{"id":"5673cebe-6407-4701-b000-c948bceb6ec6","arxiv_id":"2506.02609","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"TEDDN combines learned time embeddings, a sigmoid disentangle gate, channel attention, and residual dynamic graph convolution to forecast 60-minute traffic flow on four PEMS datasets.","lead":"This paper presents TEDDN, a neural network that splits traffic flow into two learned components and uses a dynamic graph to forecast the next hour of traffic. The authors report lower errors than ten baselines on four California highway datasets, but the paper's own tables do not fully support that claim.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim of 'marked superiority over all benchmark approaches' is contradicted by the paper's own Table 2: on PEMS08, PDFormer has lower MAE and MAPE than TEDDN; on PEMS04, PDFormer has lower MAPE. The stated conclusion is false as written, independent of baseline protocol.","rationale":"I read the paper in good faith as a benchmark-oriented traffic forecasting submission. The architecture is described in reasonable detail, uses standard components, and is evaluated on four public PeMS datasets against eleven baselines. For the central claim to hold, Table 2 would need to show TEDDN as best, or at least not worse, on all metrics against all baselines on all datasets. That condition is not met: the table itself shows PDFormer beating TEDDN on PEMS08 MAE and MAPE, and on PEMS04 MAPE. Since the abstract and Section 5.4 assert unqualified 'marked superiority over all benchmark approaches', the paper's own evidence falsifies the central claim. This is the single most load-bearing concern because it is internal: no assumption about external baseline protocols, hidden code, or seeds is needed. If the table is taken at face value, the conclusion is wrong; if the table is not taken at face value, the paper lacks any reliable evidence for the conclusion. Either way, the central claim is unsupported. The reader's weakest_assumption focused on whether baselines were rerun under the same protocol. That is a real and important concern, especially because several baseline numbers match published official values, suggesting they were copied rather than rerun. However, the direct internal contradiction in Table 2 is even more decisive, which is why I mark agreement as partial rather than full. Table 3 adds further concern: its 'Average' columns contain RMSE/MAPE values identical to Table 2 baseline rows and are inconsistent with the horizon-wise results, so the ablation evidence is also unreliable. No code, seeds, or error bars are provided, so variance and statistical significance cannot be assessed, but the decisive issue does not depend on those. I agree with the reader's REJECT verdict. The method might be salvageable with a corrected evaluation and a more carefully-scoped claim, but as submitted the central assertion is contradicted by the paper's own results.","tokens_in":8924,"tokens_out":6384,"duration_ms":57171,"concrete_test":"Check Table 2's PEMS08 and PEMS04 rows against the abstract's claim: on PEMS08, TEDDN MAE 13.70 > PDFormer 13.58 and TEDDN MAPE 9.17% > PDFormer 9.04%; on PEMS04, TEDDN MAPE 12.17% > PDFormer 12.10%. This arithmetic falsifies 'marked superiority over all benchmark approaches' without further computation. To test the weaker 'best overall' version, rerun both methods under the identical 6:2:2 split, normalization, and evaluation code; if the PEMS08 MAE/MAPE deficits persist, the claim must be revised to 'best on most metric-dataset pairs' or reported with statistical significance. Also recompute Table 3's 'Average' columns from the horizon-wise rows: the reported values do not match the means, and three entries duplicate Table 2 baseline rows, indicating a likely copy-paste error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim, repeated in the abstract, introduction, and Section 5.4, is that TEDDN shows 'marked superiority over all benchmark approaches' across four datasets. The only evidence is Table 2. Reading Table 2 at face value, this is false. On PEMS08, PDFormer reports MAE 13.58 vs TEDDN 13.70 and MAPE 9.04% vs 9.17%, so TEDDN loses on two of three metrics. On PEMS04, PDFormer reports MAPE 12.10% vs TEDDN 12.17%, so TEDDN loses there as well. Thus TEDDN is not superior to all benchmarks on all metrics. If 'superiority' is meant only as MAE/RMSE averages or as 'best overall', the paper