{"id":"bc27fc1e-2fe6-42c7-a910-03a02c5856e2","arxiv_id":"1908.00673","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A weight-shared, max-pooled hybrid of low- and high-order graph convolutions reports top accuracy on several text and citation benchmarks with fewer parameters than comparable models.","lead":"This paper presents a graph neural network that combines first-, second-, and third-order neighbor information with a max-pooling layer and shared weights, reporting accuracy gains on text and citation classification benchmarks. The approach matters because it suggests high-order graph context can be captured without multiplying trainable parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central claim of 'highest classification accuracy' is contradicted by its own reported results on MR and Pubmed; without a controlled re-evaluation, the headline assertion is not supported.","rationale":"I read the paper as proposing a concrete architectural variant: share one weight matrix across graph powers in each layer and fuse the order channels by elementwise max-pooling. That is a reasonable idea, and the parameter-count reduction is plausible because the weights are shared. However, the central claim in the abstract is empirical and unqualified: the model achieves the highest classification accuracy with a small set of parameters. The most load-bearing support for that claim is the experimental comparison, and that comparison fails internally. The paper's own Table 3 and Table 7 show datasets where the proposed models are clearly worse than the best baselines, specifically MR and Pubmed. The paper acknowledges these cases in prose but still concludes it outperforms the benchmark. This contradiction is stronger than the modeling assumption singled out by the reader: even if one believes that adding high-order information should help, the reported numbers do not establish the central claim. Additionally, the baselines are copied from prior papers instead of being retrained under the same protocol, so the remaining claimed improvements on Cora and Citeseer could be due to hyperparameter tuning, implementation differences, or different random seeds. No code is released, so independent replication is not currently possible. For these reasons, the rejection verdict is appropriate and should be unchanged. The suggested concrete test, a controlled reimplementation with shared hyperparameter search, would settle whether the claimed accuracy advantage survives under fair comparison.","tokens_in":612,"tokens_out":4310,"duration_ms":56944,"concrete_test":"Reimplement HLHG-2 and HLHG-3, and also run Text GCN and MixHop (learned), on the standard splits used by the paper: for text networks, the Yao et al. splits; for citation networks, the 20-labels-per-class Planetoid splits. Use a shared hyperparameter search budget and at least 100 random seeds. The decisive check is whether HLHG-3 exceeds Text GCN on MR and exceeds MixHop (learned) on Pubmed. If either comparison fails, the paper's 'highest accuracy' claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Section 5 claim that HLHG achieves the highest classification accuracy across the tested benchmarks. However, the paper's own tables contradict this. In Table 3, on MR, HLHG-2 achieves 75.95 and HLHG-3 achieves 76.49, both below Text GCN's 76.74. In Table 7, on Pubmed, HLHG-2 achieves 79.1 and HLHG-3 achieves 79.3, below MixHop (learned)'s 80.8 and also below HO-4's 80.1. These are not marginal differences, and they are on the same datasets the paper claims to cover. Moreover, the reported wins on Cora and Citeseer (82.7 vs. 81.9 and 71.5 vs. 71.4) come from copying baseline numbers from prior work without retraining those baselines under the same evaluation protocol, and no code is provided. Consequently, the central claim is not merely weakened by a questionable high-order-fusion assumption from Section 3.1; the empirical evidence presented does not support the unqualified statement that the proposed model achieves highest accuracy. A controlled comparison is required before the claim can be assessed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid low-order and higher-order graph convolutional network (HLHG) that combines first-, second-, and optionally third-order neighborhood information through an element-wise max fusion pooling layer and shares trainable weights across the different order convolutions. The model is evaluated on five text classification datasets under supervised learning and three citation network datasets under semi-supervised learning. The authors claim that HLHG achieves the highest classification accuracy with a smaller number of trainable parameters compared to prior GCN, HO, and MixHop models, and they compare computational complexity and parameter counts theoretically. The paper also reports that the code and data will be made available.","tokens_in":10295,"tokens_out":3342,"duration_ms":33081,"significance":"If the empirical claims were fully supported, the paper would make a modest but useful contribution: the proposed max-pooling fusion over neighborhood orders with weight sharing is a simple, parameter-efficient alternative to concatenation-based higher-order GCNs such as MixHop, and the complexity analysis in Section 3.3 is a reasonable first step. The authors also provide standard deviations over 100 runs for their own models, which is commendable. However, the central accuracy claim is not supported by the paper's