{"id":"4fd41587-d30b-4038-a78f-e07a9e380437","arxiv_id":"2501.00756","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A low-rank fast graph computation plus a dynamic spatio-temporal convolution kernel gives FasterSTS competitive traffic forecasts at a fraction of the compute, though not the universal state-of-the-art the paper claims.","lead":"FasterSTS is a traffic-forecasting graph network that replaces the usual N-by-N spatial transform with a low-rank N-by-n step, cutting computational cost. The paper reports large speedups and strong accuracy on four highway datasets, but its claim of beating all baselines is contradicted by its own tables.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 'significantly leads all baselines on all four datasets' is contradicted by Table IV, where STWave beats FasterSTS on PEMS07 in MAE and MAPE.","rationale":"The reader's weakest assumption concerned baseline protocol comparability. I agree that is a serious issue, but the more decisive problem is internal: the paper's own Table IV contradicts the strongest claim. This is not a disagreement with external consensus; it is a mismatch between the stated conclusion and the reported evidence. The central claim of universal superiority on all metrics and datasets cannot survive without discarding or correcting Table IV. The architecture also has specification gaps (Eqs. 3, 6, 13 use shape-incompatible Hadamard products), but the empirical contradiction alone is sufficient to block the claim as stated. The efficiency results are the paper's strongest asset; because they are reported for a single dataset and without code, they are not enough to carry the paper. My verdict does not move from the reader's rejection; the contradiction reinforces it.","tokens_in":17979,"tokens_out":4383,"duration_ms":41166,"concrete_test":"Recompute the PEMS07 average row in Table IV from the 15/30/60 min columns, and check the STWave and FasterSTS entries for internal consistency. Then rerun FasterSTS and the published STWave model on PEMS07 under a single protocol (6:2:2 split, 12-step history, identical normalization and evaluation). If STWave reproduces MAE below 19.67 and MAPE below 8.30, the 'all four datasets / all three indexes' claim is false; if the table entries are arithmetically wrong or unreproducible, the paper needs corrected tables and shared code before the claim can be assessed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Taking the paper's reported results at face value, the central empirical claim fails. Table IV (PEMS07, average row) lists STWave with MAE 19.60, MAPE 8.21%, RMSE 33.29 and FasterSTS with MAE 19.67, MAPE 8.30%, RMSE 32.75. STWave is better on two of three metrics on one of four datasets, directly contradicting Section IV-D's assertion that FasterSTS 'significantly leads all baseline models in three evaluation indexes on all four datasets.' If the table entries are accurate, the claim is false as stated. If they are not accurate, the paper still has not established the claim, because no shared-protocol reruns, code, seeds, or repeated-run statistics are provided. Either way, the load-bearing empirical claim is unsupported. The efficiency evidence in Table VI is independently plausible and could survive a correction, but the prediction-performance claim must be narrowed or re-established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FasterSTS, a spatio-temporal synchronous graph convolutional network for traffic flow forecasting. The core ideas are (1) a 'fast graph computation' that replaces the N×N adaptive adjacency with an N×n low-rank aggregation followed by a 1×1 convolution projection, reducing graph convolution cost from O(N²) to O(nN); (2) a spatio-temporal synchronous convolution kernel Ψ built from static learnable embeddings and a data-dependent dynamic part, intended to capture temporal correlations during graph convolution; and (3) per-hidden-dimension adaptive graphs using global and local node embeddings. The model is evaluated on PEMS03/04/07/08 and compared with eight baselines on MAE, MAPE, and RMSE, plus an efficiency comparison in Table VI. The paper claims FasterSTS 'significantly leads all baseline models in three evaluation indexes on all four datasets' (Section IV-D).","tokens_in":18100,"tokens_out":3948,"duration_ms":37587,"significance":"If substantiated, the model would be a meaningful practical contribution: it promises state-of-the-art forecasting accuracy at substantially lower training and inference cost, which is relevant for real-time and resource-constrained applications. The low-rank graph aggregation idea is intuitively appealing and the efficiency numbers in Table VI (e.g., 37 s/epoch and 1759 MB GPU memory versus 256 s and 6789 MB for STGODE on PEMS07) are suggestive. However, the significance is heavily undercut by the fact that the paper's own Tables III–IV contradict the headline empirical claim, and by the under-specified formal definition of the architecture. The efficiency advantage alone, without a reliable accuracy claim, is not sufficient to establish the paper's central thesis.","major_comments":[{"comment":"The central claim that FasterSTS 'significantly leads all baseline models in three evaluation indexes on all four datasets' is directly contradicted by the paper's own Table IV for PEMS07. On the average row, STWave achieves MAE 19.60 and MAPE 8.21%, while FasterSTS achieves MAE 19.67 and MAPE 