{"id":"2d8b3f20-7d4f-4fab-a665-2e0c4cc810d0","arxiv_id":"2508.00884","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"TSFusion, a hybrid of graph convolution and graph transformer with gated fusion, reports modest MAE and RMSE improvements over seven baselines on PeMSD4 and PeMSD8 traffic datasets.","lead":"This paper introduces TSFusion, a traffic-forecasting model that combines local graph-convolution patterns with global graph-transformer attention through a learned gate. It reports small accuracy gains over several baselines on two California highway datasets, but without shared code, error bars, or readable figures.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported TSFusion results for the same PeMSD4/15-min condition disagree (Table III MAE 8.59 vs Table IV MAE 8.35), a spread larger than the 0.19 margin over Graphormer-T; the outperformance claim is therefore not yet supported.","rationale":"TSFusion is a plausible hybrid of local graph convolution and global transformer attention, and the architecture narrative is coherent at a high level. But the central claim is empirical: the conclusion says TSFusion achieves the best accuracy among strong baselines. That claim rests on Table III. The ablation table (Table IV) reports full TSFusion at PeMSD4/15-min with MAE 8.35, while Table III reports 8.59 for the same setting. The unexplained spread exceeds the 0.19 MAE margin over Graphormer-T, so the tables do not support a stable 'best' conclusion at 15 minutes; the longer horizons are less affected but inherit the same protocol uncertainty. The workday-only vs sample-size mismatch independently shows the reported data protocol is not self-consistent. I agree with the reader's weakest assumption; no code, no error bars, and no configuration note leave the contradiction unresolved. A controlled 10-seed replication with mean and standard deviation, plus a sample-count check, would settle it. Because this is a load-bearing empirical inconsistency, the reader's REJECT verdict should stand unchanged; the manuscript should not be accepted or conditionally accepted without the replication.","tokens_in":17131,"tokens_out":11271,"duration_ms":109159,"concrete_test":"Re-run TSFusion on PeMSD4 15-min with the Section V.B data protocol and the grid-searched hyperparameters for 10 independent seeds, reporting mean and standard deviation of MAE and RMSE. Check whether Table III's 8.59/20.05 and Table IV's 8.35/19.82 are both within the same seed distribution, and whether the distribution's spread exceeds the 0.19 MAE margin over Graphormer-T. As an additional inexpensive check, count the number of 5-min samples after the stated workday-only filtering; if it does not equal 16,992, the preprocessing description and Table II cannot both be correct.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the empirical conclusion in Section VI that TSFusion 'achieves the best forecasting accuracies among the strong baselines.' The load-bearing evidence is the model-comparison table (Table III) and the ablation table (Table IV). These two tables are internally inconsistent for what should be the same condition. Table III reports TSFusion on PeMSD4 at the 15-min horizon as MAE 8.59 / RMSE 20.05; Table IV, also PeMSD4 15-min, reports the full TSFusion model as MAE 8.35 / RMSE 19.82. No sentence in the paper explains whether the two tables use different seeds, different selected checkpoints, or different hyperparameter settings. The numeric gap, 0.24 MAE, is larger than the reported margin between TSFusion and the strongest baseline Graphormer-T at the same horizon (8.78 - 8.59 = 0.19 MAE). So the published evidence is compatible with TSFusion being either first or not-first at 15 minutes, depending on which table is correct. A second internal inconsistency compounds this: Section V.B says only workday traffic data were used, but Table II reports sample sizes 16,992 and 17,856, which are exactly 288 x 59 and 288 x 62, i.e., all calendar days in the stated Jan-Feb and Jul-Aug intervals, not weekdays. Since 16,992 / 288 = 59 and 17,856 / 288 = 62, the workday-only filter cannot produce these sample sizes. Either the preprocessing description or the data statistics are wrong, so it is unclear what data generated Tables III and IV. With no code and no error bars, the claim of consistent outperformance is not verifiable from the manuscript as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes TSFusion, a hybrid traffic-flow