{"id":"3abf4472-6b65-44c0-9fef-3f92ce2f13b0","arxiv_id":"2501.01630","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A generative probabilistic classifier using node text, in/out-degree, and first-order neighbor-label counts achieves competitive accuracy on the Math Genealogy and ogbn-arxiv datasets.","lead":"The authors propose a probabilistic model that classifies nodes in directed graphs using node attributes, in/out-degree, and first-order neighbor label counts. The model is interpretable, unlike graph neural networks, and the authors introduce a new dataset adapted from the Math Genealogy Project.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 'superior to state-of-the-art' claim is not yet supported: ogbn-arxiv baselines use 128-d embeddings and 2020-era models, while the proposed model uses raw text; a matched-feature, repeated-seed re-benchmark is needed.","rationale":"The reader's weakest-assumption is the conditional independence in Eq. (3). That is a genuine limitation, but it is not the most load-bearing part of the central claim because a misspecified generative classifier can still rank labels well; the paper makes an explicitly comparative claim ('superior to state-of-the-art'). The evaluation in Tables 7 and 3 is where that claim either holds or collapses. Table 7 compares against 2020 OGB baselines that use only 128-d embeddings, while the proposed model uses raw title/abstract text; thus the reported accuracy gap is confounded by input features. Table 3 has no error bars, and the largest reported advantage over GCN is 0.0016 F1. A matched-feature, repeated-seed re-run on the official split is a single decisive experiment. If the model maintains its edge under matched inputs and against current leaderboard methods, the central claim stands; if not, the paper should be reframed as an interpretable baseline rather than a SOTA competitor. I therefore keep the reader's CONDITIONAL verdict: the work is promising but the headline performance claim needs this check before acceptance.","tokens_in":19797,"tokens_out":15512,"duration_ms":158110,"concrete_test":"Run a matched-input re-benchmark on ogbn-arxiv with the official split and 10 random seeds: (i) train GCN and GraphSage on the same raw-text unigram/bigram TF-IDF features used by the proposed model; (ii) train the proposed model with the standard 128-d embeddings (replacing omega_i with a class-conditional Gaussian/Naive Bayes on embeddings); (iii) include at least one current leaderboard method (e.g., GIANT+MLP or a recent GAT variant) with reported test accuracy. Report mean +/- std test accuracy for every method. The concern is settled if the proposed model beats the matched-feature GCN/GraphSage and the current leaderboard method by more than 1 point; if the advantage disappears or falls within noise, the SOTA claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is explicitly comparative: the abstract says the model is 'competitive with, and even superior to, state-of-the-art methods.' The evidence for that claim on ogbn-arxiv is Table 7, which reproduces 2020 baseline numbers from Hu et al.: MLP, Node2Vec, GCN, and GraphSage. Two confounds make this comparison the weakest link. First, input mismatch: those baselines consume the standard 128-dimensional average word embeddings, whereas the proposed model consumes raw title+abstract text expanded to unigrams/bigrams with TF-IDF. The +2.6-point gap over GCN could therefore reflect feature richness, not the graph mechanism being proposed. Second, the baselines are from 2020; the current OGB leaderboard contains published methods with test accuracy above 0.7432, so calling the model 'state-of-the-art' is inaccurate without qualification. In addition, no error bars or repeated-seed statistics are reported on either dataset; on Math Genealogy the ML F1 advantage over GCN is 0.0016, within plausible run-to-run noise. The conditional-independence assumption in Eq. (3) is a real limitation, but it is secondary: a misspecified generative score can still produce good rankings, whereas an unmatched comparison cannot support a comparative superiority claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a probabilistic generative model for node classification in directed attributed graphs. Labels are predicted by an ML or MAP score that combines multinomial terms for predecessor and successor label counts, per-label degree distributions, and text-attribute likelihoods under a conditional-independence assumption. Parameters are estimated from labeled training nodes, and predictions for unlabeled nodes are produced by an iterative pseudo-labeling procedure with validation-based selection of the iteration. The model is evaluated on a newly scraped Math Genealogy Project dataset and on ogbn-arxiv, with