{"id":"6f38ae66-6883-4e65-ae37-8cdb809d45ce","arxiv_id":"2505.02020","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"GCNIII, a wide linear plus deep graph convolution hybrid, achieves slightly higher node classification accuracy than GCNII on several benchmarks by adding a linear branch and tuning hyperparameters per dataset.","lead":"This paper proposes GCNIII, a graph neural network that combines a simple linear classifier with a deep graph convolution branch, borrowing the Wide & Deep idea from recommender systems. It reports small accuracy gains over prior state-of-the-art models on node classification benchmarks and analyzes why very deep graph networks sometimes show lower training error than validation error, a pattern it calls over-generalization.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Full-supervised SOTA claim relies on a weakened GCNII baseline: GCNII is retrained, underperforms Chen et al. (2020) on the last four datasets, and those lower numbers are used in Table 4; large heterophily gains may vanish against the original GCNII. Appendix B per-dataset configs compound this.","rationale":"I read the paper as making an empirical claim: GCNIII, a Wide & Deep combination of a linear branch and a deep GCN branch with three optional techniques, achieves state-of-the-art node classification. The most load-bearing support for that claim is the experimental comparison, not the auxiliary Theorem 4.1. The reader's weakest_assumption focused on per-dataset configuration choices being potential test-set fitting; I agree that is a serious issue, but I find an even more direct problem: for the full-supervised setting, baselines are reused from Chen et al. (2020) while GCNII is retrained in a framework that the authors say produces much lower results on four of the seven datasets. Using those lower GCNII numbers as the reference makes the large claimed improvements on heterophily datasets difficult to interpret. This is not an ad hominem concern; the authors transparently report the degradation, but then proceed to compare against the degraded numbers rather than against the reported GCNII results they themselves treat as unreliable to reproduce. I also examined Appendix A and agree with the reader's secondary concern: the proof of Theorem 4.1 assumes shared parameters, aligned identity-mapping weights, non-negative activations, and bounded products, assumptions not stated in the theorem. However, the theorem supports the 'ultra-deep is not necessary' narrative, not the headline SOTA claim. Because the central issue is empirical and fixable by a fair baseline comparison, I would keep the reader's CONDITIONAL verdict rather than moving to ACCEPT or REJECT. My agreement is partial because the reader's stated weakest assumption emphasizes per-dataset test-set fitting, while my primary concern is the asymmetric treatment of the GCNII baseline; both concerns point to the same section of the paper and jointly undermine the full-supervised SOTA claim.","tokens_in":22119,"tokens_out":9660,"duration_ms":94146,"concrete_test":"Obtain the full-supervised GCNII results originally reported by Chen et al. (2020), or rerun GCNII with the original PyG code and original hyperparameters instead of the DGL reimplementation, and recompute Table 4 with those numbers as the baseline while keeping GCNIII's reported per-dataset configurations. If GCNIII does not beat the original GCNII on all seven datasets—especially on Chameleon, Cornell, Texas, and Wisconsin—the state-of-the-art claim is not established. As a secondary control, fix one GCNIII configuration (e.g., techniques [1,1,1], 8 layers, gamma=0.02) and compare against a fixed GCNII configuration under validation-based model selection to test whether the Wide & Deep component itself, rather than per-dataset tuning, drives the gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim—'achieves state-of-the-art results on various node classification tasks'—is not supported by the full-supervised comparison in §6.3/Table 4. For GCN, GAT, Geom-GCN variants, and APPNP the authors reuse metrics from Chen et al. (2020), but for GCNII they retrain within their own DGL framework and explicitly state that on Chameleon, Cornell, Texas, and Wisconsin 'the results we get ... are much lower than those reported in Chen et al. (2020)'. These degraded GCNII numbers are then the baseline against which GCNIII is shown to improve by 10–18 points on those datasets. If the original GCNII numbers from Chen et al. (2020) are used, the claimed large improvements—and possibly the 'outperforms GCNII on all 7 datasets' statement—need not hold. The comparison is further muddied by Appendix B: GCNIII's architecture changes per dataset (2–32 layers; αl ∈ [0,0.8]; λ ∈ [0,1.5]; γ ∈ [0.02,0.1]; techniques [1,0,0] on Chameleon, [1,1,0] on Cora/Citeseer, [1,1,1] elsewhere), so the result is not a single fixed Wide & Deep model beating a fixed baseline, but