{"id":"ef7760d4-605c-424a-9ff9-0637fae26ada","arxiv_id":"2601.23061","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"On two economic networks, maximum-entropy network models predict missing links about as accurately as gradient-boosting machine learning, and adding geographic distance makes the physics-style model the better performer.","lead":"This paper compares physics-style network models with a machine-learning algorithm for predicting missing links in trade and banking networks, finding that the simple models perform about as well. A generalist might read it to see whether interpretable, low-cost models can replace black-box AI for recovering hidden connections in economic and financial networks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random-deletion benchmark assumes missing links are MCAR; real trade/financial missingness is biased, so the white-box vs GBDT ranking may not generalize.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern: the random-deletion protocol is the empirical basis for the claim, and the footnote-1 transferability premise is asserted but untested. I agree this is the most important soft spot because the abstract and discussion generalize beyond the specific MCAR benchmark. Other issues—missing code/data, unstated GBDT hyperparameters, absent significance tests—are real reproducibility and evidential gaps, but they are secondary: even with perfect code, the central qualitative claim would still rest on the MCAR assumption. The proposed non-random deletion test is a direct, feasible check: it uses the same datasets and models and would show whether the ranking is an artifact of the deletion mechanism. Since the reader already conditioned the verdict on this concern, no verdict change is needed.","tokens_in":20537,"tokens_out":5662,"duration_ms":73106,"concrete_test":"Re-run the WTW and eMID experiments with non-random deletion protocols: (i) delete edges with probability inversely proportional to GDP or bilateral trade volume, emulating under-reporting by small economies; (ii) delete edges incident to low-degree/low-strength nodes with higher probability; (iii) delete edges with weight below a threshold, emulating reporting thresholds in eMID. Keep all other settings identical and compare CM, CMD, and fit2SM against the GBDT using the same TPR, JI, ACC, and AUROC metrics. If CM/CMD remain comparable or superior under these protocols, the transferability concern is resolved; if the GBDT pulls ahead or the ranking reverses, the paper's conclusion must be restricted to MCAR settings.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central comparison is performed under a missing-completely-at-random (MCAR) protocol: 10–50% of edges are deleted uniformly at random and models are scored on those deleted edges. The paper's own premise (footnote 1, §III.C) is that accuracy on the observed portion transfers to the missing portion, an exchangeability assumption that is never tested. In both empirical settings this is implausible. For the WTW, the binary matrix is derived from reported trade volumes; real missingness is concentrated among small/under-reporting economies and low-weight flows, not a uniform sample of edges. For eMID, the €50,000 reporting threshold and the restriction to Italian banks mean missingness correlates with exposure size and bank activity. Under such non-MCAR missingness, the observed degree sequence—the core input of CM, CMD, and fit2SM—is a systematically biased estimate of the true degrees, and the ranking of white-box models against the GBDT could change. The headline conclusion that white-box algorithms are viable competitors is therefore supported only for random deletion, not for the realistic missingness mechanisms that motivate link prediction in economic and financial networks. This is not an internal inconsistency, but it is a correctness risk for the general claim as stated in the abstract.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares two families of link-prediction algorithms on binary, undirected representations of the World Trade Web (WTW, 1990–2000) and the electronic Market for Interbank Deposits (eMID, 1999–2014). The physics-rooted family includes the Gravity Model, Configuration Model, Chung-Lu model, fitness models, fitness models with distances, and a fitness-induced 2-star model; the machine-learning family is a Gradient Boosting Decision Tree implemented with LightGBM. The protocol randomly deletes 10–50% of links (10 replicates), trains on the remaining graph, and evaluates TPR, JI, ACC, and AUROC. The central claim is that likelihood/entropy-based models—especially the Configuration Model and its distance-augmented variant—achieve accuracy comparable to, and sometimes slightly higher than, the GBDT when given the same input information, and that white-box models are therefore viable competitors that are more interpretable and computationally faster.","tokens_in":20882,"tokens_out":3198,"duration_ms":41070,"significance":"If