reported F1/accuracy values that are competitive with or higher than several baselines, including GCN.","tokens_in":20162,"tokens_out":9184,"duration_ms":93396,"significance":"Interpretable, non-neural baselines for node classification are useful, and the new Math Genealogy dataset with benchmark numbers is a genuine asset to the community. The model is computationally cheap and the code is public. If the comparisons were feature-matched and repeated over seeds, the paper could support a credible claim that a simple generative classifier is competitive with GNNs. As it stands, however, the central comparative claim is not yet supported, and the definitional inconsistency between the stated event-level parameters and the edge-level estimators used in the likelihood must be resolved before the model is fully coherent.","major_comments":[{"comment":"There is a mismatch between the parameters as defined and as used. Equations (5)-(6) define Theta_{i,j} and Xi_{i,j} as probabilities of a node's label given the existence of a predecessor or successor with a given label. In the likelihood, however, the multinomial factors g(p; d_in, xi_i) and g(s; d_out, theta_i) use xi_i and theta_i as per-edge categorical distributions over the labels of a single predecessor or successor conditioned on the node's own label. The estimators in Section 4 are edge-frequency ratios, e.g., \\hat{Theta}_{i,j} = #{(u,v): y_u=i, y_v=j} / #{(u,v): y_u=i}, which estimate the edge-level transition P(Y_v=j | Y_u=i), not the event-level probability in Eq. (5). These quantities are generally different in directed graphs. The paper should either define Theta and Xi directly as the per-edge transition probabilities used by the multinomial model, or derive the likelihood from the event-level definitions; otherwise the score being optimized is not the likelihood of the stated generative model.","section":"Section 3, Eqs. (5)-(6); Section 4; Section 5.1.1, Eqs. (17)-(18)"},{"comment":"The comparative claim in the abstract is not supported by the evidence on ogbn-arxiv. The baselines reported from Hu et al. consume the standard 128-dimensional average word embeddings and an undirected graph, whereas the proposed model consumes raw title+abstract text expanded to unigrams/bigrams with TF-IDF. The accuracy gap (0.7432 vs 0.7174 for GCN) could be due entirely to the richer input features or to the different graph convention; it does not isolate the proposed model's mechanism. On the Math Genealogy dataset (Table 3), the ML F1 of 0.5705 exceeds GCN's 0.5689 by only 0.0016, which is within plausible run-to-run variability, and no repeated-seed statistics or confidence intervals are reported on either dataset. A matched-feature comparison with the same text representation and graph directionality for all methods, across multiple seeds, is needed before the paper can claim superiority to state-of-the-art methods.","section":"Section 7.3, Table 7; Abstract"},{"comment":"The model's likelihood rests on the conditional-independence and i.i.d. multinomial assumptions: given a node's label, the predecessor-label counts, successor-label counts, degrees, and attributes are independent, and each neighbor label is drawn independently from a single label-specific categorical distribution. In real citation and advisor graphs, neighbor labels are typically correlated beyond the ego node's label (e.g., topic communities and co-citation patterns), and text and degrees may be correlated with neighbor-label composition. Because the ML and MAP scores are products of these likelihood factors, a violation changes the ranking among candidate labels. The paper provides no diagnostic or ablation testing the sensitivity of the results to this assumption; I would like to see either a synthetic-data validation of the generative model or an ablation comparing the full model against its text-only, degree-only, and neighbor-only components.","section":"Section 3, Eq. (3); Section 5.1.1"}],"minor_comments":[{"comment":"The product index appears as \\prod_{j=i}^{\\tau}, which should presumably be \\prod_{j=1}^{\\tau}.","section":"Eq. (13)"},{"comment":"The sentence 'These estimators correspond to the maximum likelihood estimators and provide unbiased estimates' is inaccurate after additive smoothing is introduced; smoothed estimators are not MLEs, and ratios of counts need not be unbiased.","section":"Section 4"},{"comment":"The update line is written as 'Update \\hat y_v^{(t)} \\leftarrow \\hat y_v', which is missing the iteration index; it should state that \\hat y_v^{(t)} is computed using the previous iteration's labels for neighbors.","section":"Algorithm 1"},{"comment":"When TF-IDF vectorization is used, the vector \\bar x_v contains real-valued weights, so Equations (12)-(14), which are based on integer