8.30%; STWave is also better on MAE and MAPE at the 15-, 30-, and 60-minute horizons. Only RMSE favors FasterSTS. Thus the stated claim is false as written. If the table entries are accurate, the conclusion must be narrowed; if they are not accurate, the paper has not supported the claim with corrected data.","section":"IV-D, Table IV"},{"comment":"The formal specification of the spatio-temporal synchronous convolution kernel is internally inconsistent and cannot be reconstructed from the text. The kernel Ψ is first defined as Ψ ∈ R^{T din × T dout}, then later as Ψ ∈ R^{T din × dΨ}, with dΨ said to be much less than T dout; Eq. (6) writes X ∈ R^{N × dΨ} immediately after stating X ∈ R^{N × T × din}. The fast graph computation in Eq. (3) defines A ∈ R^{N × n} via A(i,j)=[Softmax(E_{:,j})]_i, but Eq. (13) applies A ⊙ X with X of shape N × T × C, which is a dimension mismatch unless additional reshaping is assumed and never stated. The 'dimension transformation' R() is described only verbally, and the sizes of CΨ and cΨ are not given. These omissions mean the proposed architecture is not precisely defined.","section":"III-E, Eqs. (6)–(9)"},{"comment":"The comparison against baselines is not established under a shared protocol. Section IV-B states the 6:2:2 split for FasterSTS, but the paper does not state that the baseline metrics in Tables III–IV were obtained by rerunning those models under the identical preprocessing, input length (12 steps), normalization, evaluation horizon, or random seed protocol. The baseline numbers appear to be imported from prior publications (e.g., the STWave values match its ICDE paper). Without shared-protocol reruns or a statement that all baselines use the same pipeline, the claim of 'significantly leads all baselines' is unsupported even setting aside the PEMS07 counterexample.","section":"IV-B and IV-C, Tables III–IV"},{"comment":"The complexity analysis is not rigorous. The paper claims the spatio-temporal kernel cost is reduced 'from T din × T dout' to 'T dindΨ + T doutdΨ', but the parenthetical 'T din × T dout is equal to T din T 2 dout + T dout T 2 din' is syntactically meaningless, and no derivation is given for the claimed costs. The analysis also omits the costs of the dynamic kernel generation steps (Eqs. (7)–(9)), the Softmax, the gating operation, and the per-hidden-dimension adaptive graph construction. The O(N²)-to-O(nN) reduction for graph aggregation is plausible, but the overall claim that the model is 'reduced from quadratic to linear level' is not demonstrated.","section":"IV-H"},{"comment":"The ablation results do not support the statement that the fast graph operation 'achieves a similar effect to traditional graph operations'. In Table V, the variant without fast graph computation (w/o FGC) is better than FasterSTS on all three metrics on PEMS04 (MAE 18.40 vs. 18.49, RMSE 29.90 vs. 29.92, MAPE 12.16 vs. 12.21). The paper's claim of parity is contradicted by its own numbers, and this weakens the argument that the efficiency gain is obtained without sacrificing accuracy.","section":"IV-E, Table V"}],"minor_comments":[{"comment":"The algorithm says 'the parameters are updated using gradient descent to minimize the Eq.11', but Eq. (11) is a residual connection, not a loss function; the loss is defined in Eq. (17). The reference should be to Eq. (17).","section":"Algorithm 1, line 15"},{"comment":"The 'Reinforcement Learning Component' is described as two fully connected layers with ReLU; no reinforcement learning signal, policy, or reward is involved. The name is misleading and should be changed to something like 'temporal embedding projection'.","section":"III-E, 'Reinforcement Learning Component'"},{"comment":"Two consecutive paragraphs in Section II-C are duplicated verbatim, beginning 'Compared with the spatio-temporal asynchronous modeling methods mentioned above...'. One copy should be removed.","section":"II-C"},{"comment":"References [35] and [39] are identical (both cite Chen et al., 'Bidirectional spatial-temporal adaptive transformer'); [35] is used for Bi-STAT in the tables and [39] is also cited for Bi-STAT. Duplicate references should be consolidated.","section":"References"},{"comment":"There are numerous typos and inconsistent notations: 'Sof tmax' for Softmax, 'caputred' for captured, 'Traning' and 'Interferce' in Table VI, 'a 1×1 convolutional layer (whose time complexity is n ∗ N)' which conflates the convolution kernel with the complexity, and the undefined symbol K in the abstract's complexity claim (O(KN)).","section":"Throughout"},{"comment":"The hyperparameter analysis refers to Fig. 5, but no figure content is described and no quantitative results are given; the text only states generic overfitting behavior. Please provide actual plots or tables with the sensitivity values.","section":"IV-F"}],"recommendation":"reject","confidential_remarks":"The paper's central empirical claim is refuted by its own Table IV on PEMS07, and the architecture is under-specified to the point of non-reproducibility. These are load-bearing problems that cannot be fixed by a local revision. The low-rank graph computation idea is worth pursuing, and a future version that reruns all baselines under a common protocol, corrects the dimensional inconsistencies, and narrows the performance claim could be a viable submission. I would also note for the editor that the paper's related-work discussion leans heavily on the authors' own HTVGNN preprint [12], and the incremental novelty over that work is not sharply delineated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has one solid engineering idea buried under an overclaimed empirical headline and equations that don't type-check. The efficiency result is worth a look; the prediction claim, as stated, is false by the paper's own tables.