{"id":"f30ef10b-853a-442c-8514-479c228769cb","arxiv_id":"2411.12330","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Graph-aware Logistic Regression, a linear classifier on node features concatenated with each node's adjacency row, ranks first on average across 13 node classification datasets, ahead of eight GNN baselines.","lead":"This paper introduces GLR, a plain logistic regression model that receives, for each node, its own attributes plus a raw list of its connections in the graph. Across 13 real-world node classification benchmarks, GLR often matches or beats graph neural networks in accuracy while running up to two orders of magnitude faster.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GLR's 'ranks first' claim rests on a 0.15 average-rank margin over H2GCN that the 5-hour timeout rule can invert: completing H2GCN's timed-out Wikivitals+ run at any rank at or below 13 would put H2GCN ahead.","rationale":"The central claim is an empirical superiority claim, not a mathematical theorem, so its validity hinges entirely on the comparison protocol. The reader's weakest_assumption correctly identifies the fairness of the GNN baseline settings and the 5-hour timeout rule as the load-bearing premise. I agree, and I find the timeout rule to be the sharper issue because it is quantifiable: with GLR's average rank 3.23 and H2GCN's 3.38, a single completed H2GCN run on Wikivitals+ that lands at rank 13 or better flips the ordering. That is a very low bar for a model that is otherwise competitive across the board. The fixed hyperparameters are a secondary contributor but point in the same direction: the GNNs are not given per-dataset tuning while GLR has essentially no hyperparameters to tune. The paper's ablation studies and feature homophily analysis are useful and credible, but they do not rescue the headline 'ranks first' if the ranking is an artifact of the timeout penalty. The reader's CONDITIONAL verdict is appropriate: acceptance should require code verification and a demonstration that the rank advantage survives a fairer protocol. My check targets exactly the condition the reader already attached to acceptance, so it does not move the verdict.","tokens_in":16492,"tokens_out":7368,"duration_ms":76365,"concrete_test":"Re-run the Table 1 experiment for H2GCN on Wikivitals+ (and the other timed-out model–dataset pairs) without the 5-hour cap, record the actual ranks, and recompute the average ranks in the last column. If H2GCN's average rank falls below GLR's 3.23, the 'ranks first' claim is unsupported. As a robustness check, also recompute average ranks after excluding all datasets on which any model timed out.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline result, 'GLR ranks first across the 13 datasets' (Section 7.2), is an average-rank claim: GLR at 3.23 versus H2GCN at 3.38, a margin of only 0.15 per dataset. The evaluation protocol assigns the lowest possible rank (15) to any model exceeding the 5-hour training limit. H2GCN timed out on Wikivitals+; replacing that rank 15 with even a mid-table rank such as 8 lowers H2GCN's average by (15−8)/13 = 0.54, moving it to 2.84, ahead of GLR. In fact, any completed rank at or below 13 flips the ordering, so the outcome is sensitive to a single unobserved run. GraphSage, SGC, and GCNII also accumulate timeout ranks on Wikivitals, Wikivitals-fr, and Wikivitals+, so the ranking systematically penalizes slower GNNs. The paper frames this as weighting scalability equally with accuracy, but the abstract and conclusions state unqualified outperformance. This is not a flaw in the GLR model equation; it is a methodological premise of the comparison. The fixed, author-recommended GNN hyperparameters (Appendix C, Table 4), applied unchanged to all 13 datasets, further weaken that premise because larger or structurally different graphs may require different settings. The central claim therefore depends on a protocol that conflates accuracy with speed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Graph-aware Logistic Regression (GLR), a non-neural node classifier that concatenates each node's adjacency row with its feature vector and feeds the result into logistic regression. The authors evaluate GLR against eight GNN baselines and several non-neural baselines on 13 datasets spanning diverse sizes, densities, and homophily levels, using a unified k-fold cross-validation protocol. They report that GLR ranks first in average rank (3.23 vs. 3.38 for H2GCN), achieves competitive