counts, no longer define a valid multinomial likelihood; this should be acknowledged explicitly as a heuristic approximation.","section":"Section 6.2.1"},{"comment":"The first example is given as node ID 1526 in the text but as ID 12408 in the Table 4 caption; please make these consistent.","section":"Section 6.5 and Table 4"},{"comment":"The claim that ML achieves better F1 'due to balanced predictions across classes' is asserted without supporting evidence; a brief explanation of the mechanism would help the reader.","section":"Section 6.4.2"},{"comment":"For out-degree on ogbn-arxiv, only 23 of 40 labels pass the goodness-of-fit test, yet Equation (21) is retained as one option; since the frequency-based alternative is selected by validation, this is acceptable, but the text should state more explicitly that the parametric form is an approximation for many labels.","section":"Section 7.2.1"},{"comment":"Minor typos include 'lustrates' (Section 3), 'Algoritm 1', 'to to additive smoothing' (Section 2.2.1), and 't he' (Section 7.3).","section":"Throughout"},{"comment":"The paper introduces a new dataset but provides no direct data release or download link; the GitHub repository appears to contain code only. Please state data availability explicitly.","section":"Section 6/Data availability"}],"recommendation":"major_revision","confidential_remarks":"The core idea is a reasonable interpretable baseline, and the new dataset is a useful contribution. However, the paper's headline claim of superiority to state-of-the-art methods is currently based on a mismatched comparison, and the definitional inconsistency in the model's parameters is load-bearing for the likelihood derivation. Both issues are fixable within the scope of a revision, so I do not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate non-neural generative classifier for directed attributed graphs, and the Math Genealogy dataset is a real resource. But the headline claim of superiority over state-of-the-art is overreach. The ogbn-arxiv comparison in Table 7 pits the model's raw text (unigrams/bigrams, TF-IDF) against 2020 baselines that consume 128-d average word embeddings. The 2.6-point gap over GCN is just as explainable by feature richness as by the graph mechanism. There are also no error bars or repeated-seed results on either dataset; the Math Genealogy F1 advantage over GCN (0.5705 vs 0.5689) is inside run-to-run noise. The honest claim is \"competitive with standard baselines under different features,\" not \"superior to state-of-the-art.\"\n\nWhat the paper does well: the model is clearly specified, the estimation is standard and reproducible (code is public), and the authors do real diligence on the degree distributions with goodness-of-fit tests, reporting failures rather than sweeping them away. The per-node discrepancy tables are a nice, genuinely interpretable output. The new Math Genealogy benchmark with several baselines is a contribution on its own.\n\nThe soft spots, in proportion: the definitional inconsistency flagged in the reader's report is real. Equations (5)-(6) define Theta and Xi as probabilities conditioned on the existence of a neighbor with a given label, but the estimators in Section 4 and the likelihood terms in Equations (17)-(18) treat them as edge-level conditional probabilities in a multinomial draw. Those are different objects. It does not necessarily wreck the empirical results, but the model as written is not self-consistent. The conditional independence assumption in Eq. (3) is strong and untested; for a generative classifier that is a limitation, not a fatal flaw, but a sensitivity check (e.g., comparing against a model that drops one information source) would be easy and would strengthen the paper. The validation-selected iteration in the iterative prediction is acceptable model selection, not circularity, but it should be reported as a selected hyperparameter with its variance.\n\nWho this is for: readers working on interpretable graph models, statistical relational learning, or the Math Genealogy dataset. The paper deserves a serious referee, provided the authors are pushed to re-benchmark with matched features, report repeated-seed error bars, and reconcile the Theta/Xi definitions. I would send it out rather than desk-reject, and I would probably cite the dataset even if I remained skeptical of the SOTA claim.","headline":"A solid, interpretable generative model and a useful new dataset, but the 'superior to state-of-the-art' claim is not yet supported because the main comparison uses mismatched features and dated baselines.","tokens_in":20621,"tokens_out":1957,"would_cite":true,"duration_ms":22546,"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 generative classifier built from neighbor label counts, degrees, and