forecasting model that combines a local temporal-spatial graph convolutional encoder with a graph-transformer encoder and adaptively fuses their embeddings through a gated residual mechanism. The model is evaluated on PeMSD4 and PeMSD8 against seven baselines, with MAE/RMSE reported at 15-, 30-, and 45-minute horizons, plus an ablation study and robustness experiments under Gaussian noise, missing data, and adversarial perturbations. The central claim, stated in Section VI, is that TSFusion achieves the best forecasting accuracy among strong baselines and defends against these perturbations in the feature space of the traffic graph dataset.","tokens_in":17530,"tokens_out":5112,"duration_ms":51681,"significance":"If the reported results are reliable, the paper offers a modest but reasonable architectural contribution: fusing local graph-convolutional patterns with global graph-transformer attention through a learned gate is a sensible design, and the multi-horizon comparisons, ablations, and robustness scenarios are the right kind of evidence for a traffic-forecasting paper. The method is not circular: the gated fusion and attention parameters are learned on training data and evaluated on held-out test data, and the paper does not claim a parameter-free derivation. However, the empirical evidence as presented does not yet support the central claim because of internal numerical inconsistencies and the absence of statistical reliability measures; the claimed improvements over the strongest baseline are small relative to the observed inconsistencies across tables.","major_comments":[{"comment":"The same condition, PeMSD4 at the 15-minute horizon for the full TSFusion model, is reported as MAE 8.59 / RMSE 20.05 in Table III and as MAE 8.35 / RMSE 19.82 in Table IV. The 0.24 MAE discrepancy is larger than TSFusion's reported margin over Graphormer-T at that horizon (8.78 - 8.59 = 0.19), so the evidence does not establish that TSFusion is the best model at 15 minutes. Please clarify whether the two tables use identical data splits, random seeds, and hyperparameters; if they differ, report both configurations and state which one supports the headline conclusion in Section VI.","section":"Table III vs. Table IV"},{"comment":"Section V.B states that only workday traffic data were used in all experiments and that redundant stations were removed, leaving 307 and 170 stations for PeMSD4 and PeMSD8, yet Table II reports sample sizes of 16,992 and 17,856, which are exactly 288 x 59 and 288 x 62, i.e., all calendar days in the stated two-month intervals. A workday-only filter cannot produce these counts, and the text also says the two datasets span January-February and July-August while Table II says 'Time duration 1 month'. Please reconcile the preprocessing description with the actual samples used, because this determines what generated all subsequent tables and figures.","section":"Section V.B and Table II"},{"comment":"The paper states that each model was evaluated 10 times and the average was taken as the final result, but it reports no standard deviations, confidence intervals, error bars, or significance tests. The reported margins over Graphormer-T are small (0.19 MAE at PeMSD4/15min and 0.36 MAE at PeMSD8/15min), while the ablation variants in Table IV differ from the full model by up to 0.67 MAE, so the claimed consistent outperformance may lie within run-to-run noise. Please add variance information or paired significance tests across the 10 runs for the main comparisons and the ablations.","section":"Section V.D"},{"comment":"The robustness analysis is presented only through figures with no numerical values, error bars, or experimental details. In particular, the adversarial attack setup is not specified: there is no perturbation budget, attack algorithm, or iteration count, and the TSFusion-adv variant does not appear in the main tables. As a result, the conclusion's statement that the model 'can defend against' Gaussian noise, missing data, and adversarial attacks is not quantitatively supported. Please report the underlying numbers, the exact attack configuration, and the uncertainty of the robustness estimates.","section":"Section V.G and Fig. 6"}],"minor_comments":[{"comment":"The baseline list in Section V.C names 'DCRNN', but Table III reports 'DCRGNN' for that row; if this is a typo, correct it, and if it is a different model, define it in the baseline description.","section":"Section V.C and Table