accuracy on homophilous and heterophilous graphs, and is up to two orders of magnitude faster. They also introduce 'feature homophily' as a complementary explanatory variable to label homophily and argue that GNNs underuse node features when those features are informative.","tokens_in":16819,"tokens_out":4475,"duration_ms":43099,"significance":"If the central claim is sustained, the paper is significant: it shows that a simple linear model that does not aggregate messages can match or beat specialized GNNs on a broad suite of benchmark graphs, while being far cheaper. The controlled evaluation framework, the inclusion of non-neural baselines, and the publicly available code are concrete strengths. The feature homophily analysis, though exploratory, is a useful step beyond label homophily. These strengths make the paper potentially important for the node-classification literature. The main reservation is that the headline 'ranks first' claim depends on a small average-rank margin that is sensitive to the handling of timed-out runs and to the choice of fixed GNN hyperparameters; these points need to be addressed before the stronger claims can be accepted.","major_comments":[{"comment":"The claim that 'GLR ranks first across the 13 datasets' rests on an average-rank margin of only 0.15 over H2GCN (3.23 vs. 3.38). Because models that exceed the 5-hour limit are assigned the lowest rank, H2GCN's timeout on Wikivitals+ contributes a rank of 15. If H2GCN had completed that run at any rank at or below 13 (e.g., a mid-table rank of 8), its average would improve by (15-8)/13 = 0.54 and it would overtake GLR. The result is therefore sensitive to a single unobserved run of a competing method. The paper should report a sensitivity analysis (e.g., a table of average ranks under different treatments of timeouts, or a per-dataset win/loss count) and qualify the 'ranks first' statement accordingly. Without this, the headline claim is not robust.","section":"Section 7.2, Table 1"},{"comment":"The abstract and Section 7.2 state that GLR 'outperforms both foundational and sophisticated state-of-the-art GNN models.' Table 1 does not support an unqualified reading: H2GCN has higher accuracy on Cora, Pubmed*, Citeseer, and CS, and GLR's advantage over H2GCN on many other datasets is within one standard deviation. The accurate statement is that GLR has the best average rank under the adopted timeout rule and is competitive with or better than H2GCN on most datasets, while being much faster. The abstract and conclusions should reflect this qualified claim rather than a blanket outperformance.","section":"Abstract and Section 7.2"},{"comment":"The GNN hyperparameters are fixed to the values proposed in the original papers for all 13 datasets, with no per-dataset tuning. In contrast, GLR's regularization strength appears to be a free parameter (though the paper does not state whether it is tuned per dataset). If GLR's regularization is tuned on each dataset while GNNs use fixed defaults, the comparison is not a fair test of accuracy under equal tunability. The paper should disclose the GLR regularization values and either tune the GNNs as well (e.g., with a small validation-based search) or explicitly justify that the fixed defaults are representative and that tuning would not change the conclusions.","section":"Appendix C, Table 4 and Section 6.2"},{"comment":"The complexity analysis is not rigorous enough to support the scalability claims. The paper states that GLR's training time complexity is O(n(n+L)), which is quadratic in node count and would be worse than the stated GNN cost O(mL+nLd) for large sparse graphs. If the intended meaning is that the number of parameters is n+L, that is not a training-time complexity. The empirical speedups in Figure 3 may be valid for the specific implementations and dataset sizes, but the asymptotic comparison as written is misleading and should be corrected or removed.","section":"Section 4.2"}],"minor_comments":[{"comment":"The evaluation framework states that k-fold cross-validation is used, but the value of k is never specified. Appendix C gives a 75%-25% split, which suggests k=4, but this should be stated explicitly.","section":"Section 6.1"},{"comment":"There are several typos: 'In this work, propose combining' is missing 'we'; 'compared to it best neural competitor' should be 'its best'; 'computation time reduction up to two orders' should be 'reduction of up to two orders'; 'results in a total of k training and testing