never defines this weaker claim, and Section 5.4 states the stronger claim without qualification. This concern is more load-bearing than the reader's protocol concern because it does not depend on whether baselines were rerun locally: the paper's own reported numbers refute the stated conclusion. Additionally, Table 3's 'Average' columns are unreliable: three entries exactly duplicate Table 2 baseline rows (w/o TE RMSE 34.06/MAPE 14.20 = FC-LSTM PEMS08; w/o DG RMSE 26.96/MAPE 11.32 = DSANet PEMS08; w/o GRU RMSE 30.05/MAPE 12.15 = GraphWaveNet PEMS08), and the averages do not match the means of Horizons 3/6/12. These anomalies further undermine the empirical support for the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TEDDN, a traffic flow forecasting model that combines a disentangle gate, a time-enhanced module (CWAM/TE), and residual graph convolution with learnable dynamic graphs. The model is evaluated on four PeMS datasets (PEMS03/04/07/08) for 60-minute-ahead prediction and compared with eleven baselines. The paper claims that TEDDN shows marked superiority over all benchmark approaches and that ablation studies confirm the contribution of the disentangle gate and TE module.","tokens_in":9360,"tokens_out":3686,"duration_ms":33972,"significance":"The disentanglement-and-temporal-enhancement architecture is a plausible idea, and the paper includes a fairly wide comparison (eleven baselines) and four real-world datasets. If the empirical claims were reliable, the contribution could be of interest to the traffic-forecasting community. However, the paper's headline claim of superiority is contradicted by its own reported numbers, and the ablation table contains values that exactly duplicate baseline rows from the main comparison table. These problems bear directly on the central claim and on the credibility of the experimental evidence, so the manuscript is not acceptable in its current form.","major_comments":[{"comment":"The claim that TEDDN shows 'marked superiority over all benchmark approaches' is false as stated when checked against the paper's own Table 2. On PEMS08, PDFormer achieves a lower MAE (13.58 vs. 13.70) and a lower MAPE (9.04% vs. 9.17%) than TEDDN; on PEMS04, PDFormer also achieves a lower MAPE (12.10% vs. 12.17%). Thus TEDDN does not outperform all benchmarks on all metrics on two of the four datasets. If the intended claim is 'best on average' or 'best on most metrics', the paper never defines that weaker claim and Section 5.4 presents the unqualified stronger claim.","section":"Abstract, Section 5.4, Table 2"},{"comment":"The 'Average' columns in Table 3 are internally inconsistent and contain exact duplications of baseline rows from Table 2. For example, the w/o TE row lists Average RMSE 34.06 and MAPE 14.20%, which are exactly the PEMS08 values for FC-LSTM in Table 2; the w/o DG row lists 26.96 and 11.32%, which are exactly the PEMS08 values for DSANet; and the w/o GRU row lists 30.05 and 12.15%, which are exactly the PEMS08 values for GraphWaveNet. Moreover, the stated averages do not match the means of the Horizon 3, 6, and 12 columns (e.g., w/o TE Average MAE is 13.89, while the mean of 13.17, 13.87, and 15.49 is 14.18). This makes the ablation evidence unreliable and undermines the conclusion that the TE module, Disentangle Gate, and GRU layer contribute to performance.","section":"Section 5.5, Table 3"},{"comment":"The baseline comparison is not documented as a controlled experiment. Section 5.2 lists ten baselines only with citations, and Section 5.3 specifies the training settings for TEDDN (6:2:2 split, 60-minute history and horizon, Adam optimizer, etc.) but nowhere states whether the baseline methods were rerun under the same protocol or whether the numbers were taken from previous publications. Different PeMS papers often use different train/validation/test splits (e.g., 7:2:1) and different horizon settings. Since Table 2 is the sole evidence for the paper's central claim, the absence of protocol documentation is a load-bearing omission.","section":"Section 5.2 and Section 5.3"}],"minor_comments":[{"comment":"The column header 'MASE' appears where 'MAE' is intended (e.g., PEMS04 column in Table 2 and the MAE columns in Table 3).","section":"Tables 2 and 3"},{"comment":"The baseline list writes 'DCRNN ,' without a citation, even though DCRNN is described in Section 2.1 with reference [10]; also 'GraphwaveNet' should be 'Graph WaveNet' for