text can match or beat graph neural networks on directed graphs while keeping predictions interpretable.","keywords":["node classification","directed graphs","generative classifier","interpretability","maximum a posteriori","Math Genealogy Project","ogbn-arxiv","conditional independence"],"falsifier":"Generate a synthetic directed graph whose true generative process includes correlations among the labels of a node's predecessors beyond the conditioning on the node's own label, then compare the model's MAP predictions with the true conditional probabilities. If prediction accuracy on such graphs stays near the ogbn-arxiv level, the independence assumption is not load-bearing; if it drops sharply, the reported benchmark numbers rely on that assumption.","tokens_in":19584,"feed_emoji":"📊","tokens_out":7356,"duration_ms":72646,"temperature":0.7,"pith_summary":"The paper tries to establish that a simple generative probabilistic classifier, rather than a learned neural network, can match or beat graph neural networks for node classification on directed attributed graphs while remaining interpretable. The model predicts an unlabeled node from four signals: the label counts among its predecessors, the label counts among its successors, its in- and out-degrees, and its own text attributes. The authors estimate parameters by maximum likelihood or MAP and test the model on the ogbn-arxiv citation network and on a newly assembled graph built from the Math Genealogy Project. On both benchmarks the model matches or exceeds the reported GCN and GraphSage baselines, and each prediction can be decomposed into per-component discrepancy terms that show why a label was chosen.","feed_headline":"Simple probabilistic classifier beats GNN baselines on two directed graphs","feed_subtitle":"Using only neighbor label counts, degrees, and text, its MAP version hits 0.7432 accuracy on ogbn-arxiv, above GCN and GraphSage.","key_machinery":"The load-bearing object is the factorized generative likelihood of Eq. (3), $f_{P,S,X|Y} = f_{P|Y}f_{S|Y}f_{X|Y}$, combined with multinomial models for the predecessor and successor label-count vectors and label-conditional degree and text distributions. Each component contributes a separate negative-log discrepancy term in the prediction objective, which is what makes the model interpretable: the winning label is the one whose total discrepancy is smallest, and the per-component terms show which signals supported or opposed it. The degree distributions are modeled parametrically when the data supports it, with a truncated power law with an atom at zero for the Math Genealogy out-degree and discrete log-normals for ogbn-arxiv in- and out-degrees, with empirical frequencies kept as an alternative.","core_discovery":"The paper's central claim is that a generative classifier built on a conditional-independence factorization of the joint distribution over a node's predecessor-label counts, successor-label counts, and attributes can serve as a competitive alternative to GNNs. Conditioned on a node's label, the predecessor counts follow a multinomial with class-dependent probabilities, the successor counts follow a multinomial with a possibly different class-dependent matrix, the degrees follow label-specific parametric or empirical distributions, and the text follows a multinomial-Naive-Bayes-style attribute model. Prediction becomes a sum of negative-log terms, one per component, and choosing the label that minimizes the total is the ML or MAP estimate. Reported results: on ogbn-arxiv the MAP version reaches 0.7432 test accuracy versus 0.7174 for GCN and 0.7149 for GraphSage; on the Math Genealogy graph the ML version reaches 0.5705 F1 versus 0.5689 for GCN, and MAP attains 0.7463 accuracy, the highest of the compared methods.","pith_inferences":["The ML/MAP choice effectively tunes a prior over class frequencies; on imbalanced labels, MAP's preference for frequent classes explains its higher accuracy and lower F1, and the same knob could be tuned per application cost.","The discrepancy decomposition suggests a natural active-learning rule: label next the node whose top-two total discrepancies are closest, since that margin is a readymade uncertainty estimate.","Because the transition matrices $\\Theta$ and $\\Xi$ encode cross-label edge propensities, the generative model is not inherently limited to homophily and could be used on heterophilous graphs in the same iterative refinement loop.","A stronger comparison would give the GNN baselines the same raw-text features the model uses; the reported ogbn-arxiv baselines rely on precomputed 128-dimensional embeddings and an undirected version of the graph, so part of the gap may reflect feature choice rather than model class."],"forward_implications":["Node classification