the central claim holds, the paper provides a useful benchmark result for economic and financial networks: a simple, interpretable maximum-entropy model can match a state-of-the-art gradient-boosting classifier in link prediction when both use the same coarse inputs. The strength of the paper is its careful specification of the iterative recipes for the maximum-entropy models (Eqs. 28, 32, 35, 38, 45), the use of multiple metrics and deletion fractions, and the additional robustness checks in Appendices A–C, including a different train/test split protocol. These elements make the comparison reproducible in principle. However, the general conclusion is currently supported only under a missing-completely-at-random protocol, and the ML baseline is not fully specified, so the scope of the claim needs tightening.","major_comments":[{"comment":"The benchmark deletes edges uniformly at random (10–50%, repeated 10 times), and the paper's own premise (footnote 1, §III.C) is that a model reproducing the observed portion will reproduce the missing portion with the same accuracy. This exchangeability assumption is not tested. In both empirical settings, real missingness is plausibly non-MCAR: WTW links are derived from reported trade volumes (missingness concentrated among small/under-reporting economies), and eMID has a €50,000 reporting threshold and an Italian-bank subgraph. Under biased missingness, the observed degree sequence k_i(A_obs), the core input of CM/CMD/fit2SM, is a systematically biased estimate of the true degrees, and the ranking of white-box models against the GBDT could change. The abstract's general claim that white-box algorithms are 'viable competitors' is therefore only established for random deletion. Please","section":"§III.C, footnote 1; §II"},{"comment":"The GBDT implementation is described only as 'LightGBM' (ref. [26]); no hyperparameters, tuning procedure, or validation strategy are reported. Since the paper's conclusion is that white-box models perform comparably to this specific GBDT instance, the absence of the hyperparameter configuration makes the baseline unverifiable and potentially unrepresentative. In addition, the abstract and Discussion claim that white-box models are 'computationally faster' than the GBDT, but no runtime measurements or complexity comparison are provided anywhere in the manuscript. Please either report the LightGBM configuration and a runtime table, or remove/qualify the speed claim.","section":"§III.B; Abstract"},{"comment":"The central comparative claims—'comparable', 'slightly higher', 'outperforms'—are based on averages over 10 random link deletions, with standard deviations shown as error bars. No statistical significance tests are reported for any metric or dataset. Given that the bar overlaps are substantial in several panels (e.g., Fig. 2, TPR panels for CMD vs GBDT; Fig. 6, eMID AUROC), the claimed differences could be within sampling noise. Please add paired tests across the 10 realizations (e.g., Wilcoxon signed-rank tests) or otherwise demonstrate that the observed differences are not due to chance, and adjust the wording where significance is lacking.","section":"§IV, Figs. 2–6; Appendix A"}],"minor_comments":[{"comment":"The section numbering is duplicated: 'III. MISSING LINKS PREDICTION' is followed by another 'III. DATA DESCRIPTION'. The subsequent sections should be renumbered.","section":"Overall structure"},{"comment":"The expression for symmetrisation reads 'a_ij = Θ[w_ij + w_ij]'; this should presumably be Θ[w_ij + w_ji].","section":"§III (Data Description), eMID binarisation"},{"comment":"The notation assigns α = ln ε but uses ε inside the logarithm; the definition of ε is implicit. Please clarify whether α is a fitted constant or derived from ε.","section":"Eq. (10)"},{"comment":"The eMID daily snapshots have average degree as low as 4.61, yet the same 20% deletion fraction is used for all aggregation levels. A brief justification of why 20% is appropriate at daily scale, or a sensitivity check, would strengthen the appendix.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the experimental design is mostly standard, but the three major issues—the unqualified MCAR-to-real-world generalisation, the underspecified GBDT baseline, and the absence of significance testing—are load-bearing for the headline claim. The core derivations appear sound, and the appendices show effort toward robustness. With targeted revisions (non-MCAR experiments or a scoped claim, LightGBM configuration and runtime, and significance tests) the paper could become a solid contribution. I would not reject it at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe take-home: this paper gives practitioners a genuinely new benchmark — with matched feature sets, the maximum-entropy Configuration Model (and the distance-enriched CMD variant) matches a LightGBM gradient-boosting classifier on missing-link recovery, and on the WTW the CMD beats it. That specific comparison isn't in the earlier papers, and the setup is careful about putting each model on the same informational footing. If you work on economic or financial network reconstruction, this is worth knowing.