III"},{"comment":"Equation (8) says 'the matrix V is the learnable weight matrix,' but Eq. (8) uses V only through the value embeddings defined in Eq. (6), while the output projection multiplied by V appears in Eq. (9); please align the notation to avoid confusing the value matrix with the output projection matrix.","section":"Eq. (8) and Eq. (9)"},{"comment":"References [23] and [34] are the same DCRNN paper; please remove the duplicate and renumber accordingly.","section":"Section II.A and references"},{"comment":"The notation for the history window is inconsistent: Eq. (2) uses M for the window length while Definition 2 introduces τ+1; please use one symbol throughout.","section":"Section III, Eq. (2)"},{"comment":"In the manuscript version provided, Figures 5 and 6 do not render as plots but appear as uninterpretable encoded strings; please ensure the final PDF contains legible figures with axis labels and legends.","section":"Figures 5 and 6"},{"comment":"The sentence 'lots of researchers determine to curve the inherent geographical connections' appears to contain a typo; 'curve' should likely be 'capture' or 'mine'.","section":"Section II.B"}],"recommendation":"major_revision","confidential_remarks":"The core architectural idea is coherent, but the empirical foundation needs substantial rework: the Table III/IV discrepancy and the sample-size contradiction with the workday-only description are load-bearing and must be resolved with a consistent experimental protocol. It would be valuable for the editor to request the code or a detailed configuration table, since the current report does not let a reader determine which numbers are canonical. I am not alleging misconduct; the discrepancies may arise from different seeds, checkpoints, or preprocessing variants that the authors simply forgot to disclose."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead TSFusion. The idea is sound: combine a local spatio-temporal graph convolution encoder with a Graphormer-style global transformer and let a learned gate balance them. That's a natural hybrid, clearly described, and the ablation study is a legitimate attempt to dissect the components. The robustness experiments (Gaussian noise, missing data, adversarial features) are a plus. So the architecture itself is worth a look.\n\nThe problem is the evidence. Table III gives TSFusion on PeMSD4 at 15 minutes as MAE 8.59 / RMSE 20.05. Table IV, same condition, gives 8.35 / 19.82. The gap is 0.24 MAE, bigger than the 0.19 edge over Graphormer-T in Table III. The paper never explains this. Also, Section V.B says only workday data were used, but Table II reports 16,992 and 17,856 samples, which are exactly 288×59 and 288×62 — all calendar days, not weekdays. Either the preprocessing description or the data statistics are wrong.\n\nNo error bars or significance tests are given despite claiming 10 runs, and there's no code. The robustness section introduces TSFusion-adv, an adversarially trained variant, but never says how the adversarial training was done. The baseline set is thin: no Graph WaveNet, STSGCN, or other recent strong models.\n\nThese are fixable issues. The architecture is coherent, the writing is mostly clear, and the idea isn't a repackaged known method. But the numbers as reported don't support the conclusion that TSFusion consistently beats strong baselines. If the authors can reconcile the tables, release code, and add error bars, this could be a modest incremental contribution. As it stands, I wouldn't rely on the results.\n\nFor us: this is a useful case study in how reporting inconsistencies can undermine a plausible empirical claim. I'd bring it to reading group, but I wouldn't cite it yet. I think a serious editor should send it to peer review, precisely so these issues get caught and fixed. My verdict is that it needs major revision before we can trust its numbers.