consistency with reference [19].","section":"Section 5.2"},{"comment":"Several typos and mechanical errors should be corrected: 'ciyies' in the Fig. 1 caption, 'Insistute' in the affiliation block, 'frature' in Section 4.3, and 'Madeptly' in Section 2.2.","section":"Throughout"},{"comment":"Equation (4) defines Ω_{t,i} with a node index i, but the text in the following paragraph refers to 'node j'; also the dimensional consistency between Ω ∈ R^{T_h×N×1} and the product X ⊙ Ω in Eq. (5) is not explained.","section":"Section 4.3"},{"comment":"The text says 'PeMS datasets (PEMS03-08)', but the experiments use PEMS03, PEMS04, PEMS07, and PEMS08; the naming should be made precise.","section":"Section 5.1"}],"recommendation":"reject","confidential_remarks":"The exact duplication of baseline rows in Table 3 (FC-LSTM, DSANet, and GraphWaveNet PEMS08 values appearing as the 'Average' columns of the ablation variants) is a data-integrity concern that goes beyond ordinary presentation issues. The editors may wish to ask the authors for the raw experimental logs. Even setting that aside, the paper's stated conclusion of universal superiority is contradicted by its own Table 2, so the central claim fails on the evidence presented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the model is a reasonable assembly of known parts—SE-style channel attention, SGE-like time enhancement, a learned sigmoid gate splitting the input, and adaptive residual graph convolution. Nothing fundamentally wrong with the mechanism. The problem is that the paper's central claim, repeated in the abstract and in Section 5.4, that TEDDN shows 'marked superiority over all benchmark approaches', is false on the paper's own numbers. On PEMS08, PDFormer reports MAE 13.58 vs TEDDN 13.70 and MAPE 9.04% vs 9.17%. On PEMS04, PDFormer has lower MAPE (12.10% vs 12.17%). So TEDDN does not beat all benchmarks on all metrics. If the authors meant 'best overall average' or 'best on most metrics,' they need to say so and defend that weaker claim.\n\nThe other red flag is Table 3. The 'Average' columns for the three ablated variants appear to be copied from Table 2 baseline rows: w/o TE shows RMSE 34.06/MAPE 14.20, which is exactly FC-LSTM's PEMS08 row; w/o DG shows 26.96/11.32, exactly DSANet's row; w/o GRU shows 30.05/12.15, exactly GraphWaveNet's row. These are not the means of the H3/H6/H12 columns. I can't rule out a formatting accident, but as presented it looks like copy-paste, and it undermines the ablation evidence—the main support for the disentangle gate and TE module.\n\nWhat the paper does well: the module descriptions are clear, the disentangle gate is a simple and testable idea, and the ablation direction—TE and DG help—is plausible. The processing time table is useful. But there is no code, no seeds, no error bars, so the gains are not independently checkable. The baseline protocol is also uncertain: the paper doesn't state whether baselines were rerun under the same split and normalization, and many PEMS baselines in the literature use different splits.\n\nThis paper is for someone who wants to try the disentangle gate in their own traffic model. It deserves a serious referee because the architecture is sane and the idea is worth checking, but the submission needs major revision: correct the overclaim, fix or redo Table 3, and either rerun baselines under a controlled protocol or qualify the comparison. I'd engage after those fixes.","headline":"Plausible architecture, but the paper's own Table 2 contradicts its 'superiority over all benchmarks' claim and Table 3's average columns look copy-pasted.","tokens_in":9953,"tokens_out":2603,"would_cite":false,"duration_ms":21006,"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 time-enhanced data disentanglement network claims better traffic forecasts by splitting flow into stable patterns and residual trends.","keywords":["traffic flow forecasting","data disentanglement","time-enhanced module","dynamic graph learning","spatiotemporal graph neural networks","PEMS datasets","gated disentanglement","multi-step traffic prediction"],"falsifier":"Re-run FC-LSTM, DSANet, GraphWaveNet, DCRNN, ASTGCN, STFGNN, STGODE, STG-NCDE, DSTAGNN, ST-AE, and PDFormer under TEDDN's exact 6:2:2 split, 12-step horizon, and evaluation code; if their PEMS03/04/07/08 errors drop to or below the reported TEDDN numbers, the central superiority claim