{"id":"4d74ae92-de12-4c46-9f9d-81b27666079d","arxiv_id":"2501.16656","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A review of graph neural network applications in transportation networks, covering traffic prediction, operations, industry practice, future directions, and public datasets and code.","lead":"This preprint surveys how graph neural networks are used to mine data from transportation networks, covering traffic prediction, traffic operations, and industry deployments such as Google Maps, Amap, and Baidu Maps. It is a useful entry point for researchers and practitioners who want a structured map of methods, datasets, and open problems in this area.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'since 2023' industry perspective is unsupported: all cited industry deployments in Table 4 and elsewhere are from 2020-2022, undercutting the review's claimed currency and gap-filling contribution.","rationale":"The reader's conditional verdict is appropriate. Our stress-test agrees with the reader's weakest assumption: the industry strand of the claimed scope is not current. We did not find a more fundamental flaw: the GNN background and academic 2023-2024 prediction coverage are broadly accurate, and the citation key errors, while real and in need of correction, do not by themselves invalidate the review's synthesis. The 'comprehensive' claim is also somewhat underspecified because no search protocol is given; however, the concrete falsifiable gap is the industry 'since 2023' claim, which is directly contradicted by the dates of the works cited. Therefore the verdict should remain CONDITIONAL (unchanged), with the condition that the authors correct the citation keys and either supply post-2023 industry evidence or adjust the stated scope.","tokens_in":29852,"tokens_out":6049,"duration_ms":53320,"concrete_test":"Audit the publication year of every industry deployment cited in Table 4 and Section 5.5 (Derrow-Pinion 2021, Dai 2020, Fang 2020, Huang 2022); if none is from 2023 or later, the abstract's 'since 2023' industry claim is unsupported and must be revised, and the review should either include post-2023 industry cases or explicitly lower the temporal claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims the review highlights progress 'from academic and industry perspectives since 2023,' and the introduction presents industry involvement as the key differentiator from prior reviews. The actual industry evidence does not support that temporal claim. Table 4 contains three deployments: Derrow-Pinion et al. (2021, Google), Dai et al. (2020, Amap), and Fang et al. (2020, Baidu). The only later industry item, DuETA (Huang et al., 2022), appears in Section 5.5 as a future-direction example, not in the industry review. No cited industry practice is from 2023 or 2024. Since the central claim is a comprehensive, up-to-date summary relevant to industry and filling a gap left by reviews that omitted industry deployments, the absence of any post-2022 industry material means the 'since 2023' industry perspective is currently an assertion without evidence. The citation key errors compound this: Table 3 assigns Xu et al. (2024b) to the BEGAN air-traffic model, while Xu et al. (2024b) in the reference list is the U.S. 20 dataset; Li et al. (2023a) is used for IG-Net, but the reference Li et al. (2023a) is the DCRNN benchmark/solution paper. These errors weaken the audit trail for exactly the tables the comprehensiveness claim rests on.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript reviews graph neural network (GNN) methods for data mining in transportation networks. It positions itself against earlier surveys by covering both traffic prediction and traffic operation, adding an industry-practice section on Google Maps, Amap, and Baidu Maps, and closing with future research directions and a collection of open datasets, code libraries, and tutorials. The review is organized around three research questions, with a 31-entry table of 2023–2024 prediction works, a shorter treatment of traffic operation and accident prediction, and a resource list. The authors claim this fills a gap left by prior prediction-focused reviews and offers an up-to-date, since-2023 academic and industry perspective.","tokens_in":30150,"tokens_out":5856,"duration_ms":55082,"significance":"If the reference and recency issues are fixed, this review is genuinely useful to the community: it consolidates dispersed works, provides structured tables of models, datasets, and industry deployments, and articulates testable hypotheses (urban streets versus highways, sensor placement effects, prediction-interval drivers) in Section 5.1. The industry-practice section is a real differentiator relative to the prior reviews listed in Table 2, and the resource compilation in Section 6 lowers the entry barrier for newcomers. The paper introduces no new methods or predictions, which is appropriate for a review; its value lies in coverage and organization, and these are currently weakened by citation inconsistencies and an unsupported temporal claim.","major_comments":[{"comment":"The abstract and Section 1 state that the review highlights progress since 2023 from academic and industry perspectives, but all industry deployments reviewed in Table 4 are from 2020 or 2021 (Derrow-Pinion et al., 2021; Dai et al., 2020; Fang et al., 2020). The only more recent industry item, DuETA (Huang et al., 2022), appears