\n\nBest.","headline":"Reasonable hybrid architecture, but the reported numbers don't agree across tables, so the central outperformance claim is unverifiable as written.","tokens_in":18076,"tokens_out":3283,"would_cite":false,"duration_ms":31362,"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":"Traffic forecasts improve when local graph convolution and global transformer attention are adaptively gated together.","keywords":["traffic flow forecasting","spatio-temporal graph convolutional networks","graph transformer","adaptive gated fusion","multi-grained representation learning","robustness to feature-space noise","PeMSD4","PeMSD8"],"falsifier":"A direct re-run of the same protocol (workday-only PeMSD4 and PeMSD8, stations separated by more than 3.5 miles, 60-minute history, 15/30/45-minute targets) that finds any of Graphormer-T, STGCN, or DCRNN matching or beating TSFusion on MAE or RMSE, or an ablation in which removing the gate or the transformer branch leaves the error unchanged, would contradict the central claim.","tokens_in":16959,"feed_emoji":"🚦","tokens_out":6772,"duration_ms":61499,"temperature":0.7,"pith_summary":"The paper claims that accurate traffic flow forecasting needs both local and global spatial-temporal evidence, and that neither a graph convolutional encoder nor a transformer encoder alone is enough. It presents TSFusion, a hybrid network that extracts local patterns from stacked graph-convolution blocks and global patterns from a graph-enhanced transformer, then combines the two through a learned gated fusion unit with residual connections. On the PeMSD4 and PeMSD8 highway datasets, the authors report that TSFusion achieves the lowest MAE and RMSE among seven baselines at 15-, 30-, and 45-minute horizons, and that it degrades less than the baselines when input features are corrupted by Gaussian noise, missing values, or adversarial perturbations. A sympathetic reader would take the paper's claim to be that adaptive multi-grained fusion is the operative ingredient for both accuracy and stability in graph-based traffic forecasting.","feed_headline":"Gated fusion of local and global patterns wins traffic forecasting","feed_subtitle":"Hybrid graph network beats seven baselines on PeMSD4 and PeMSD8 and resists noise, missing data, and attacks.","key_machinery":"Three components carry the argument. The temporal-spatial encoder stacks 'sandwich' blocks, each placing a 1-D causal convolution followed by gated linear units between two graph convolution layers, producing time-aware local node embeddings $L_t$. The graph transformer encoder augments these embeddings with learnable in-degree/out-degree centrality vectors and injects edge information through averaged shortest-path features $c_{ij}$ added inside the self-attention score before softmax, so attention can retrieve long-range 'invisible' relations between nodes not connected in the road graph. The adaptive fusion module learns a per-feature gate $Gate = \\sigma(MLP(\\tilde{B}_t) W_g^\\top + b_g)$ and forms $H_t = Gate \\odot MLP(\\tilde{B}_t) + (1 - Gate) \\odot MLP(L_t)$, with residual connections around the transformer layer; this gate is the mechanism that balances local and global evidence, and the centrality/edge biases are what make the global branch spatial rather than purely sequential.","core_discovery":"On its own terms, the paper establishes that single-grained encoders are the bottleneck in traffic graph forecasting. A graph convolution branch that message-passes over the pre-defined road network captures local station-to-station dependencies but misses functional similarities between distant stations (e.g., two transportation hubs); a transformer-only model captures global relations but tends to distort geographic structure. TSFusion feeds the local branch's embeddings into a graph transformer whose self-attention is biased by node in/out-degree centralities and shortest-path edge features, then merges the two embeddings with a gate $Gate = \\sigma(MLP(\\tilde{B}) W_g + b_g)$ and residual combination $H = Gate \\odot MLP(\\tilde{B}) + (1 - Gate) \\odot MLP(L)$. The reported experiments on PeMSD4 and PeMSD8 show TSFusion outperforming HA, ARIMA, LSVR, LSTM, DCRNN, STGCN, and Graphormer-T on MAE and RMSE at all tested horizons, with the gap widest at 45 minutes; ablations show removing the global branch or the gate hurts more than removing the residual connection, and robustness tests show the fused model holds up better than the compared baselines under feature-space noise, missing data, and adversarial attacks.","pith_inferences":["Beyond the paper: the same gated local/global fusion recipe should transfer to other spatio-temporal graph forecasting tasks, such as crowd flow, energy demand, or water network prediction, where distant nodes with similar functional roles are common; swapping in another public sensor dataset would test whether the gate learns a nontrivial mix.","Beyond the paper: the ablation that removing global attention hurts more than removing local aggregation suggests that on these highway