\n\nWhat it does well: the protocol is standard and transparent — random deletion of 10–50% of edges, 10 repetitions, TPR/JI/ACC/AUROC, plus an appendix that repeats the exercise under the alternative train/test split of node pairs used by Mungo et al. The feature-matching discipline is the right way to compare a black-box learner against physics-rooted benchmarks. The equations are consistent with the cited literature and the fitting recipes are straightforward. The CMD-beats-GBDT result on the WTW is new.\n\nWhere it's soft. First, no code or data are released, and the LightGBM hyperparameters are not given. For a benchmark whose value depends on the exact training setup, that's a real reproducibility gap. Second, the \"computationally faster\" claim in the abstract is asserted, not measured; it should be shown or dropped. Third, there are no significance tests — error bars are shown but overlap is eyeballed. Given the headline is \"comparable, sometimes better,\" a paired test would make the claim firmer. Fourth — and this is the most substantive issue — the whole exercise uses random deletion as the missingness mechanism. The paper itself states (footnote 1, §III.C) that accuracy on the observed portion transfers to the missing portion; that's an exchangeability assumption that is never tested. Real missingness in trade and interbank networks is likely concentrated in small flows or below-threshold exposures, not uniform. So the ranking white-box vs GBDT is established for MCAR, and the abstract's more general phrasing overreaches. I don't think this invalidates the comparison — it is a fair and useful first pass — but the authors should scope the claim and, ideally, add a robustness check with non-uniform deletion.\n\nFor whom: this is for network scientists and econometricians doing link prediction or network reconstruction on economic data. It won't change the world, but it gives a concrete, honest data point that simple white-box models are competitive.\n\nMy recommendation: send it to peer review. The referee should ask for code, data, and hyperparameters, and push on the missingness assumption and the runtime statement. With those additions, this would be a solid applied-network-science contribution.","headline":"A careful benchmark showing max-entropy models can match a GBDT on missing-link recovery, with a real reproducibility gap and an overgeneralized missingness assumption.","tokens_in":21367,"tokens_out":3232,"would_cite":true,"duration_ms":31298,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["89.75.Fb","02.50.Tt"],"model":"deepseek-v4-flash","headline":"This paper claims that physics-rooted maximum-entropy models—using only coarse structural features like node degrees and geographic distances—predict missing links in economic and financial networks with accuracy comparable to, and in one c","keywords":["link prediction","maximum-entropy models","Configuration Model","gradient boosting","World Trade Web","interbank networks","network reconstruction","white-box vs black-box"],"falsifier":"Construct a controlled test in which missing links are removed by a systematic rule rather than at random—for instance, deleting all trade links below a volume threshold or all links incident to peripheral nodes—and measure whether the Configuration Model with Distances still matches or beats the gradient-boosting classifier; if its advantage disappears, the random-deletion assumption is the load-bearing premise.","tokens_in":20416,"feed_emoji":"🔗","tokens_out":4432,"duration_ms":52956,"temperature":0.7,"pith_summary":"This paper asks whether a black-box machine-learning model is actually necessary for predicting missing links in economic and financial networks. It compares the Gradient Boosting Decision Tree algorithm with physics-rooted maximum-entropy models—the Configuration Model and its fitness-based variants—on the World Trade Web and the electronic Market for Interbank Deposits. The central finding is that the white-box entropy models, which take as input only coarse quantities such as node degrees and distances, recover missing links with accuracy comparable to the GBDT, and the Configuration Model with Distances outperforms its GBDT counterpart on the trade network. If this holds, link prediction does not require fine-grained edge-level information; simple, interpretable, and fast models are viable competitors.","feed_headline":"Simple entropy models match gradient boosting at link prediction","feed_subtitle":"Using only degrees and distances, interpretable models match a fine-grained learner on trade and bank networks.","key_machinery":"The central object is the maximum-entropy random-graph ensemble, specifically the Configuration Model and its variants, which produce a probability p_ij for each non-observed pair by