own reported results on all datasets, and the baseline comparisons are not conducted under a controlled protocol. As presented, the evidence does not establish state-of-the-art performance, and the contribution reduces to an incremental heuristic with unresolved validation issues.","major_comments":[{"comment":"The abstract and conclusion state that the proposed model achieves 'highest classification accuracy', but the reported results contradict this claim on two of the eight benchmarks. In Table 3, on MR, HLHG-2 achieves 75.95 and HLHG-3 achieves 76.49, both below Text GCN's 76.74. In Table 7, on Pubmed, HLHG-2 achieves 79.1 and HLHG-3 achieves 79.3, below MixHop (learned) at 80.8 and HO-4 at 80.1. These are not marginal gaps, and the paper itself acknowledges the MR discrepancy only in passing ('the proposed models yield worse accuracies on the dataset MR'). The unqualified state-of-the-art claim is therefore not justified by the presented evidence and must be re-scoped or supported by a controlled re-evaluation.","section":"Abstract and Section 5 vs. Tables 3 and 7"},{"comment":"The citation-network baselines are copied from prior work (Abu-El-Haija et al. and Yang et al.) rather than retrained under the same protocol, and no error bars or standard deviations are reported for those baselines. The text states that 'The benchmark test result copy from [2] and [28]' without specifying whether the train/validation/test splits, number of runs, and hyperparameter selection are identical across methods. Reported gains on Cora (82.7 vs. 81.9) and Citeseer (71.5 vs. 71.4) are within typical run-to-run variability of these baselines, and without a controlled comparison it is impossible to assess whether HLHG actually outperforms them. This weakens the central claim of superiority in semi-supervised node classification.","section":"Section 4.2 and Table 7"},{"comment":"Hyperparameters, including the maximum neighborhood order p (2 vs. 3), hidden units, dropout rate, learning rate, epochs, and early-stopping patience, are tuned separately for each dataset without a described validation procedure. Selecting p based on test-set accuracy, as implied by the sentence 'we validate the p = 2 and p = 3 in Eq. (3) for our HLHG models,' introduces selection bias. A nested validation or a fixed protocol applied uniformly across all datasets and baselines is needed to make the reported accuracy differences credible.","section":"Sections 4.1 and 4.2, Tables 2 and 6"}],"minor_comments":[{"comment":"Equation (3) contains unbalanced parentheses and applies the activation function σ only to the first argument of Pmax, which obscures the model definition. Equations (4) and (5) are clearer, but the malformed equation should be corrected.","section":"Section 3.2, Eq. (3)"},{"comment":"There is a typo: 'and andh1' should be 'and h1'. In addition, the phrase 'unsupervised learning' in Section 3.2 appears to be an error; the experiments use semi-supervised learning.","section":"Section 3.1"},{"comment":"The entries for MixHop (learned) report O(60×1×r) parameters while MixHop reports O(20×3×r); the meaning of the constants (hidden units per order vs. total hidden units) should be clarified to make the parameter-count comparison transparent.","section":"Table 8"},{"comment":"The sentence 'our HLHG-3 model has an advantage in computational complexity in dataset MR' compares O(64×3×m×r) with Text GCN's O(200×1×m×r); the difference (192 vs. 200) is minor and the phrasing overstates the advantage. Also, the promised GitHub link for code and data is not provided in the manuscript.","section":"Section 4.1"}],"recommendation":"reject","confidential_remarks":"The paper's headline contribution is the state-of-the-art accuracy claim, which is internally contradicted by the reported numbers on MR and Pubmed and by the lack of controlled baseline comparisons. The reproducibility of the experiments is further limited by the absence of code and a clear validation protocol. These issues are load-bearing and would require substantial new experiments or a major re-scoping of the claims to be addressed; I do not see a path to acceptance within the current manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a straightforward extension of GCN/MixHop: share one weight matrix across powers of the normalized adjacency matrix and fuse the order-specific representations with element-wise max. That is a sensible parameter-reduction trick, and it does seem to give a point or two on several text benchmarks. But the paper overclaims: its own results on MR and Pubmed are below the best baselines, so the abstract's 'highest accuracy' is simply false as written.\n\nWhat is new: the specific combination of weight-tied high-order propagation with max-pooling fusion. It is incremental, but not a carbon copy of MixHop, which concatenates. The complexity comparison in Section 3.3 is roughly correct: with shared weights the parameter count stays independent of the number of orders. On R52, OH, 20NG and R8 the reported gains over Text GCN are 0.3–1.0 point, which is plausible for a method that adds neighborhood context.