fails.","tokens_in":8660,"feed_emoji":"🚦","tokens_out":8284,"duration_ms":80797,"temperature":0.7,"pith_summary":"The paper proposes TEDDN, a traffic-flow forecasting network whose central idea is that visible traffic flow is a mixture of hidden travel modes and that explicitly splitting it improves prediction. A learned, time-conditioned gate separates the input $X$ into a dominant pattern $X_1=X\\odot\\Omega$ and a residual $X_2=X-X_1$, where $\\Omega\\in(0,1)$ is generated from time-slot and node embeddings. A Time-Enhanced Module then sharpens temporal features, and an adaptive residual graph convolution models the resulting streams. On the PEMS03, PEMS04, PEMS07, and PEMS08 datasets with 60-minute history and 60-minute horizon, the authors report aggregate improvements over eleven baseline models, and ablation experiments on PEMS08 attribute the gains to the Time-Enhanced Module, the Disentangle Gate, and the GRU layer.","feed_headline":"Traffic model splits flow into hidden modes to beat 11 rivals","feed_subtitle":"Time-enhanced gate disentangles stable patterns and residuals, and wins on four PEMS datasets at the 60-minute horizon.","key_machinery":"The load-bearing object is the Disentangle Gate, a learned ratio $\\Omega_{t,i}\\in(0,1)$ produced from time-slot embeddings $T^D$, $T^W$ and a node embedding $E_i$; it splits the traffic tensor into $X_1$ (one traffic mode's contribution) and $X_2$ (the unexplained residual), so later layers can model the two streams separately. The Time-Enhanced Module adds a second mechanism: global average pooling gives a descriptor $g$, the dot product $c_i=g\\odot x_i$ gives an importance coefficient, normalization and an affine transform give $a_i$, and features are scaled by $\\mathrm{Sigmoid}(a_i)$. The GC Module completes the architecture with a learnable adjacency matrix $A=\\mathrm{ReLU}(\\tanh(\\alpha(D_{E_1}D_{E_2}^T-D_{E_1}^T D_{E_2})))$ and residual convolution $H^{(k)}=\\beta H_{in}+(1-\\beta)\\tilde{D}^{-1}\\tilde{A}H^{(k-1)}$.","core_discovery":"The paper's claim, stated on its own terms, is that entangled multi-mode traffic data should be decomposed before spatiotemporal modeling, and that temporal information deserves a dedicated enhancement step rather than being learned implicitly. TEDDN computes a per-node, per-time-step gate $\\Omega_{t,i}=\\mathrm{Sigmoid}(\\mathrm{ReLU}(T_t^D\\parallel(T_t^W\\parallel E_i)W_1)W_2)$, multiplies it into the raw flow to isolate one traffic pattern, and treats the subtraction as residual flow. The Time-Enhanced Module derives normalized importance coefficients $a_i=\\gamma\\hat{c}_i+\\beta$ from a global average-pooled descriptor and scales features by $\\mathrm{Sigmoid}(a_i)$, while the GC Module constructs a dynamic non-negative adjacency matrix and applies residual graph convolution. The authors conclude from their tables that this combination yields better 12-step forecasts than adaptive spatiotemporal graph networks and sequence models on the four PEMS datasets, and that each module contributes measurably to the result.","pith_inferences":["Because the gate is conditioned only on time-slot and node embeddings, the paper does not verify that $X_1$ and $X_2$ correspond to physically identifiable modes such as passenger cars versus trucks; identifying what the gate separates is a natural follow-up.","The same split-and-residual architecture could be transferred to other multi-regime time series, such as energy load or crowd flow, to test whether the gain comes from disentanglement itself or from added model capacity.","Applying the Disentangle Gate to a simpler backbone would help isolate whether the empirical improvement is due to the gate or to the combination of stacked modules.","A metric-by-metric reading of the four PEMS tables, rather than an aggregate ranking, is a worthwhile check when comparing TEDDN to its strongest baselines."],"forward_implications":["If the reported gains hold under a controlled comparison, explicit temporal-mode disentanglement becomes a viable alternative to end-to-end spatiotemporal encoders for traffic forecasting.","The PEMS08 ablation shows that removing the Disentangle Gate raises average MAE from 13.70 to 15.06 and removing the Time-Enhanced Module raises it to 13.89, so each component carries predictive weight.","The authors' horizon-wise plots indicate that TEDDN's error grows more slowly than baselines such as STGODE, which matters for look-ahead traffic management