networks 'invisible' functional relations carry more signal than geographic proximity; inspecting the learned gate values per station would show whether transfer hubs and business districts systematically receive high global weights.","Beyond the paper: the robustness comparison includes only STGCN, Graphormer-T, and DCRNN; retraining those baselines with the same adversarial training used for TSFusion-adv would isolate whether the resilience comes from the fusion mechanism or simply from a larger, more expressive model.","Beyond the paper: since the paper reports 10-run averages without spread, a paired statistical test across seeds for the TSFusion-versus-Graphormer-T gap would tell whether the headline improvement is reliable or within noise."],"forward_implications":["Hybridizing a local graph-convolution branch with a global graph-transformer branch is, if the reported results hold, a more accurate design for short-term traffic forecasting than using either branch alone.","The performance gap grows with the forecast horizon, so the global branch contributes most where local persistence is weakest.","The ablation ordering (global branch and gate matter more than the residual connection) says future architectures should invest in adaptively mixing multi-scale evidence rather than in deeper skip paths.","The robustness results imply that fusion itself confers some protection against feature-space corruption; the added adversarial training only yields marginal improvement on Gaussian noise and no clear gain on missing data."],"supporting_citations":[{"why":"Supplies the PeMSD4 and PeMSD8 datasets and their station/feature structure on which every comparison and robustness test runs.","marker":"[39]"},{"why":"Defines the STGCN baseline and provides the spatial-temporal convolution block design that the local branch adapts into 'sandwich' blocks.","marker":"[38]"},{"why":"Defines the DCRNN baseline representing diffusion-based graph recurrent forecasting that TSFusion must outperform.","marker":"[23]"},{"why":"Provides the Graphormer baseline (adapted as Graphormer-T) against which the transformer-only approach is compared.","marker":"[43]"},{"why":"Supplies the self-attention and multi-head attention machinery that the graph transformer branch augments with centrality and edge features.","marker":"[26]"},{"why":"Provides the adversarial training method used to form the TSFusion-adv variant in the robustness experiments.","marker":"[44]"}],"fun_headline_variants":["Gated fusion of graph scales boosts traffic forecast accuracy","Merging local and global traffic patterns via learned gate","Adaptive gate fuses local and global graph patterns","Stable traffic forecasting via gated multiscale fusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the experimental comparison being trustworthy: the 10-run averages are presented without error bars, and the ranking of TSFusion over the seven baselines would lose support if run-to-run variance, preprocessing mismatches, or configuration differences between the tables are large enough to reverse the ordering.","fun_headline_variants_meta":{"raw":{"variants":["Gated fusion of graph scales boosts traffic forecast accuracy","Merging local and global traffic patterns via learned gate","Adaptive gate fuses local and global graph patterns","Stable traffic forecasting via gated multiscale fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000892,"raw_usage":{"total_tokens":3874,"prompt_tokens":998,"completion_tokens":2876,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":2811}},"tokens_in":614,"tokens_out":2876,"duration_ms":19265,"temperature":1.0,"reasoning_tokens":2811,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:02:08.563920+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct re-run of the same protocol (workday-only PeMSD4 and PeMSD8, stations separated by more than 3.5 miles, 60-minute history, 15/30/45-minute targets) that finds any of Graphormer-T, STGCN, or DCRNN matching or beating TSFusion on MAE or RMSE, or an ablation in which removing the gate or the transformer branch leaves the error unchanged, would contradict the central claim.","supporting_citations":[{"cited_title":"Network level spatial temporal traffic forecasting with hierarchical-attention-lstm,","cited_arxiv_id":null,"evidence_quote":"Supplies the PeMSD4 and PeMSD8 datasets and their station/feature structure on which every comparison and robustness test runs."}],"review_version":2}