a per-dataset model-selection exercise. The theorem concern in Appendix A is real—the proof assumes parameter alignment and ReLU removal—but the Table 4 baseline mismatch is the more direct threat to the headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GCNIII, a Wide & Deep architecture for node classification that combines a linear wide branch (optionally with an \"Intersect memory\" graph-convolution operation) with a deep GCN component built from Initial residual and Identity mapping techniques. The authors introduce the term \"over-generalization\" for the observed phenomenon where a 64-layer GCNII has higher training error than validation error, present Theorem 4.1 as evidence that ultra-deep GCNII is unnecessary, and evaluate GCNIII on semi-supervised (Cora, Citeseer, Pubmed), full-supervised (those three plus Chameleon, Cornell, Texas, Wisconsin), and inductive (PPI) tasks. The paper also sketches a conceptual extension in which large language models construct node features. Code is released at https://github.com/CYCUCAS/GCNIII.","tokens_in":22608,"tokens_out":6937,"duration_ms":63862,"significance":"If the empirical claims were fully supported, GCNIII would be a simple and useful contribution: it unifies existing GCNII techniques and a linear branch in one framework, the semi-supervised baselines are retrained in a shared framework with 10 seeds, raw per-run numbers are reported in Appendix F, and the code is publicly available. The current manuscript, however, does not yet support the headline \"state-of-the-art\" claim. The full-supervised comparison in Table 4 reuses published baselines but retrains GCNII under a different framework with explicitly lower results; the GCNIII configuration changes across datasets in Appendix B to the point that on Chameleon the model is no longer a deep Wide & Deep network; the ablation in Table 6 uses a single seed; and the semi-supervised improvements are small relative to the reported standard deviations. Theorem 4.1 is heavily assumption-laden and is not load-bearing for the architecture's empirical value. These issues are fixable within the scope of a major revision, but the paper's central claims currently outrun its evidence.","major_comments":[{"comment":"The headline claim of outperforming GCNII on all seven full-supervised datasets rests on an inconsistent baseline comparison. For GCN, GAT, Geom-GCN variants, and APPNP the paper reuses metrics from Chen et al. (2020), but GCNII is retrained inside the authors' DGL framework, and the text states that on Chameleon, Cornell, Texas, and Wisconsin \"the results we get ... are much lower than those reported in Chen et al. (2020)\" (e.g., GCNII 54.4 on Chameleon, 59.46 on Cornell, 65.68 on Texas, 65.69 on Wisconsin). The 10-18 point margins over GCNIII on these datasets are therefore an artifact of a weakened baseline. Please report the original GCNII numbers from Chen et al. (2020) alongside the retrained ones, or retrain all baselines under the same framework; otherwise the \"state-of-the-art\" claim in the abstract and conclusion is not supported.","section":"§6.3, Table 4"},{"comment":"The GCNIII configuration is changed per dataset in almost every hyperparameter: 2-32 layers, α in [0,0.8], λ in [0,1.5], γ in [0.02,0.1], dropout in [0,0.8], and the technique vector varies. On Chameleon (Table 8), GCNIII uses 2 layers, α=0, λ=0, dropout=0, and techniques [1,0,0], which is essentially a shallow GCN with an Intersect-memory linear branch rather than a deep Wide & Deep model. If these choices were selected by peeking at test performance, the comparison is a per-dataset tuning exercise, not an evaluation of a single proposed architecture. Please describe the model-selection protocol (e.g., selection on validation), or fix a single architecture and tune only standard hyperparameters.","section":"Appendix B, Tables 7-9"},{"comment":"The ablation of the three techniques uses a single random seed (the text says \"we fixed the random seed as 42, so the results have a certain randomness\"). The observed differences are small (-0.4, +1.0, +0.1, +0.6, etc.) compared with the roughly 0.4-0.6 standard deviations reported in Table 2, so the conclusions that Initial residual is \"the most influential factor\" and the other two techniques \"have destabilizing effects\" are not supported. Please run the ablation over multiple seeds and report means with standard deviations, ideally with significance tests.","section":"§6.5, Table 6"},{"comment":"Theorem 4.1 is not a general statement about GCNII depth because the proof assumes that ReLU is removed, that inputs and intermediate activations are non-negative, that products of Identity-mapping parameter matrices are bounded by a constant C, and, crucially, that the parameters of f_K and f_{K+1} satisfy the \"misaligned equality\" W_I^{(l)} = tilde W_I^{(l+1)}. This last assumption identifies the extra layer's parameters with an existing