{"id":"6c2250e9-cbc5-4ef6-b357-372c0b33adf3","arxiv_id":"2607.09372","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Mixed-distribution GNN training improves transferable betweenness rankings (τ≈0.86–0.92 across families) with large inference speedups, while closeness remains highly topology-sensitive.","lead":"Message-passing GNNs can learn node rankings that approximate betweenness and closeness centrality, with mixed synthetic training improving betweenness transfer and inference up to ~98× faster than exact methods. This is useful for ranking important nodes in large networks where exact shortest-path centrality is too slow.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Mixed-training transfer claim is load-bearing but under-specified: real-world zero-shot results appear to use ER-only models, so the headline improvement is not shown where it matters most.","rationale":"The reader correctly flags fixed-depth message passing and synthetic generators as a weak assumption for real-world transfer, and correctly keeps the verdict CONDITIONAL. The more precise load-bearing gap is narrower: the mixed-training improvement that the abstract and strongest claim advertise is demonstrated only on BA/GRP synthetics (§4.4), while the real-world zero-shot numbers (§4.5) are not explicitly attributed to the mixed model. That attribution is what would convert the synthetic transfer result into evidence for the paper’s stated goal of transferable structural representations. Degree remains a strong cheap baseline on homogeneous graphs and closeness transfer is weak—both already noted by the reader—but those do not undercut the betweenness mixed-training claim itself. Settling the ER-only vs mixed comparison on the three real networks is a single, cheap, decisive check; until it is reported, CONDITIONAL remains the right call and confidence stays high because the rest of the evaluation protocol is transparent and reproducible via the public code link.","tokens_in":12162,"tokens_out":673,"duration_ms":6641,"concrete_test":"Re-evaluate the three real-world networks of §4.5 with both the ER-only betweenness checkpoint and the mixed ER+BA+GRP checkpoint under identical preprocessing and scoring; report per-network Kendall τ and the mixed−ER delta. If mixed does not raise τ by a clear margin (e.g., ≥0.05 on at least two of three nets, or mean τ above the current 0.603–0.782 band), the transfer claim remains synthetic-only and the real-world support for the strongest claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper’s central claim is that mixed-distribution training (ER+BA+GRP) improves betweenness transfer (GRP τ from 0.552 ER-only to 0.861 mixed; §4.4, Fig. 5). That claim is well supported on the three synthetic families. The load-bearing next step for the paper’s own framing—transferable structural representations usable beyond the training generators—is zero-shot real-world ranking (§4.5, Fig. 6). There the manuscript reports betweenness τ in [0.603, 0.782] on C. Elegans, Email-Eu-Core, and the Western US power grid, but never states whether those numbers come from the mixed model or from the ER-only model. Architecture and training details in §3.1–3.2 and Table 1 treat mixed training as a separate betweenness experiment; the real-world section only says “models trained on synthetic graphs.” If the real-world numbers are ER-only, the strongest transfer claim has not been stress-tested on the topologies the introduction and conclusion treat as the target. If they are mixed, the paper should report the ER-only vs mixed delta on the same three networks. Either way, the current write-up leaves the decisive transfer result ambiguous.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper formulates betweenness and closeness centrality approximation as supervised node-ranking with message-passing GNNs, using exact NetworkX labels and Kendall τ evaluation. On held-out ER graphs the models reach τ ≈ 0.851 (betweenness) and 0.894 (closeness); a dedicated ER model at N = 5,000 reaches τ = 0.938. Dual-pathway product fusion is proposed for betweenness. The central empirical claim is that mixed training on ER + BA + GRP improves betweenness transfer across those families (GRP τ from 0.552 ER-only to 0.861 mixed), while closeness remains topology-sensitive. Zero-shot real-world results and up to 97.7× inference speedup over exact Brandes are also reported.","tokens_in":12580,"tokens_out":855,"duration_ms":6355,"significance":"If the mixed-training transfer result holds under clearer real-world controls, the work is a useful, reproducible contribution