on directed attributed graphs can be done without training a neural network, using parameters that have direct frequency-based estimates.","Predictions come with an explanation: the six discrepancy terms, for predecessor labels, successor labels, predecessor count, successor count, attributes, and prior, show which evidence supports or opposes each candidate label.","The method is inductive, since a node not present during training can be classified once the global parameters are estimated.","The iterative refinement scheme allows predictions to be updated in parallel, and the stopping iteration is chosen by validation performance.","The new Math Genealogy graph gives the community a benchmark where graph-aware methods clearly beat text-only methods, with BERT at 0.5028 F1 falling below GCN at 0.5689 F1 and the proposed model at 0.5705 F1."],"supporting_citations":[{"why":"Supplies the ogbn-arxiv dataset, its split, and the baseline accuracy numbers for MLP, Node2Vec, GCN, and GraphSage that the model is compared against.","marker":"[15]"},{"why":"Provides the GCN architecture and the GCN baseline used on both benchmarks.","marker":"[20]"},{"why":"Provides the GraphSage baseline and the inductive node-classification setting that the paper also targets.","marker":"[14]"},{"why":"Provides the Node2Vec embedding baseline reported for ogbn-arxiv.","marker":"[12]"},{"why":"Supplies the discretisation of the log-normal distribution used to model label-conditional in- and out-degree distributions.","marker":"[35]"},{"why":"Supports the use of discrete log-normal distributions for citation counts, justifying the parametric degree model on ogbn-arxiv.","marker":"[34]"},{"why":"Provides the Label Propagation baseline measured on the Math Genealogy dataset.","marker":"[16]"}],"fun_headline_variants":["Probabilistic model beats GNNs on directed graph node classification","Interpretable generative classifier outperforms GCN and GraphSage","New Math Genealogy benchmark: simple model tops GNN baselines","MAP estimate hits 74.3% on ogbn-arxiv, surpassing GNN baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes that once you know a node's own label, the labels of its predecessors, the labels of its successors, and its text attributes are completely independent of one another; real directed graphs often violate this, and the paper gives no sensitivity analysis showing how much the benchmark scores depend on that assumption.","fun_headline_variants_meta":{"raw":{"variants":["Probabilistic model beats GNNs on directed graph node classification","Interpretable generative classifier outperforms GCN and GraphSage","New Math Genealogy benchmark: simple model tops GNN baselines","MAP estimate hits 74.3% on ogbn-arxiv, surpassing GNN baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1524,"prompt_tokens":889,"completion_tokens":635,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":555}},"tokens_in":505,"tokens_out":635,"duration_ms":6541,"temperature":1.0,"reasoning_tokens":555,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:23:40.495051+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a synthetic directed graph whose true generative process includes correlations among the labels of a node's predecessors beyond the conditioning on the node's own label, then compare the model's MAP predictions with the true conditional probabilities. If prediction accuracy on such graphs stays near the ogbn-arxiv level, the independence assumption is not load-bearing; if it drops sharply, the reported benchmark numbers rely on that assumption.","supporting_citations":[{"cited_title":"Hamilton, Rex Ying, and Jure Leskovec","cited_arxiv_id":null,"evidence_quote":"Provides the GraphSage baseline and the inductive node-classification setting that the paper also targets."},{"cited_title":"node2vec: Scalable feature learning for networks, 2016","cited_arxiv_id":null,"evidence_quote":"Provides the Node2Vec embedding baseline reported for ogbn-arxiv."},{"cited_title":"The discretised lognormal and hooked power law distribu- tions for complete citation data: Best options for modelling and regression","cited_arxiv_id":null,"evidence_quote":"Supplies the discretisation of the log-normal distribution used to model label-conditional in- and out-degree distributions."},{"cited_title":"Are the discretised lognormal and hooked power law distri- butions plausible for citation data? Journal of Informetrics, 10(2):454–470, 2016","cited_arxiv_id":null,"evidence_quote":"Supports the use of discrete log-normal distributions for citation counts, justifying the parametric degree model on ogbn-arxiv."},{"cited_title":"Ben- son","cited_arxiv_id":null,"evidence_quote":"Provides the Label Propagation baseline measured on the Math Genealogy dataset."}],"review_version":1}