in Section 5.5 as a future-direction example rather than in the industry review. The claimed industry perspective since 2023 is therefore not supported by the cited evidence. Please add post-2022 industry material or revise the temporal claim to match the actual coverage.","section":"Abstract and §4.4/Table 4"},{"comment":"The BEGAN air-traffic model is cited as Xu et al. (2024b) in Table 3 and in the text of Section 4.1.1, but the reference list's Xu et al. (2024b) is the U.S. 20 traffic-assignment dataset paper; BEGAN is Xu et al. (2023). This is not a cosmetic slip because Table 5 and Section 6.1 rely on Xu et al. (2024b) for the U.S. 20 dataset, so the same citation key now refers to two different works. Please reassign BEGAN to Xu et al. (2023) and re-verify every Table 3 citation against the bibliography.","section":"Table 3 and §4.1.1"},{"comment":"IG-Net is listed as Li et al. (2023a) in Table 3, but reference Li et al. (2023a) is the DCRNN benchmark/solution paper by Fuxian Li et al., while IG-Net is Pei Li et al. (2023b), which is the key used in Section 4.1.1. The table and the main text therefore use inconsistent keys, and a reader cannot resolve the intended work from the reference list. Please correct the key and check the surrounding rows for similar mismatches.","section":"Table 3, metro-demand row"}],"minor_comments":[{"comment":"The phrase \"interoperability analysis\" in the discussion of Tygesen et al. (2023) should read \"interpretability analysis\" to match the topic of the subsection.","section":"§4.1.5"},{"comment":"The text \"canner gene modules\" appears to be a typo for \"cancer gene modules\".","section":"§3.5"},{"comment":"The bullet on robust learning refers to \"Property GNN,\" but the cited work (Jin et al., 2020) is generally known as ProGNN; please clarify or correct the model name.","section":"§3.3"},{"comment":"There are several small typos, including \"Bejing\" and \"Flordia\" in Table 3 and \"predcition\" in Section 6.3; these should be corrected in a final pass.","section":"Table 3 and §6.3"},{"comment":"The checkmark-based layout of Table 2 is hard to parse because the column headers are not visually aligned with the rows; consider using explicit yes/no entries in a normal table format.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a survey venue and has no apparent circularity problem. Its acceptance depends on fixing the reference-key collisions and calibrating the recency claim so that the stated contribution matches the actual table contents."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this one. First, it's a genuinely useful review: it covers traffic operation (vehicle routing, relocation, traffic assignment) alongside prediction, which earlier surveys mostly skipped, and it compiles datasets, code, and tutorials that practitioners will actually use. Second, the headline claim doesn't survive contact with the tables. The abstract says it covers industry progress 'since 2023,' but every industry deployment in Table 4 and Section 4.4 is from 2020-2022 (Google 2021, Amap 2020, Baidu 2020; DuETA is 2022 and appears only in the outlook). So the 'since 2023' industry perspective is asserted, not demonstrated.\n\nThe citation errors compound the problem. Table 3 assigns Xu et al. (2024b) to the BEGAN air-traffic model, but in the reference list Xu et al. (2024b) is the U.S. 20 dataset; the BEGAN paper is Xu et al. (2023). Similarly, Li et al. (2023a) is listed for IG-Net, but that key points to the DCRNN benchmark paper, not the IG-Net paper (which is Li et al. 2023b). These are easy to fix, but they sit in the exact tables that the comprehensiveness claim rests on, so they deserve attention before this is used as a reference.\n\nWhat's genuinely new and good: the industry-vs-academic comparison (graph construction, prediction targets, evaluation metrics like negative ETA) is a useful synthesis that I haven't seen in prior surveys. The future directions section is thoughtful—interval prediction, model simplification, physics-informed GNNs, end-to-end predict-then-optimize—and it names specific datasets (LargeST, U.S. 20) that could support those directions. The model taxonomy in Section 3 is mostly standard GNN material, which is fine for a survey, but the novelty is in the application and resource sections, not there.\n\nThe central descriptive content is broadly accurate and the structure is coherent. The main soft spots are the overstated temporal claim and the audit-trail issues. Both are fixable in revision. I'd send this to peer review rather than desk reject it; it deserves referee time, but the authors should either adjust the 'since 2023' framing or add post-2022 industry evidence, and they must clean up the citation keys. For my own work, I wouldn't cite it until those fixes land.","headline":"A competent review of GNNs for transportation that overstates its industry coverage: the 'since 2023' claim rests on deployments from 2020-2022, and the citation keys in the core tables are unreliable.","tokens_in":30641,"tokens_out":2265,"would_cite":false,"duration_ms":23833,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A survey maps graph neural networks across traffic prediction, operations, and industry practice since 2023.","keywords":["graph neural networks","transportation networks","traffic prediction","traffic operation","travel time estimation","survey"],"falsifier":"A checker could audit Table 3 and