\n\nThe soft spots are in the evidence, not the math. First, the headline claim is contradicted in the paper's own tables: MR (76.49 vs Text GCN 76.74) and Pubmed (79.3 vs MixHop-learned 80.8, HO-4 80.1). That is not a minor quibble; it changes the conclusion from 'state-of-the-art' to 'competitive on some datasets.' Second, the baselines are copied from earlier papers rather than rerun under the same split and hyperparameter protocol, so the +0.8 on Cora and +0.1 on Citeseer could easily be evaluation noise. Third, hyperparameters and the pooling order p are chosen per dataset with no validation split described; the reader is told p=4 'was tested' but no numbers are shown. Fourth, the promised code and data are not publicly available, so the numbers can't be checked. The typesetting slips in Eqs. (3) and (6) make those equations hard to parse, though the intent is clear from the surrounding text.\n\nNone of this is fatal to the underlying idea. The weight-sharing/max-pooling design is easy to implement and might be a useful baseline for higher-order GCN papers. But in its current form the paper does not support its central claim, and the evaluation is not rigorous enough to be cited as a reliable result.\n\nIf I were handling it, I would send it to reviewers only with the expectation of a major revision: rerun baselines under the same protocol, add error bars everywhere, fix the claim, and release code. If this is for your own reading: it's worth skimming for the architecture, but I wouldn't build on the reported numbers.","headline":"A simple weight-sharing, max-pooling higher-order GCN with modest gains on some benchmarks, but the paper's own tables contradict its state-of-the-art claim and the evaluation protocol is too loose to trust the numbers.","tokens_in":10776,"tokens_out":2274,"would_cite":false,"duration_ms":21706,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that element-wise max fusion of low- and high-order neighborhoods, with weights shared across orders, lets graph convolutional networks beat prior higher-order models while using fewer trainable parameters.","keywords":["graph convolutional network","higher-order neighborhood","weight sharing","fusion pooling","text classification","semi-supervised node classification","computational complexity"],"falsifier":"An ablation that swaps $P_{\\max}$ for element-wise sum on Cora, Citeseer, and the text datasets would settle whether the max-pooling mechanism is load-bearing; if accuracy does not change, the claimed fusion mechanism is not the cause of the improvement. The paper's own MR and Pubmed results already provide a partial negative: on those datasets the hybrid high-order models fall short of the best baselines reported in the same tables.","tokens_in":9845,"feed_emoji":"🔗","tokens_out":7621,"duration_ms":72986,"temperature":0.7,"pith_summary":"The paper is trying to establish that a graph convolutional network can capture the benefits of higher-order neighborhood information, meaning nodes several hops away, without the usual parameter explosion, by fusing first- through third-order propagated features with element-wise maximum pooling and reusing a single weight matrix across all orders. If the claim is right, the practical payoff is that high-order graph convolution becomes affordable on large graphs: the trainable parameter count stays fixed as the neighborhood order grows, instead of multiplying by the order as in concatenation-based higher-order GCNs. The authors test this on five text networks under supervised learning and three citation networks under semi-supervised learning, reporting the best accuracy on most of those benchmarks with fewer trainable weights than the Text GCN baseline. Their own tables show MR and Pubmed are exceptions where baselines do better, so the claim is not that the fusion wins universally.","feed_headline":"Max-fused graph neighborhoods beat GCNs with fewer weights","feed_subtitle":"Element-wise max fusion of low and high order neighborhoods lifts text and citation classification accuracy.","key_machinery":"The load-bearing object is the fusion pooling operator $P_{\\max}$, an element-wise maximum over the first $p$ neighborhood propagations; it keeps the propagated tensor shape unchanged, so adding another order does not add a new weight matrix. Weight sharing is the companion mechanism: one trainable $W_l$ is applied to every order in a layer, so the parameter count does not scale with $p$. The third piece is right-to-left iterative multiplication, $\\hat{A}^p X W_1 = \\hat{A}(\\cdots(\\hat{A}X))W_1$, which avoids explicitly forming dense high powers of the adjacency matrix. Together these pieces carry the argument that high-order neighborhood information can be fused at the parameter cost of a first-order GCN.","core_discovery":"The central discovery, stated on the paper's own terms, is that element-wise max over the first $p$ powers of the normalized adjacency matrix, applied to feature-transformed representations, is a cheap and effective nonlinear way to mix low- and high-order neighborhood information. The HLHG-2 and HLHG-3 models stack two graph convolutional layers and one fusion pooling layer, with propagation $Y = \\mathrm{softmax}[P_{\\max}(\\hat{A}\\sigma(H)W_2,\\ldots,\\hat{A}^p \\sigma(H)W_2)]$ where $H = P_{\\max}(\\hat{A}XW_1,\\ldots,\\hat{A}^p XW_1)$ and $P_{\\max}$ is element-wise maximum. Because the same $W_1$ and $W_2$ are shared across orders, the weight count per layer is $O(r_l r_{l-1})$, independent of the order $p$; the complexity stays $O(r_l p m r_{l-1})$, with $m$ nonzeros in the sparse adjacency matrix. The paper reports that these models reach top accuracy on R52, OH, 20NG, and R8 among text classifiers and on Cora and Citeseer among citation baselines while using fewer parameters than Text GCN.","pith_inferences":["A