{"id":"df6fe276-2b7a-45fd-973f-cc05cfe4b569","arxiv_id":"2507.20127","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"AMLP is a simple unsupervised framework that makes a single MLP adaptive to graph aggregation through a neighbor-consistency loss and a reconstructed graph, and it reports strong clustering and classification performance.","lead":"A new unsupervised graph learning method, AMLP, trains a single-layer MLP to be aware of the graph's aggregation rule, then uses the learned embeddings for clustering and classification. It reports strong results on homophilic and heterophilic graphs, but the theoretical proof and the experimental protocol have gaps that need attention.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The heterophilic gains depend on the thresholded reconstruction S in Eq. (3), but the paper gives no validation-based selection for epsilon and its only ablation shows S hurts Cora NMI; Theorem IV.2 is too weak and technically flawed to establish class-relevance.","rationale":"The reader identified the same load-bearing premise: Eq. (3) is where the method's heterophilic advantage is manufactured, and the paper's own ablation is the clearest evidence. I agree with that reading. The theoretical section does not rescue it, because it proves a smoothness/grouping bound on a smoothed surrogate rather than class separation, and the proof has concrete technical errors: Hoeffding's inequality is applied with the wrong dimension, no union bound over output coordinates appears, and the analysis applies to a randomly initialized W rather than the trained weights. I would not move the verdict: the empirical claim is conditionally plausible, and a clean reconstruction comparison with validation-based tuning could confirm or refute it. However, because the clustering protocol selects hyperparameters on the test metrics without error bars, the SOTA statement should be read as provisional pending that test.","tokens_in":17541,"tokens_out":8293,"duration_ms":94821,"concrete_test":"Controlled reconstruction comparison with validation-based hyperparameter selection. On the six heterophilic datasets plus Cora and Citeseer, hold out 20% of labels for validation, choose epsilon from {0.001, 0.05, 0.1}, k from {1, 2, 3, 5, 7, 8, 9, 10}, and lambda from {1, 1e-1, 1e-2, 1e-3} on the validation split, then compare AMLP using (a) S of Eq. (3), (b) original A, (c) a feature-cosine-only thresholded graph, and (d) an adjacency-row-cosine-only thresholded graph, with 10 random seeds and mean/std reporting. If Eq. (3) is not consistently better than A and the simpler alternatives under this protocol, the reconstruction assumption fails; if it is, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method's heterophilic gains are carried by the reconstructed graph S of Eq. (3), not by the MLP alone: in Table IV, using the original adjacency A instead of S drops AMLP's Texas ACC from 74.32 to 57.38 and Washington ACC from 74.35 to 62.61, while on Cora S actually hurts NMI (57.61 vs 59.43). The paper never shows that the specific thresholded product of feature cosine and adjacency-row cosine is class-relevant; the only support is Theorem IV.2, which (i) analyzes a sigmoid-smoothed approximation of S, (ii) proves a Lipschitz/grouping bound, not that kept edges connect same-class nodes, and (iii) is flawed: Assumption IV.4 applies Hoeffding with the input dimension d in place of the output dimension c, omits a union bound over output coordinates, and concerns a randomly initialized W, not the trained one. The selection of epsilon ('0.001 or 0.05') and the grid search without a validation split (Sec. V-C) mean reported margins may be test-set-selected. If S is not actually class-relevant, neither k-hop aggregation nor Lagg can recover the discarded signal.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AMLP, an unsupervised graph representation learning method. It first constructs a thresholded graph S from the product of normalized feature cosine similarity and normalized adjacency-row cosine similarity (Eq. (3)), then computes k-hop aggregated features eS^k XW, and trains a single-layer MLP with an aggregation-aware loss Lagg = ||eS^k XW - XW||_F^2 plus an inner-product reconstruction loss Lrec = ||Yhat Yhat^T - eA||_F^2, with output Y = eS^k XW + XW. Node embeddings are evaluated by K-means clustering and by a linear classifier. The manuscript claims state-of-the-art clustering results on six heterophilic and six homophilic datasets, competitive node classification against supervised and self-supervised baselines, scalability to Ogbn-arXiv, and a theoretical 'high-order grouping effect' together with a spectral interpretation of Lagg.","tokens_in":17880,"tokens_out":6184,"duration_ms":66214,"significance":"If the