Table 4 by looking for GNN transportation papers published in 2023-2024 that are absent and would alter the main trends.","tokens_in":29678,"feed_emoji":"","tokens_out":2024,"duration_ms":17558,"temperature":0.7,"pith_summary":"This review argues that graph neural networks (GNNs) have become essential tools for data mining in transportation networks, and that prior surveys missed a large share of the action by concentrating on traffic prediction alone. The authors set out to establish a more complete map: GNN methods applied to traffic prediction, traffic operation, accident analysis, and large industrial deployments such as Google Maps, Amap, and Baidu Maps. A reader who accepts the review's scope can rely on it as a current guide to GNN use in transportation since 2023, complete with datasets and code resources.","feed_headline":"GNNs in transportation: prediction, operation, industry","feed_subtitle":"A review since 2023 mapping graph neural networks across traffic prediction, operations, and deployments like Google Maps.","key_machinery":"The organizing object is the GNN layer, written as $H^{l+1} = f(H^l, A, W)$, which combines neighborhood aggregation, a linear transformation, and a nonlinear activation. The review uses this single formulation to compare architectures across academic models and industry systems.","core_discovery":"The central claim is that a survey covering prediction, operation, and industry practice together gives a fairer and more useful picture of GNNs in transportation than the prediction-only surveys that preceded it. The review therefore catalogs GNN applications in traffic speed, flow, demand, transit, air traffic, accidents, vehicle routing, traffic assignment, and travel time estimation, while highlighting how industry models differ from academic ones.","pith_inferences":["A useful follow-up would be to verify whether the trends highlighted since 2023 continue into 2025 and beyond, since the review is a snapshot.","The review's comparison of graph construction suggests that standardizing road-segment-level benchmarks would help bridge academic and industry research.","The emphasis on GNN+LLM integration hints that transportation applications of that combination are still largely unexplored.","The described industry studies being proprietary suggests that open datasets with similar granularity, such as road-segment-level travel times, could accelerate progress."],"forward_implications":["Readers can locate GNN models for a specific transportation task and compare their base architectures and temporal modules.","Practitioners can see how industry models differ from academic models, particularly in graph construction and evaluation metrics.","Researchers can identify open problems, such as interval prediction, model simplification, and end-to-end learning for traffic operations.","The compiled datasets and code resources lower the entry barrier for new researchers entering GNN-based transportation studies."],"supporting_citations":[{"why":"Google Maps ETA prediction case study that motivates the industry GNN section.","marker":"(Derrow-Pinion et al., 2021)"},{"why":"Amap H-STGCN example of industry road-segment-level travel time estimation.","marker":"(Dai et al., 2020)"},{"why":"Baidu Maps ConSTGAT example of industry road-segment-level travel time estimation.","marker":"(Fang et al., 2020)"},{"why":"DCRNN is the foundational traffic prediction GNN and source of widely used benchmark datasets.","marker":"(Li et al., 2017)"},{"why":"GCN is the base architecture used by a majority of the surveyed traffic prediction studies.","marker":"(Kipf and Welling, 2016)"},{"why":"Prior review covering traffic prediction and some operations, which the present review extends.","marker":"(Rahmani et al., 2023)"},{"why":"Prior traffic prediction review used as a baseline comparison for coverage.","marker":"(Jiang and Luo, 2022)"},{"why":"Graph Network is the base module for Google's industrial ETA model.","marker":"(Battaglia et al., 2018)"}],"fun_headline_variants":["GNNs span transportation: prediction to operations","Graph neural networks: traffic prediction, operations, industry","Transportation GNNs reviewed: beyond just traffic prediction","GNNs in transit: prediction, operation, and industry use","GNN review: traffic prediction, operations, and real apps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's usefulness hinges on the selected 2023-2024 studies being representative of the whole field and the industry table being up to date.","fun_headline_variants_meta":{"raw":{"variants":["GNNs span transportation: prediction to operations","Graph neural networks: traffic prediction, operations, industry","Transportation GNNs reviewed: beyond just traffic prediction","GNNs in transit: prediction, operation, and industry use","GNN review: traffic prediction, operations, and real apps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1357,"prompt_tokens":884,"completion_tokens":473,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":392}},"tokens_in":500,"tokens_out":473,"duration_ms":4376,"temperature":1.0,"reasoning_tokens":392,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T11:38:09.032272+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A checker could audit Table 3 and Table 4 by looking for GNN transportation papers published in 2023-2024 that are absent and would alter the main trends.","supporting_citations":[],"review_version":1}