rather than one-step prediction.","Table 4 reports inference times under eight seconds on PEMS04 and about three seconds on PEMS08, suggesting the added mechanisms do not make the model impractical for near-real-time use.","The architecture is dataset-agnostic and could be retrained on any sensor network with the same 60-minute history and horizon setup."],"supporting_citations":[{"why":"Supplies the PEMS dataset protocol that the experiments reuse.","marker":"[16]"},{"why":"FC-LSTM baseline representing sequence-only forecasting, which TEDDN must improve upon.","marker":"[17]"},{"why":"DCRNN baseline, a widely used spatiotemporal benchmark for traffic forecasting.","marker":"[10]"},{"why":"GraphWaveNet baseline whose adaptive graph ideas motivate TEDDN's graph learning.","marker":"[19]"},{"why":"ASTGCN baseline that combines attention with spatiotemporal convolution.","marker":"[20]"},{"why":"STFGNN baseline that builds data-driven graphs using dynamic time warping.","marker":"[21]"},{"why":"STGODE baseline using graph neural ordinary differential equations.","marker":"[22]"},{"why":"STG-NCDE baseline using neural controlled differential equations for traffic.","marker":"[23]"},{"why":"DSTAGNN baseline that models dynamic spatial-temporal dependencies.","marker":"[24]"},{"why":"PDFormer baseline, the strongest competitor in the comparison and the one TEDDN is most directly measured against.","marker":"[26]"}],"fun_headline_variants":["Disentangle traffic modes to forecast better","Time-enhanced splitting of traffic flow boosts accuracy","TEDDN separates stable patterns and residuals for traffic","Decompose traffic data before spatiotemporal modeling","New traffic forecast model wins by separating patterns"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison in Table 2 assumes the eleven baseline results were obtained under the same data split, history and horizon length, normalization, and evaluation code as TEDDN, but the paper does not state whether the baselines were rerun locally.","fun_headline_variants_meta":{"raw":{"variants":["Disentangle traffic modes to forecast better","Time-enhanced splitting of traffic flow boosts accuracy","TEDDN separates stable patterns and residuals for traffic","Decompose traffic data before spatiotemporal modeling","New traffic forecast model wins by separating patterns"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000187,"raw_usage":{"total_tokens":1316,"prompt_tokens":921,"completion_tokens":395,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":326}},"tokens_in":537,"tokens_out":395,"duration_ms":4481,"temperature":1.0,"reasoning_tokens":326,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:20:36.601193+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run FC-LSTM, DSANet, GraphWaveNet, DCRNN, ASTGCN, STFGNN, STGODE, STG-NCDE, DSTAGNN, ST-AE, and PDFormer under TEDDN's exact 6:2:2 split, 12-step horizon, and evaluation code; if their PEMS03/04/07/08 errors drop to or below the reported TEDDN numbers, the central superiority claim fails.","supporting_citations":[{"cited_title":"In: Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining","cited_arxiv_id":null,"evidence_quote":"Supplies the PEMS dataset protocol that the experiments reuse."},{"cited_title":"Ieee Access 9, 35973–35983 (2021)","cited_arxiv_id":null,"evidence_quote":"ASTGCN baseline that combines attention with spatiotemporal convolution."},{"cited_title":"In: Proceedings of the AAAI conference on artificial intelligence","cited_arxiv_id":null,"evidence_quote":"STFGNN baseline that builds data-driven graphs using dynamic time warping."},{"cited_title":"In: Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining","cited_arxiv_id":null,"evidence_quote":"STGODE baseline using graph neural ordinary differential equations."},{"cited_title":"In: Proceedings of the AAAI conference on artificial intelligence","cited_arxiv_id":null,"evidence_quote":"STG-NCDE baseline using neural controlled differential equations for traffic."},{"cited_title":"In: International conference on ma- chine learning","cited_arxiv_id":null,"evidence_quote":"DSTAGNN baseline that models dynamic spatial-temporal dependencies."},{"cited_title":"In: Proceedings of the AAAI conference on artificial intelligence","cited_arxiv_id":null,"evidence_quote":"PDFormer baseline, the strongest competitor in the comparison and the one TEDDN is most directly measured against."}],"review_version":1}