to neural algorithmic approximation of centrality rankings. Strengths include exact supervision, held-out and cross-family evaluation, dual-pathway ablation (Table 3), Brandes/degree baselines (Tables 4–5), an N = 5,000 scalability run, and public code. The practical value is a fast ranking approximator whose training diversity is shown to matter for betweenness; the paper also cleanly documents that closeness is harder to transfer. Novelty is incremental relative to prior GNN centrality work, but the controlled isolation of training-distribution diversity is a clear, citable finding for the community.","major_comments":[{"comment":"§4.5 / Fig. 6: The load-bearing transfer claim is mixed training (§4.4: GRP betweenness τ 0.552 → 0.861). Real-world zero-shot numbers (τ ∈ [0.603, 0.782]) only say “models trained on synthetic graphs” and never state whether they come from the mixed model or the ER-only model (Table 1 treats mixed as a separate experiment). Report ER-only vs mixed τ on the same three networks, or state explicitly which model produced Fig. 6; without that, the headline transfer result is not stress-tested where the introduction and conclusion claim it matters most.","section":null},{"comment":"Tables 4–5: On ER graphs degree centrality already reaches τ = 0.886 (betweenness) and 0.923 (closeness), above the GNN; mixed GNN beats degree on BA/GRP but the paper still frames GNN ranking quality as the primary success. Clarify when and why a learned model is preferable to degree (e.g., under topology shift, for non-degree-correlated centrality, or for amortized multi-graph inference), and report degree on the real-world networks in §4.5 so the practical gain is measurable.","section":null},{"comment":"§4.6 Table 6 vs §4.7: Scalability is split across settings. Table 6 reports speedups and τ for models evaluated at N up to 1,000 (with a drop to τ = 0.712 at N = 500), while the N = 5,000 result (τ = 0.938) is a separately trained ER-only model with no corresponding inference-time comparison at that size. Align the claims: either report inference time and ranking quality for the N = 5,000 model against exact Brandes, or avoid presenting 97.7× and τ = 0.938 as a single continuous scalability story.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a competent methods paper that isolates training-distribution diversity for GNN centrality ranking and shows mixed ER+BA+GRP training lifts betweenness transfer on community graphs (GRP from ~0.55 to ~0.86). The N=5000 run (τ≈0.938) and the 97.7× inference speedup are useful numbers. It does not invent the problem—Maurya, Zou/CNCA-IGE, and BRAVA-GNN already do GNN centrality approximation—but it does measure transfer gaps more carefully than most of that line.\n\nWhat works: ranking loss + Kendall τ is the right objective; dual-pathway ablation is reported; Brandes sampling and degree are included; code is linked; limitations on closeness are stated plainly. Degree is already strong on ER (τ~0.89), so the GNN’s in-distribution win is modest, but the mixed model beats Brandes k=50 on BA/GRP, which is the more interesting comparison.\n\nSoft spots, in proportion. First, the load-bearing claim is mixed training for transferable structure. That claim is solid on the three synthetic families. Real-world zero-shot (§4.5) reports betweenness τ in 0.60–0.78 but never says whether those numbers are from the mixed model or ER-only. The text only says “models trained on synthetic graphs.” If ER-only, the headline transfer story was not stress-tested where the intro says it matters; if mixed, they should show the ER-only vs mixed delta on the same three nets. Second, closeness transfer is weak and they own that. Third, fixed L and the chosen generators leave open whether the representations survive high-diameter or spatial graphs—the power-grid closeness collapse is consistent with that. None of these sink the paper; they bound how far the “transferable structural representations” claim currently reaches.\n\nMath and data look standard and transparent: exact NetworkX labels, held-out splits, reported means/stds. Citations cover Brandes, sampling, GNN message-passing, and recent inductive centrality models; no obvious citation games.