empirical claims hold, the paper makes a meaningful point: a simple, label-free MLP with a hand-constructed graph filter can match or beat sophisticated aggregation-based GNNs on several heterophilic benchmarks, and the method is lightweight and scalable. The paper includes useful ablations, an efficiency comparison, parameter sensitivity studies, and an attempt at theoretical justification. However, the theoretical analysis is not sound as written, and the experimental protocol leaves open the possibility that reported margins partly reflect test-set selection. The central claim is therefore not yet established, though the method itself is simple enough that a corrected analysis and a validated tuning protocol could make the contribution solid.","major_comments":[{"comment":"The proof of the high-order grouping effect is mathematically unsound. Assumption IV.4 asserts E[||r·W||_p] = 0 for a random matrix W with independent uniform entries; this cannot hold unless r·W = 0 almost surely, since a norm is nonnegative. The stated Hoeffding concentration P(||rW||_p >= t) <= 2 exp(-d t^2 / (2||r||_p^2)) also mixes dimensions: r is in R^c and W is in R^{d*c}, so the number of independent summands is c, not d, and no union bound over the c output coordinates is provided. In addition, the theorem's bound O(sqrt(ln(2/(δ(M+N))))) diverges as M+N -> 0, which is precisely the limit in which two nodes are most similar and in which the intended grouping property should give close embeddings. Finally, the theorem analyzes a sigmoid-smoothed approximation of S rather than the binary threshold in Eq. (3), so it does not establish that the specific reconstruction rule keeps class-relevant edges.","section":"§IV-C, Assumption IV.4 and Theorem IV.2"},{"comment":"The experimental protocol does not describe a validation-based selection of the hyperparameters. Section V-C states that k is searched in {1,2,3,5,7,8,9,10} except on Ogbn-arXiv, λ is searched in {1,1e-1,1e-2,1e-3}, the learning rate in {1e-2,1e-3,1e-4}, and the MLP width in {100,500}, but no validation split or repeated-selection procedure is described, and epsilon in Eq. (3) is reported as '0.001 or 0.05' without selection details. The clustering results in Tables II and III are also reported without standard deviations or error bars. Under this protocol, the reported state-of-the-art margins may be partly the result of selecting the best configuration on the test data, and the significance of the gains cannot be assessed.","section":"§V-C, Tables II-IV"},{"comment":"The ablation shows that the reconstructed graph S is load-bearing for the heterophilic results: replacing S with A drops Texas ACC from 74.32 to 57.38 and Washington ACC from 74.35 to 62.61, while on Cora the original A gives higher NMI than S (59.43 vs 57.61). This means the method's heterophilic advantage is carried by the particular thresholded product of feature and adjacency-row cosines in Eq. (3), not by the adaptive MLP per se, and the reconstruction is not uniformly beneficial. No evidence is given that this specific thresholding rule is class-relevant, and Theorem IV.2 does not fill that gap because it bounds a smoothed surrogate in terms of essentially the same cosine similarities that define S. If S discards the class-discriminative signal, neither the k-hop aggregation nor Lagg can recover it.","section":"§V-F, Table IV and Eq. (3)"},{"comment":"The derivation of the claimed high-pass filtering effect is not justified. The transition from min ||(eS^k-I)XW||_F^2 to a spectral sum with terms [(eS^k-I)^2-I]u_j ||(XW)_u - (XW)_j|| is presented without an algebraic derivation and is not a standard spectral identity; as written, Eq. (9) does not establish the claimed relationship. Without a correct proof, the statement that Lagg 'exhibits a high-pass filtering effect' is unsupported, and this weakens the theoretical interpretation in Section IV-C.","section":"§IV-C, Proposition IV.6 and Eq. (9)"}],"minor_comments":[{"comment":"The notation in the preliminaries is inconsistent: eA is defined as the degree-normalized adjacency with self-loops, but the graph Laplacian is written as L = I - A rather than using the normalized variant, and Eq. (1) applies the MLP only to raw features rather than to transformed neighbor features, which may confuse the message-passing formalization.","section":"§III-A"},{"comment":"The captions of Figures 1 and 2 are too terse; they do not describe the axes, the quantities plotted, or the meaning of the subfigures, making the empirical study difficult to verify from the figures alone.","section":"§III-B, Figures 1 and 2"},{"comment":"Equation (3) defines S_ij for all i,j, but the text later says eS has no self-loop; it should be stated explicitly whether diagonal entries are excluded before