layer's parameters, so the difference f_{K+1} - f_K collapses essentially by construction. The theorem should be reframed as a heuristic or proved under substantially weaker and more explicit assumptions; the notation also appears to confuse node count with hidden width (W_I^{(l)} is said to be in R^{n×n}).","section":"Theorem 4.1, Appendix A"},{"comment":"The semi-supervised SOTA claim rests on differences that are small relative to the reported variability: GCNIII vs GCNII is 85.6±0.4 vs 85.2±0.4 on Cora, 73.0±0.5 vs 72.8±0.6 on Citeseer, and 80.4±0.4 vs 79.8±0.4 on Pubmed. With 10 seeds these gaps are not shown to be statistically significant. Please add significance tests or confidence intervals, and soften the \"new state-of-the-art on all three datasets\" claim unless the improvements are supported.","section":"§6.2, Table 2"}],"minor_comments":[{"comment":"The \"Dropout is the key\" claim that removing all dropout from GCNII drops accuracy from \"over 85%\" to 82% is reported without configuration details or number of seeds; please add experimental details.","section":"Section 4"},{"comment":"\"Graph Transoformer\" is a typo for \"Graph Transformer\".","section":"Section 5"},{"comment":"\"Mdoel\" is a typo for \"Model\".","section":"Appendix F, Table 14 caption"},{"comment":"The column header \"Pumb.\" should be \"Pubmed\", and \"Wide Componet\" should be \"Wide Component\".","section":"Table 4 and Section 6.5"},{"comment":"The conclusion that Batch Normalization reduces generalization is based on single numbers without seed variability; please report means and standard deviations.","section":"Appendix C, Tables 10-11"},{"comment":"Section 6 says LLMs are not used in the experiments, yet the abstract says the paper \"explores the use of large language models (LLMs) for node feature engineering\"; please clarify that the LLM part is a conceptual proposal, not an evaluated contribution.","section":"Section 3.3 and Section 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is transparent about the GCNII retraining issue in §6.3, but transparency does not remedy the invalid comparison. Given the per-dataset configuration in Appendix B, I would ask the authors to clarify how hyperparameters were selected and to show that selection was validation-based rather than test-based. The LLM portion promises more than it delivers and should be trimmed or explicitly deferred to future work. The core Wide & Deep idea is plausible and the fixes (consistent baselines, multi-seed ablations, clearer model-selection protocol, and a weakened theoretical claim) are within the scope of a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Tom,\n\nYou should know two things about this one. First, the semi-supervised experiments are competently done: ten seeds, retrained baselines, a public implementation, and the wide-plus-deep GCNIII does make a small, consistent improvement over GCNII on Cora/Citeseer/PubMed. Second, the full-supervised SOTA claim does not survive contact with the paper's own Table 4. The authors reuse published numbers for GCN, GAT, APPNP, and Geom-GCN, but they retrain GCNII in their own framework and get numbers on Chameleon, Cornell, Texas, and Wisconsin that they explicitly say are much lower than Chen et al. (2020). Against that weakened baseline, GCNIII's 10–18 point gains are not evidence of SOTA. Use the original GCNII numbers and the improvement likely shrinks or disappears. This is the main load-bearing flaw, and the stress-test note has it right.\n\nWhat is genuinely new: adding a linear wide branch with batch norm (and optional graph propagation via their 'Intersect memory') to a GCNII-style deep branch is a simple combination I have not seen before, and it seems to help slightly when used with care. The analysis of dropout in deep GCNII is also useful—calling the train-error-higher-than-val-error pattern 'over-generalization' is a bit of a name for what dropout does, but the observation that the pre-embedding dropout is the critical one is a concrete and reproducible finding.\n\nThe soft spots beyond Table 4: Appendix B shows the architecture is tuned per dataset to the point of collapse. On Chameleon, GCNIII uses 2 layers, alpha=0, lambda=0, and only Intersect memory—that is effectively a linear model plus a two-layer propagation, not a deep Wide&Deep model. So the 'single flexible framework' claim is not tested as a fixed architecture; it is model selection. And Theorem 4.1 is a weak theoretical contribution: the proof removes ReLU, assumes aligned identity-mapping parameters between the K and K+1 layer models, and bounds products of parameters; those assumptions largely produce the conclusion. There is also a small algebraic slip in the derivation of K0. None of this breaks the empirical semi-supervised story, but the theorem should either be tightened or dropped.