experiments' should be 'k training and testing experiments per fold'.","section":"Section 1 and Section 8"},{"comment":"The SGC row is labeled '1-layer GAT(K=2)' which appears to be a copy-paste error; it should say SGC.","section":"Table 4"},{"comment":"The dataset name is written as 'Ogbn-arixv' in the text; it should be 'Ogbn-arxiv'.","section":"Section 7.4"},{"comment":"The paper does not report the regularization strength of the logistic regression used in GLR. Since this is the only free hyperparameter of the proposed method, reporting its value (or the tuning procedure) would improve reproducibility.","section":"Section 4.2 and Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a strong claim that is close to being supported but is undermined by the sensitivity of the average-rank statistic to the timeout rule and by the asymmetry in hyperparameter tuning. The authors should be encouraged to add a robustness analysis and to soften the wording. The complexity claim in Section 4.2 also needs correction. The scope (node classification, transductive setting) is appropriate for the journal, and the code availability is a plus."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful takeaway from this paper is empirical, not methodological: the concatenation of the adjacency row and the node features, fed into a plain logistic regression, is a deceptively strong baseline across a broad set of node-classification benchmarks, often matching or beating more elaborate GNNs while running much faster. That is worth knowing, and the paper's evaluation is generally careful, with k-fold CV, a diverse set of 13 graphs, and non-neural baselines included. The feature homophily diagnostic is a nice addition, though it reads as post-hoc explanation rather than a sharp test.\n\nThe method itself is a trivial concatenation of LINK and feature-based logistic regression, and the authors acknowledge its closeness to SGC. The new thing is the benchmark result, not the model.\n\nNow the soft spots. The abstract says GLR 'outperforms both foundational and sophisticated state-of-the-art GNN models.' Table 1 does not support that unqualified claim: GNNs win on Cora, Pubmed*, and Citeseer, and H2GCN beats or ties GLR on several datasets. The 'ranks first' claim rests on a 0.15 average-rank margin (3.23 vs 3.38). More importantly, the protocol assigns the lowest possible rank to any model that exceeds the 5-hour training limit. H2GCN timed out on Wikivitals+; if that run had completed at any rank at or below 13, the ordering flips and H2GCN would be ahead. The timeout rule is a defensible policy for a scalability-focused benchmark, but the paper presents the rank as if it were purely about accuracy. That needs to be stated more carefully.\n\nThe fixed GNN hyperparameters, taken from original papers and applied unchanged across all 13 datasets, are a known weakness in this literature; they make the baseline comparison less 'rigorous' than claimed, though I don't think it's fatal. There's also a questionable complexity statement: GLR is described as O(n(n+L)) per training step, which would be prohibitive on ogbn-arxiv (n=169k); the authors must be exploiting sparsity, and the stated bound is misleading as written.\n\nThe paper is honest about its limitations, and the ablation showing that concatenation beats either input alone is useful. The GitHub repo is linked; I did not verify it, but the protocol is described well enough to reproduce.\n\nWho is this for? Practitioners wanting a cheap, strong baseline, and anyone building graph benchmarks that omit non-neural baselines. It deserves a serious referee despite the over-claims. I would accept it for review but require the abstract to be toned down and the timeout sensitivity to be analyzed or at least acknowledged.","headline":"A simple logistic regression on concatenated adjacency and features is a strong, fast baseline, but the paper's 'outperforms GNNs' claim is a 0.15 average-rank margin that a single timeout rule can flip.","tokens_in":17317,"tokens_out":4205,"would_cite":true,"duration_ms":38731,"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":"Graph-aware Logistic Regression, a non-neural model that concatenates each node's adjacency row with its feature vector, ranks first across 13 node-classification datasets and runs up to two orders of magnitude faster than the best neural…","keywords":["node classification","graph neural networks","logistic regression","graph-aware logistic regression","homophily","feature