normalization, since the formula as written would include i = j with a value depending on the cosine of a row with itself.","section":"§IV-A, Eq. (3)"},{"comment":"The caption says the best performance is marked in bold, but in the rendered table the Cora NMI value for 'AMLP w A' (59.43) is higher than that of AMLP (57.61); either the bold marking is missing or the claim that AMLP is best on Cora needs qualification.","section":"§V-F, Table IV"},{"comment":"The claim that this is 'the first work to shift the focus from designing the aggregation mechanism to making learned representations aggregation-adaptive' is too strong and is not supported by a comparison with existing learnable or adaptive aggregation approaches; the claim should be softened or accompanied by a more thorough discussion.","section":"§I and §II"}],"recommendation":"major_revision","confidential_remarks":"The manuscript presents a simple, potentially useful unsupervised method and a large set of experiments, but the theoretical section contains a serious mathematical error and the experimental protocol lacks validation-based hyperparameter selection and error bars for the clustering results. I recommend major revision: the authors should either provide a corrected theorem and proof or remove the flawed theoretical claims, and they should clarify or reform the tuning protocol. I do not think the issues are irreparable within the manuscript's scope, but in the current form the central claims of SOTA performance and theoretical support are not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read it: the empirical story is stronger than the theory, and the method's heterophilic edge is carried by a hand-built graph reconstruction that the paper never properly validates.\n\nWhat's actually new is modest. The aggregation-aware loss is close to the neighbor-consistency objectives in the works they cite, and the graph reconstruction follows the same group's DGCN. The specific combination — thresholded product of feature and adjacency cosines, single-layer MLP, inner-product decoder — is not in the literature verbatim, and the empirical study is broad: 12 datasets, 35 baselines, clustering and classification, an ablation, and an efficiency table. The classification results come with error bars and look competitive against supervised GNNs on heterophilic benchmarks; that's the most credible part of the paper. The Dirichlet-energy experiment motivating Lagg is also a nice touch.\n\nSoft spots, in proportion. The theory section is not sound. Theorem IV.2 analyzes a sigmoid-smoothed S, not the thresholded S; it proves a Lipschitz-type bound, not class relevance; and the Hoeffding application has the wrong dimension in the exponent, no union bound, and assumes a random W rather than the trained one. The proof also conflates S^k with S when factoring (S_i - S_j). None of this supports the claim that the specific thresholding rule produces a class-relevant graph. The clustering evaluation has no error bars, and k, lambda, and epsilon are grid-searched per dataset without a validation split, so the SOTA margins may be partly test-set selection. The ablation shows the reconstruction hurts Cora NMI, which the authors admit, but that undercuts the universality claim. These are real problems, but they don't falsify the empirical observation; the classification experiments are more solid.\n\nCitation pattern is fine — they build on their own prior work, but that's legitimate here. The paper is for people working on unsupervised graph learning or heterophilic benchmarks. It deserves peer review because the empirical claim is important enough, but the theory should be fixed or removed, and the hyperparameter selection needs a validation protocol.","headline":"The empirical results are worth a look, but the load-bearing graph reconstruction and the flawed theory don't support the SOTA claims as stated.","tokens_in":18326,"tokens_out":3673,"would_cite":false,"duration_ms":38586,"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":"One MLP adapts to any graph's aggregation, no labels needed","keywords":["graph clustering","heterophily","unsupervised graph representation learning","aggregation-aware MLP","graph reconstruction","message passing","node classification"],"falsifier":"Compute clustering performance with the original adjacency $A$ in place of the reconstructed $S$ across all datasets; the central claim that $S$ is more class-relevant than $A$ would be falsified if the $A$ variant wins on most heterophilic graphs, and the paper's own ablation already shows reconstruction lowers NMI on Cora from 59.43 to 57.61.","tokens_in":17349,"feed_emoji":"🕸️","tokens_out":6469,"duration_ms":61254,"temperature":0.7,"pith_summary":"The paper proposes AMLP, an unsupervised graph