\n\nBottom line: this is an incremental but honest empirical paper. The right reader is someone working on deep GCN training recipes or cheap modifications to GCNII. It deserves a serious refereeing because the flaws are fixable and the code and seeds are there. I would send it to an empirical GNN venue with a request to compare against the original GCNII numbers, report GCNIII under a fixed (or properly nested-selected) architecture, and repair or remove the theorem.","headline":"A competent incremental GCN variant whose full-supervised SOTA claim rests on a retrained GCNII baseline that underperforms the published numbers; worth reviewing, but the headline needs rework.","tokens_in":23049,"tokens_out":4880,"would_cite":false,"duration_ms":47291,"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":"A graph net that pairs a wide linear branch with a deep GCN reaches new state-of-the-art results in node classification.","keywords":["node classification","graph convolutional networks","Wide & Deep architecture","GCNIII","GCNII","over-generalization","initial residual","node features"],"falsifier":"Fix one GCNIII configuration per dataset by selecting layers, $\\gamma$, and the technique set on validation accuracy alone; if the resulting accuracy no longer beats GCNII on the same splits, the paper's unified-architecture state-of-the-art claim is refuted.","tokens_in":21947,"feed_emoji":"🕸️","tokens_out":9137,"duration_ms":83774,"temperature":0.7,"pith_summary":"The paper argues that the best way to improve node classification is to stop choosing between shallow and deep graph networks and instead run both at once. It introduces GCNIII, which combines a wide linear classifier whose output can be smoothed by the graph via Intersect memory with a deep GCN built from Initial residual and Identity mapping connections, merging the two with a single weight $\\gamma$. On semi-supervised, full-supervised, and inductive benchmarks, GCNIII reports accuracy gains over GCNII on every dataset, with the largest gains on heterophilous web networks. The paper also diagnoses a phenomenon it calls over-generalization—training error persistently above validation error in deep GCNII—and argues that dropout is the key driver, with a theorem showing that beyond a point adding layers barely changes the output.","feed_headline":"Wide plus deep beats deep-only graph nets on node classification","feed_subtitle":"GCNIII pairs a memorizing linear branch with a graph convolution branch and edges out GCNII across 11 benchmarks.","key_machinery":"The object is the joint prediction $P=\\mathrm{Softmax}(\\gamma W(X)+(1-\\gamma)D(A,X))$, where $W$ is a linear classifier on node features (optionally multiplied by the graph operator $\\tilde{G}$ for Intersect memory) and $D$ is a deep GCN with Initial residual (mixing $H^{(0)}$ into every layer with weight $\\alpha_l$) and Identity mapping (weight $\\beta_l=\\lambda/l$, pushing each layer's linear transform toward the identity). The scalar $\\gamma$ controls how much the output leans on memorization versus graph-based generalization. The convergence claim is Theorem 4.1: for every $\\epsilon>0$ there is a $K_0$ such that for all $K>K_0$, $\\|f_{K+1}(A,X)-f_K(A,X)\\|_2<\\epsilon$, which the paper uses to argue that ultra-deep GCNII is unnecessary.","core_discovery":"The central claim is that the Wide & Deep recipe transfers from recommender systems to node classification: a wide, essentially linear branch memorizes raw feature-class associations, a deep GCN branch generalizes through graph propagation, and jointly training them inside one softmax output balances the two. The paper presents GCNIII as the first graph version of this architecture and repackages Initial residual, Identity mapping, dropout, and Intersect memory—a graph-convolution smoothing applied to the wide branch's output—as switchable hyperparameters. Empirically, it reports higher accuracy than GCNII on all ten benchmark datasets plus PPI, and it explains GCNII's behavior through over-generalization, identifying dropout as the key component that keeps deep GCNII training error above validation error.","pith_inferences":["If the per-dataset configuration is essential, the paper's real contribution may be a searchable space of techniques plus a diagnostic, rather than one fixed model; a testable extension is to select a single configuration per dataset on validation only and lock it for all datasets.","Theorem 4.1 suggests a practical effective-depth rule: stop adding layers once consecutive GCNII outputs are close in spectral norm, which could be checked cheaply on any dataset and compared with validation accuracy.","The LLM feature-engineering proposal is untested in this paper, but it implies a concrete experiment: replace bag-of-words features with text-attribute LLM embeddings on the same datasets, keep the wide branch on the sparse features, and see whether the gap over GCNII widens."],"forward_implications":["Practitioners can treat depth, wide-branch weight, and the three