maximising entropy under constraints such as the observed degree sequence (and optionally total distances). These probabilities are used as ranking scores to select the |E_miss| most likely missing links. The comparison protocol—random deletion of a fixed share of links, averaging over ten realizations, and measuring TPR, JI, ACC, and AUROC—lets each white-box model be compared to a GBDT instance trained on exactly the same input features, isolating the contribution of the model family from the information fed t","core_discovery":"Running likelihood-based entropy benchmarks alongside a gradient-boosting classifier under the same training protocol, the authors find that the Configuration Model—which assigns each pair of nodes a connection probability from the observed degree sequence—matches the GBDT when both are fed the same 'endogenous' features, and that enriching the Configuration Model with geographic distances (CMD) yields higher TPR, JI, and AUROC than the GBDT on the World Trade Web. On the interbank market, each endogenous GBDT instance performs comparably to the purely structural white-box models. The paper interprets this as evidence that topological information is decisive in shaping network topology and t","pith_inferences":["A natural next test is to delete links in a biased way—e.g., removing all trade links below a volume threshold or all links incident to peripheral nodes—to see whether the white-box models' standing changes; real-world missingness in trade and financial data may not be random.","The success of the Configuration Model with Distances suggests that explicitly encoding spatial costs into entropy benchmarks could be a general principle for other geographic or transportation networks.","The comparable accuracy of the single-constraint CM to the fit2SM indicates that, for these dense economic networks, higher-order degree-correlation information adds little; testing on sparser networks would reveal where non-linear models start to matter.","The protocol compares models under equal information, but a practical comparison would also account for the cost of acquiring input features, which would likely strengthen the case for white-box models."],"forward_implications":["The Configuration Model alone, using only node degrees, can reach the accuracy of a gradient-boosting classifier trained on the same degree features, so structural information can do much of the work.","Adding a single exogenous quantity (geographic distance) to the Configuration Model produces a white-box model that outperforms the corresponding GBDT on the World Trade Web.","Because white-box models need only aggregated quantities (degrees, total links) rather than the exact pattern of observed edges, they can be applied when fine-grained edge-level information is unavailable or unreliable.","The paper's results suggest that the machine-learning advantage over simple models in link prediction is not general; it depends on the features and the training protocol.","White-box reconstruction-style models generalize more robustly across datasets, while the GBDT performs poorly or cannot run when trained on one network to predict another."],"fun_headline_variants":["Entropy models match gradient boosting in link prediction","White-box link prediction rivals black-box machine learning","Simple structural models compete with GBDT on missing links","Distance-enriched entropy model beats ML on trade networks","Physics-rooted methods hold their own against ML for links"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The benchmark's validity rests on the assumption that a model's accuracy on the observed portion of a network transfers to the missing portion—that is, the hidden links are missing at random and behave like the visible ones.","fun_headline_variants_meta":{"raw":{"variants":["Entropy models match gradient boosting in link prediction","White-box link prediction rivals black-box machine learning","Simple structural models compete with GBDT on missing links","Distance-enriched entropy model beats ML on trade networks","Physics-rooted methods hold their own against ML for links"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000493,"raw_usage":{"total_tokens":2233,"prompt_tokens":694,"completion_tokens":1539,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":1463}},"tokens_in":438,"tokens_out":1539,"duration_ms":16425,"temperature":1.0,"reasoning_tokens":1463,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T06:13:09.931723+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a controlled test in which missing links are removed by a systematic rule rather than at random—for instance, deleting all trade links below a volume threshold or all links incident to peripheral nodes—and measure whether the Configuration Model with Distances still matches or beats the gradient-boosting classifier; if its advantage disappears, the random-deletion assumption is the load-bearing premise.","supporting_citations":[],"review_version":1}