\n\nWho it’s for: people building or using neural approximators for network ranking who care about distribution shift. Worth a serious referee. I’d engage if I’m working on graph algorithmic reasoning or large-network ranking; otherwise skim the mixed-training and real-world sections and move on. Send to peer review.","headline":"Clean empirical GNN ranking study with real mixed-training gains on synthetic families and honest closeness limits; real-world transfer write-up is ambiguous and novelty is incremental.","tokens_in":13141,"tokens_out":602,"would_cite":false,"duration_ms":11930,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Training graph neural nets on mixed random-graph families yields transferable betweenness rankings and nearly 100× faster inference than exact algorithms.","keywords":["Graph Neural Networks","centrality approximation","node ranking","transfer learning","graph representation learning","complex networks","betweenness centrality","closeness centrality"],"falsifier":"Train the identical mixed-distribution architecture on the same three generators, then measure zero-shot Kendall tau on a held-out suite of real networks whose average shortest-path lengths and modularity lie well outside the synthetic range; a collapse of betweenness tau below the ER-only baseline would falsify the transfer claim.","tokens_in":13065,"feed_emoji":"🔗","tokens_out":647,"duration_ms":7556,"temperature":0.7,"pith_summary":"This paper treats betweenness and closeness centrality as supervised node-ranking problems for message-passing graph neural networks. Exact centrality scores label the training graphs; the models are judged by how well they recover the induced ranking (Kendall tau). On held-out Erdős–Rényi graphs the networks reach tau about 0.85 for betweenness and 0.89 for closeness; a dedicated model trained at N = 5 000 still reaches tau = 0.938. The central claim is that a single model trained on a balanced mix of Erdős–Rényi, Barabási–Albert and community-partition graphs generalizes far better across those families than an ER-only model, especially on modular graphs. Closeness remains more fragile under topology shift and transfers poorly to real networks. Inference is up to 97.7 times faster than exact betweenness, so the learned rankers can be used wherever relative node importance matters more than absolute scores.","feed_headline":"Mixed-graph training lifts betweenness ranking transfer","feed_subtitle":"GNN inference is up to 97.7× faster than exact algorithms; closeness stays topology-sensitive","key_machinery":"Dual-pathway message-passing network (one pathway on A, one on Aᵀ) whose layer-wise MLP scores are fused by element-wise product, trained with pairwise ranking loss and evaluated by Kendall tau.","core_discovery":"Message-passing GNNs can learn transferable structural representations for betweenness ranking when they are trained on a deliberately mixed distribution of random-graph families; the same mixed training does not rescue closeness, which stays highly sensitive to community structure and real-world topology.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Mixed training enables GNN betweenness rank transfer","GNNs transfer betweenness ranks via mixed-graph training","Mixed-graph training improves betweenness ranking transfer","Message-passing GNNs learn transferable betweenness ranks","Betweenness transfers across topologies with mixed GNN training"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That the fixed-depth message-passing layers and the three chosen synthetic generators already capture enough structural variety for the learned rankings to transfer to real networks whose diameters and modularity differ markedly.","fun_headline_variants_meta":{"raw":{"variants":["Mixed training enables GNN betweenness rank transfer","GNNs transfer betweenness ranks via mixed-graph training","Mixed-graph training improves betweenness ranking transfer","Message-passing GNNs learn transferable betweenness ranks","Betweenness transfers across topologies with mixed GNN training"]},"model":"grok-4.5","effort":"low","cost_usd":0.005858,"raw_usage":{"total_tokens":1513,"prompt_tokens":757,"num_sources_used":0,"completion_tokens":58,"cost_in_usd_ticks":58580000,"prompt_tokens_details":{"text_tokens":757,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":698,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":757,"tokens_out":58,"duration_ms":7681,"temperature":1.0,"reasoning_tokens":698,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T03:29:17.056811+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the identical mixed-distribution architecture on the same three generators, then measure zero-shot Kendall tau on a held-out suite of real networks whose average shortest-path lengths and modularity lie well outside the synthetic range; a collapse of betweenness tau below the ER-only baseline would falsify the transfer claim.","supporting_citations":[],"review_version":1}