techniques as modular hyperparameters rather than choosing between shallow and deep GCN families.","Deep GCNII's effective behavior is concentrated in the layers nearest the output, so ultra-deep 64-layer configurations can be replaced by 8-to-32-layer GCNIII variants with comparable or better accuracy and faster training.","Dropout on the input feature embedding is a major source of deep GCN generalization; removing it drops Cora accuracy from above 85% to about 82%.","Node feature quality, not only graph structure, drives performance: sparse bag-of-words features markedly outperform random, dense, and one-hot features in the paper's 2-layer GCN experiments.","The over-generalization observation identifies a distinct failure mode—training error persistently above validation error—that deep GCN designers should monitor alongside over-smoothing."],"supporting_citations":[{"why":"Supplies the Wide & Deep architecture and the memorization-versus-generalization framing that GCNIII transfers to graphs.","marker":"Cheng et al. (2016)"},{"why":"Provides GCNII, the deep baseline, the Initial residual and Identity mapping techniques, and the hyperparameter settings GCNIII builds on.","marker":"Chen et al. (2020)"},{"why":"Supplies APPNP's initial-residual propagation that makes deep GCNs trainable and is a key baseline.","marker":"Gasteiger et al. (2019)"},{"why":"Supplies the residual and identity-mapping concepts used in GCNII and GCNIII.","marker":"He et al. (2016)"},{"why":"Supplies the vanilla GCN layer and the graph-convolution operator used as the deep branch.","marker":"Kipf & Welling (2017)"},{"why":"Supplies dropout, which the paper identifies as the key component behind over-generalization in deep GCNII.","marker":"Srivastava et al. (2014)"},{"why":"Provides the claim that classic GNNs are strong baselines and the hyperparameter settings the paper reuses while critiquing its training tricks.","marker":"Luo et al. (2024b)"},{"why":"Provides the citation-network datasets and the sparse bag-of-words node features used in most experiments.","marker":"Sen et al. (2008)"},{"why":"Provides Geom-GCN baselines and the random 60/20/20 split protocol for full-supervised node classification.","marker":"Pei et al. (2020)"}],"fun_headline_variants":["Wide & Deep architecture reaches node classification","GCNIII: Wide & Deep for graph node classification","From recommenders to graphs: Wide & Deep node classification","Wide + Deep branches balance memorization and generalization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim stands or falls on whether the per-dataset architecture and hyperparameter choices reported in Appendix B reflect principled model selection rather than tuning to the test set; if they were test-set choices, the comparison against fixed GCNII baselines is not a fair test of one unified model.","fun_headline_variants_meta":{"raw":{"variants":["Wide & Deep architecture reaches node classification","GCNIII: Wide & Deep for graph node classification","From recommenders to graphs: Wide & Deep node classification","Wide + Deep branches balance memorization and generalization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00106,"raw_usage":{"total_tokens":4414,"prompt_tokens":880,"completion_tokens":3534,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":3471}},"tokens_in":496,"tokens_out":3534,"duration_ms":24646,"temperature":1.0,"reasoning_tokens":3471,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:03:59.488188+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix one GCNIII configuration per dataset by selecting layers, $\\gamma$, and the technique set on validation accuracy alone; if the resulting accuracy no longer beats GCNII on the same splits, the paper's unified-architecture state-of-the-art claim is refuted.","supporting_citations":[{"cited_title":"Wide & deep learning for recommender systems","cited_arxiv_id":null,"evidence_quote":"Supplies the Wide & Deep architecture and the memorization-versus-generalization framing that GCNIII transfers to graphs."},{"cited_title":"Simple and deep graph convolutional networks","cited_arxiv_id":null,"evidence_quote":"Provides GCNII, the deep baseline, the Initial residual and Identity mapping techniques, and the hyperparameter settings GCNIII builds on."},{"cited_title":"Predict then propagate: Graph neural networks meet personalized pagerank","cited_arxiv_id":null,"evidence_quote":"Supplies APPNP's initial-residual propagation that makes deep GCNs trainable and is a key baseline."},{"cited_title":"Collective classification in network data","cited_arxiv_id":null,"evidence_quote":"Provides the citation-network datasets and the sparse bag-of-words node features used in most experiments."},{"cited_title":"C.-C., Lei, Y., and Yang, B","cited_arxiv_id":null,"evidence_quote":"Provides Geom-GCN baselines and the random 60/20/20 split protocol for full-supervised node classification."}],"review_version":1}