representation method in which a single-layer MLP is made 'aggregation-aware' rather than relying on a fixed Mean/Max/Sum aggregator. It first rebuilds the adjacency graph with a thresholded product of feature and topology similarities, then trains the MLP with a loss that pulls the k-hop aggregated embedding back toward the raw embedding. The authors claim this unifies homophilic and heterophilic graph learning without labels and report state-of-the-art clustering on all six heterophilic and most homophilic datasets tested, plus competitive node classification against supervised aggregation GNNs. If the claim holds, a simple parameter-light MLP can replace supervised aggregation design in many graph tasks.","feed_headline":"One MLP adapts to any graph's aggregation, no labels needed","feed_subtitle":"An unsupervised aggregation-aware loss lets a single MLP out-cluster supervised GNNs on heterophilic graphs.","key_machinery":"The load-bearing object is the reconstructed graph $S$ of Eq. (3), whose binary entries keep edge $(i,j)$ only when the squared product of the feature cosine similarity and the degree-normalized adjacency-row cosine similarity exceeds $\\epsilon$; together with the aggregation-aware loss $L_{\\mathrm{agg}}$, this is what lets the single MLP adapt to the graph. $S$ is supposed to inject a high-order grouping effect by correlating feature and topology spaces, while $L_{\\mathrm{agg}}$ penalizes mismatch between $k$-hop aggregated and raw embeddings, and the optional residual $XW$ compensates for the absence of self-loops. The theoretical development proves the grouping bound for a sigmoid-smoothed version of $S$ and shows that the loss has a high-pass spectral effect.","core_discovery":"AMLP claims that a single linear transform can adapt to graph structure if its training objective is aware of aggregation. Concretely, with reconstructed graph $S$ and normalized adjacency $\\tilde S$, the method forms $Y = \\tilde S^k XW + XW$ and trains $W$ by minimizing $L_{\\mathrm{agg}} + \\lambda L_{\\mathrm{rec}}$, where $L_{\\mathrm{agg}}=\\|\\tilde S^k XW - XW\\|_F^2$ makes the MLP responsive to the graph filter and $L_{\\mathrm{rec}}$ reconstructs the original adjacency through inner products. The paper argues that $L_{\\mathrm{agg}}$ acts as a high-pass filter, balancing the low-pass smoothing of aggregation, and that the reconstructed $S$ induces a high-order grouping effect that keeps similar nodes close while separating dissimilar ones. Extensive experiments on 12 datasets are reported as evidence that this unsupervised scheme reaches state-of-the-art clustering and beats label-hungry aggregation GNNs on several heterophilic classification benchmarks.","pith_inferences":["Editorial extension: because graph reconstruction is done as a preprocessing step and $L_{\\mathrm{agg}}$ is additive, one could apply the same reconstruction and loss to supervised GNNs; the likely testable benefit is improved robustness on heterophilic graphs with few labels.","Editorial extension: the product form in Eq. (3) is hand-chosen and the paper only tries $\\epsilon \\in \\{0.001, 0.05\\}$. A natural extension is to learn $S$ by backpropagation or to tune $\\epsilon$ per dataset; the ablation's Cora result suggests that a hard threshold can discard useful structure, so an adaptive choice between $A$ and $S$ may improve worst-case performance.","Editorial extension: the theoretical grouping guarantee is for a sigmoid-smoothed $S$, not the implemented hard threshold; checking empirically whether the hard-thresholded $S$ satisfies the same bound on real graphs would connect the proof to the deployed algorithm."],"forward_implications":["Clustering on graphs can be done without labels and without choosing an aggregator: AMLP reports the best ACC and NMI among the compared unsupervised methods on all six heterophilic datasets and on most homophilic datasets, including the large Ogbn-arXiv graph.","The aggregation-aware loss is a drop-in training term: the authors state it can be added to any existing message-passing GNN, so homophily-versus-heterophily tuning could become a training-time adjustment rather than a network-design choice.","A single-layer MLP with $k$-hop filtering and a residual connection is enough to beat supervised aggregation baselines on three of the four classification datasets reported.","Balancing the low-pass aggregation filter with the high-pass $L_{\\mathrm{agg}}$ prevents both over-smoothing and over-sharpening, which the paper supports by showing that Dirichlet energy moves in the expected direction on homophilic versus heterophilic graphs."],"supporting_citations":[{"why":"Defines the weighted-sum graph convolution and supplies the Cora, Citeseer, and Pubmed citation benchmarks used throughout the