homophily","scalability","non-neural baselines"],"falsifier":"Run the same 13-dataset benchmark with per-dataset hyperparameter search for each GNN (e.g., 50 random trials per dataset) and no 5-hour timeout; if the average rank of H2GCN or GCNII drops below GLR's, the paper's central claim is refuted.","tokens_in":16259,"feed_emoji":"⚡","tokens_out":10787,"duration_ms":95391,"temperature":0.7,"pith_summary":"Graph-aware Logistic Regression (GLR) is a non-neural classifier that builds each node's input by concatenating its row of the adjacency matrix with its own attribute vector, then fits a single logistic regression. The paper claims this recipe beats classical graph algorithms and state-of-the-art graph neural networks on a benchmark of 13 real-world attributed graphs, ranking first overall and first or second in 9 of 13 cases, while running up to two orders of magnitude faster than the best neural model. The authors argue that message passing, the core mechanism of GNNs, can drag in uninformative neighbor signals on heterophilous graphs, whereas a linear model on concatenated topology and features can learn to weigh the two sources. If the claim is right, strong node classification does not require a neural architecture, and feature-aware linear baselines should be standard in graph benchmarks.","feed_headline":"A non-neural model outperforms GNNs on 13 graph benchmarks","feed_subtitle":"Concatenating adjacency rows to node features lets logistic regression beat message-passing nets, up to 100x faster.","key_machinery":"The load-bearing object is the concatenated representation $h_u = \\mathrm{CONCAT}(A_u, X_u)$: the binary adjacency row of node $u$ followed by $u$'s feature vector, classified by $\\hat{y}_u = \\mathrm{softmax}(\\beta^T h_u + \\beta_0)$. This 'graph as a feature' move converts structure into an ordinary input vector, so the model can weight topology and attributes with a single learned linear map instead of running message passing. The paper compares GLR to SGC, noting that SGC is also a logistic regression but on smoothed features $S^l X$, whereas GLR keeps raw adjacency and raw features separate and concatenated. Training time is $O(n(n+L))$ with $n+L$ parameters, contrasted with the $O(mL + nLd)$ per-layer cost of message passing. The paper's second mechanism is feature homophily, $H_f(u) = \\frac{1}{d_u} \\sum_{v\\in \\mathcal{N}(u)} \\mathrm{sim}(X_u, X_v)$, a diagnostic used to explain when features, rather than labels, carry the signal.","core_discovery":"The paper's central discovery is that substituting the graph itself for the first layer of a GNN—using the raw adjacency row plus node features as the input to a linear softmax—is sufficient to outperform both foundational and specialized GNNs on diverse graphs. Under a k-fold protocol with fixed seeds, GLR's average rank over the 13 datasets is 3.23, ahead of H2GCN's 3.38, the best neural competitor; GLR is first or second in 9 of the 13 datasets, and it wins outright on most Wikipedia-based and heterophilous networks. The exceptions are the classic citation networks Cora, Pubmed, and Citeseer, where GNNs keep the lead. To explain this split, the paper introduces feature homophily—the average similarity between the attributes of connected nodes—and shows that on graphs with informative features (medium to high feature homophily), GNNs frequently fail to exploit the features, while GLR does. The authors also report that three highly homophilous graphs where GLR wins (CS, Photo, Ogbn-arxiv) have higher feature homophily than the citation networks, so label homophily alone does not predict which model will win.","pith_inferences":["A direct extension the authors do not test is applying the same concatenation recipe to link prediction and graph classification; if it transfers, the 'graph as a feature' principle would hold beyond node classification.","Because GNN hyperparameters are fixed at author-recommended values across all datasets, the superiority claim is protocol-dependent; per-dataset tuning or relaxed time limits could narrow or reverse the gap.","The feature-homophily diagnostic could be used as a selection rule before training: compute $H_f$ and choose GLR when it is high, a GNN when label homophily is high and feature homophily low."],"forward_implications":["GLR becomes a baseline every node-classification benchmark should include; it is simpler, faster, and in this comparison at least as accurate as the best GNN.","The claim that