sharp ablation test suggested by, but not run in, the paper would replace $P_{\\max}$ with element-wise sum or concatenation on the same datasets; if accuracy does not drop, the max operation itself is not the source of the reported gains.","The paper compares complexity and parameters only at the first layer under the assumption $r_0 \\gg r_1$; extending the shared-weight construction to deeper or heterogeneous graphs would reveal whether the savings persist beyond the regime tested.","Reading across the results, the fusion seems to help least on datasets (MR, Pubmed) where strong first-order baselines already saturate; a natural hypothesis, not stated by the authors, is that higher-order fusion matters most when local neighborhoods are noisy or weakly informative.","The authors say they validated $p=4$ and found no significant improvement, which suggests the effective receptive field of these two-layer models is already covered by orders 1 through 3; one could test this by varying the number of layers jointly with the order."],"forward_implications":["If the paper is right, a two-layer shared-weight max-fusion GCN is sufficient to outperform or match prior higher-order GCN architectures on the citation benchmarks, so the extra parameters spent on per-order weights in concatenation models are not necessary for accuracy.","The parameter count of a graph convolutional layer no longer needs to grow with the number of neighborhood hops, making higher-order convolution more practical for large sparse graphs.","On the text benchmarks, the reported accuracy on four of five datasets implies that the hybrid fusion captures enough high-order text structure to beat both specialized text classifiers and the Text GCN baseline.","The complexity bound $O(r_l p m r_{l-1})$ per layer implies that the price of adding neighbors is a constant factor in the sparse-matrix multiplications, not a blow-up in trainable weights."],"supporting_citations":[{"why":"Supplies the classical single-order GCN propagation that HLHG extends to multiple neighborhood orders.","marker":"[14]"},{"why":"Defines the higher-order graph convolutional layer whose parameter growth HLHG is designed to avoid.","marker":"[1]"},{"why":"Presents MixHop, the concatenation-based high-order baseline that HLHG compares against on citation networks.","marker":"[2]"},{"why":"Provides the text-graph construction, benchmark datasets, and Text GCN baseline for the supervised text classification experiments.","marker":"[29]"},{"why":"Provides the semi-supervised evaluation methodology and Planetoid baseline used on the citation networks.","marker":"[28]"},{"why":"Supplies the Cora, Citeseer, and Pubmed citation network datasets used in the semi-supervised experiments.","marker":"[22]"}],"fun_headline_variants":["Weight-shared graph conv fuses low and high orders","Element-wise max fusion cuts GCN parameters, keeps accuracy","HLHG nets top text and citation accuracy with tiny weight sets","Max-pooled graph neighborhoods beat GCNs on fewer weights","Shared-weight graph conv with max fusion beats GCNs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's accuracy claims rest on the assumption, stated plainly in Section 3.1, that adding and max-fusing higher-order neighborhood features will improve classification accuracy, but no proof is given that element-wise max is the right fusion; the reported results on MR and Pubmed show the assumption does not always hold.","fun_headline_variants_meta":{"raw":{"variants":["Weight-shared graph conv fuses low and high orders","Element-wise max fusion cuts GCN parameters, keeps accuracy","HLHG nets top text and citation accuracy with tiny weight sets","Max-pooled graph neighborhoods beat GCNs on fewer weights","Shared-weight graph conv with max fusion beats GCNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1356,"prompt_tokens":949,"completion_tokens":407,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":323}},"tokens_in":565,"tokens_out":407,"duration_ms":4556,"temperature":1.0,"reasoning_tokens":323,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:38:41.928228+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An ablation that swaps $P_{\\max}$ for element-wise sum on Cora, Citeseer, and the text datasets would settle whether the max-pooling mechanism is load-bearing; if accuracy does not change, the claimed fusion mechanism is not the cause of the improvement. The paper's own MR and Pubmed results already provide a partial negative: on those datasets the hybrid high-order models fall short of the best baselines reported in the same tables.","supporting_citations":[{"cited_title":"A higher-order graph convolutional layer","cited_arxiv_id":null,"evidence_quote":"Defines the higher-order graph convolutional layer whose parameter growth HLHG is designed to avoid."},{"cited_title":"MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing","cited_arxiv_id":"1905.00067","evidence_quote":"Presents MixHop, the concatenation-based high-order baseline that HLHG compares against on citation networks."},{"cited_title":"Cohen, and Ruslan R","cited_arxiv_id":null,"evidence_quote":"Provides the semi-supervised evaluation methodology and Planetoid baseline used on the citation networks."},{"cited_title":"Collective classification in network data","cited_arxiv_id":null,"evidence_quote":"Supplies the Cora, Citeseer, and Pubmed citation network datasets used in the semi-supervised experiments."}],"review_version":1}