experiments.","marker":"[1]"},{"why":"RAW-GNN is a heterophily-specific aggregation baseline that AMLP compares against in the classification experiments.","marker":"[6]"},{"why":"GloGNN++ provides the precedent AMLP cites for pushing aggregated embeddings close to raw input features and supplies the classification splits used for aggregation-based baselines.","marker":"[10]"},{"why":"GPRGNN is a supervised adaptive-aggregation baseline that AMLP must beat on heterophilic classification.","marker":"[11]"},{"why":"The self-expressive adjacency-based reconstruction idea from which the aggregation-aware loss is drawn.","marker":"[21]"},{"why":"DGCN supplies the graph-agnostic clustering baseline, the experimental configuration, and the view that graph reconstruction is itself a fixed aggregation.","marker":"[22]"},{"why":"Motivates the local-consistency objective of matching neighbor-aggregated features to input features.","marker":"[23]"},{"why":"Supplies the difficult Roman-Empire heterophilic benchmark used in the clustering experiments.","marker":"[28]"}],"fun_headline_variants":["Unsupervised MLP learns to adapt to any graph aggregation","Aggregation-aware MLP beats supervised GNNs with zero labels","Single-layer MLP adapts to heterophily without supervision","Label-free MLP out-clusters GNNs on heterophilic graphs","Unsupervised AMLP: MLP that adjusts to graph filters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method rests on the hand-built rule that decides which edges survive graph reconstruction; if that rule removes or distorts the exact signal that separates classes, the aggregation-aware loss cannot recover it.","fun_headline_variants_meta":{"raw":{"variants":["Unsupervised MLP learns to adapt to any graph aggregation","Aggregation-aware MLP beats supervised GNNs with zero labels","Single-layer MLP adapts to heterophily without supervision","Label-free MLP out-clusters GNNs on heterophilic graphs","Unsupervised AMLP: MLP that adjusts to graph filters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000208,"raw_usage":{"total_tokens":1409,"prompt_tokens":956,"completion_tokens":453,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":363}},"tokens_in":572,"tokens_out":453,"duration_ms":4386,"temperature":1.0,"reasoning_tokens":363,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:49:43.033038+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute clustering performance with the original adjacency $A$ in place of the reconstructed $S$ across all datasets; the central claim that $S$ is more class-relevant than $A$ would be falsified if the $A$ variant wins on most heterophilic graphs, and the paper's own ablation already shows reconstruction lowers NMI on Cora from 59.43 to 57.61.","supporting_citations":[{"cited_title":"Semi-supervised classification with graph convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Defines the weighted-sum graph convolution and supplies the Cora, Citeseer, and Pubmed citation benchmarks used throughout the experiments."},{"cited_title":"Raw-gnn: Random walk aggregation based graph neural network,","cited_arxiv_id":null,"evidence_quote":"RAW-GNN is a heterophily-specific aggregation baseline that AMLP compares against in the classification experiments."},{"cited_title":"Finding global homophily in graph neural networks when meeting heterophily,","cited_arxiv_id":null,"evidence_quote":"GloGNN++ provides the precedent AMLP cites for pushing aggregated embeddings close to raw input features and supplies the classification splits used for aggregation-based baselines."},{"cited_title":"Adaptive universal generalized pagerank graph neural network,","cited_arxiv_id":null,"evidence_quote":"GPRGNN is a supervised adaptive-aggregation baseline that AMLP must beat on heterophilic classification."},{"cited_title":"Multi-view contrastive graph clustering,","cited_arxiv_id":null,"evidence_quote":"The self-expressive adjacency-based reconstruction idea from which the aggregation-aware loss is drawn."},{"cited_title":"Beyond homophily: Reconstructing structure for graph-agnostic clustering,","cited_arxiv_id":null,"evidence_quote":"DGCN supplies the graph-agnostic clustering baseline, the experimental configuration, and the view that graph reconstruction is itself a fixed aggregation."},{"cited_title":"Graph data condensation via self- expressive graph structure reconstruction,","cited_arxiv_id":null,"evidence_quote":"Motivates the local-consistency objective of matching neighbor-aggregated features to input features."},{"cited_title":"A critical look at the evaluation of gnns under het- erophily: are we really making progress?","cited_arxiv_id":null,"evidence_quote":"Supplies the difficult Roman-Empire heterophilic benchmark used in the clustering experiments."}],"review_version":1}