GNNs are the best for node classification needs qualification: on heterophilous and feature-rich graphs, a linear model on concatenated adjacency and features can win.","Feature homophily gives practitioners a cheap pre-training check: if a graph has medium-to-high feature homophily, a linear feature-aware model may outperform a message-passing network.","Scalability results imply large attributed graphs can be classified accurately without GPU training or aggressive sampling; GLR finishes where GNN baselines time out under the 5-hour budget."],"supporting_citations":[{"why":"Defines GCN, the foundational message-passing architecture that GLR is benchmarked against.","marker":"Kipf & Welling, 2017"},{"why":"Defines GraphSage, a sampled-neighbour GNN baseline in the comparison.","marker":"Hamilton et al., 2017"},{"why":"Defines GAT, the attention-based GNN baseline in the comparison.","marker":"Velickovic et al., 2018"},{"why":"Defines SGC, the simplified GNN whose logistic-regression form GLR is contrasted with.","marker":"Wu et al., 2019"},{"why":"Defines APPNP, a propagation-based GNN baseline with one of the strongest average ranks.","marker":"Gasteiger et al., 2019"},{"why":"Defines H2GCN, the best-performing neural competitor GLR must beat.","marker":"Zhu et al., 2020"},{"why":"Defines LINK, the adjacency-only logistic regression baseline that GLR extends with node features.","marker":"Zheleva & Getoor, 2009"},{"why":"Provides the diffusion model, a topology-only non-neural baseline in the benchmark.","marker":"Zhu, 2005"},{"why":"Supplies the evidence that split choice shifts model rankings, motivating the k-fold protocol.","marker":"Aleksandar & Günnemann, 2018"},{"why":"Supplies the Ogbn-arxiv dataset, the largest graph in the benchmark.","marker":"Hu et al., 2020"}],"fun_headline_variants":["Non-neural GLR outperforms GNNs on most graph benchmarks","Graph as a feature: logistic regression beats GNNs","Simple GLR model tops GNNs, up to 100x faster","Why a linear model beats GNNs: feature homophily","No neural net needed: graph-aware logistic regression wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that using each GNN's author-recommended hyperparameters unchanged across all 13 datasets, plus a 5-hour training cap with timed-out models ranked last, is a fair measure of each architecture's performance.","fun_headline_variants_meta":{"raw":{"variants":["Non-neural GLR outperforms GNNs on most graph benchmarks","Graph as a feature: logistic regression beats GNNs","Simple GLR model tops GNNs, up to 100x faster","Why a linear model beats GNNs: feature homophily","No neural net needed: graph-aware logistic regression wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000528,"raw_usage":{"total_tokens":2587,"prompt_tokens":1029,"completion_tokens":1558,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":1469}},"tokens_in":645,"tokens_out":1558,"duration_ms":15109,"temperature":1.0,"reasoning_tokens":1469,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:39:04.183649+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 13-dataset benchmark with per-dataset hyperparameter search for each GNN (e.g., 50 random trials per dataset) and no 5-hour timeout; if the average rank of H2GCN or GCNII drops below GLR's, the paper's central claim is refuted.","supporting_citations":[{"cited_title":"Semi-supervised classification with graph convolutional networks","cited_arxiv_id":null,"evidence_quote":"Defines GCN, the foundational message-passing architecture that GLR is benchmarked against."},{"cited_title":"Inductive representation learning on large graphs","cited_arxiv_id":null,"evidence_quote":"Defines GraphSage, a sampled-neighbour GNN baseline in the comparison."},{"cited_title":"Graph attention networks","cited_arxiv_id":null,"evidence_quote":"Defines GAT, the attention-based GNN baseline in the comparison."},{"cited_title":"Predict then propagate: Graph neural networks meet personalized pagerank","cited_arxiv_id":null,"evidence_quote":"Defines APPNP, a propagation-based GNN baseline with one of the strongest average ranks."},{"cited_title":"Beyond homophily in graph neural networks: Current limitations and effective designs","cited_arxiv_id":null,"evidence_quote":"Defines H2GCN, the best-performing neural competitor GLR must beat."},{"cited_title":"Semi-supervised learning with graphs","cited_arxiv_id":null,"evidence_quote":"Provides the diffusion model, a topology-only non-neural baseline in the benchmark."}],"review_version":1}