semantic prototypes + swapped couplings = better graph clusters","Optimal transport swaps enhance graph cluster separability","Momentum prototype graph lifts graph clustering quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1439,"prompt_tokens":1094,"completion_tokens":345,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":710,"completion_tokens_details":{"reasoning_tokens":272}},"tokens_in":710,"tokens_out":345,"duration_ms":3957,"temperature":1.0,"reasoning_tokens":272,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:49:11.389056+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same training recipe but replace the momentum-updated prototype graph $B$ with a fixed, hand-specified structural target such as an isolated graph with $C$ components; if the reported accuracy and Macro-F1 gains over Dink-Net persist, then the prototype-graph mechanism is not what produces the cluster separability the paper attributes to it.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the swapped cross-view assignment prediction idea (SwAV) that THESAURUS adapts to graphs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines Fused Gromov-Wasserstein OT, the core transportation problem used to derive the assignments."}],"review_version":1}