\n\nWhat's new and good: the fast graph computation is a low-rank factorization of adjacency aggregation—use an N×n adaptive matrix instead of N×N, then project back with a 1×1 convolution. Low-rank tricks are standard, but applying this to a spatio-temporal synchronous convolution kernel is a reasonable combination, and Table VI shows large speedups: 37 s/epoch and 1.76 GB GPU memory on PEMS07 versus 452 s and 9.3 GB for STWave, and a clear gap over the non-fast variant w/o FGC. The ablation is honest here: w/o FGC gets slightly better MAE (18.40 vs 18.49 on PEMS04), so the fast version trades a hair of accuracy for a big efficiency gain. That is a real, useful trade-off.\n\nSoft spots: the Section IV-D claim that FasterSTS \"significantly leads all baselines in three evaluation indexes on all four datasets\" is contradicted by Table IV. On PEMS07, STWave has lower MAE (19.60 vs 19.67) and lower MAPE (8.21% vs 8.30%); on PEMS08 at 15 min, STWave matches or beats it in MAE. So the universality claim is simply false as written. The equations are under-specified: Eq. (3) applies a Hadamard product between an N×n matrix and a T×N×C tensor, which is undefined, and similar issues appear in Eqs. (6) and (13). I could not reconstruct the forward pass exactly. The \"RL Component\" is two fully connected layers with ReLU; calling it reinforcement learning is misleading. No code, seeds, or repeated-run statistics are provided, and the baseline numbers are imported from prior papers without a shared protocol, so even corrected comparisons rest on the original sources.\n\nThat said, the efficiency contribution is plausible and the ablation suggests the core mechanism works. This is not a forget-it paper; it needs a serious referee to push for fixed equations, narrowed claims, rerun baselines, and released code. The low-rank synchronous convolution could be useful for edge deployment.\n\nRecommendation: send to peer review, but expect major revision. I would not cite it yet.","headline":"The efficiency claim is real and worth engaging, but the paper's headline prediction claim is contradicted by its own Table IV and the equations cannot be reconstructed.","tokens_in":18737,"tokens_out":2741,"would_cite":false,"duration_ms":26063,"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":"FasterSTS claims state-of-the-art traffic forecasts while cutting graph convolution cost from quadratic to linear.","keywords":["traffic flow forecasting","spatio-temporal synchronous modeling","graph convolutional networks","fast graph computation","adaptive graph","low-rank node aggregation","PEMS datasets"],"falsifier":"Rerun the ten baselines on the same machine, with the same data splits, normalization, and evaluation script used for FasterSTS, and recompute Tables III and IV; if any baseline beats FasterSTS's reported average MAE, MAPE, or RMSE on any dataset, the 'leads all baselines' claim is refuted. A cheaper check is already available in the paper: in Table IV, the STWave row has lower average MAE and MAPE than FasterSTS on PEMS07.","tokens_in":17640,"feed_emoji":"🚦","tokens_out":10601,"duration_ms":90292,"temperature":0.7,"pith_summary":"FasterSTS is a traffic-flow forecasting model whose central claim is that spatio-temporal correlations can be modeled synchronously, inside a single graph-convolution operation, rather than by separate spatial and temporal blocks. The paper introduces a fast graph computation that replaces the dense $N \\times N$ adjacency operation with an $N \\times n$ aggregation step followed by an $n \\times N$ projection, cutting the graph-convolution cost from $O(N^2)$ to $O(nN)$. On four public California highway datasets (PEMS03, PEMS04, PEMS07, PEMS08), the paper claims lower MAE, MAPE, and RMSE than ten baseline models, while using no multi-head self-attention, no recurrent neural network, and no temporal convolutional network. If those comparisons hold under a common evaluation protocol, accurate spatio-temporal forecasting would become practical on resource-limited devices and inside real-time loops such as reinforcement-learning controllers.","feed_headline":"Traffic forecast model cuts graph cost from quadratic to linear","feed_subtitle":"Claims state-of-the-art accuracy on four highway datasets while dropping attention, RNNs, and TCNs.","key_machinery":"The load-bearing object is the fast graph computation, which splits the conventional graph operation into node information aggregation, a trainable matrix $E \\in \\mathbb{R}^{N \\times n}$ with $n \\ll N$, applied as a softmax-weighted Hadamard product, and node representation projection, a $1\\times1$ convolution that lifts the $n$-dimensional aggregated representation back to $N$ dimensions. This reduces graph computation from $O(N^2)$ to $O(nN)$. The second central object is the spatio-temporal synchronous graph convolution kernel $\\Psi$, built from a static learnable embedding and a dynamic input-derived embedding, which lets graph convolution itself carry temporal correlation instead of delegating it to RNNs, TCNs, or attention. The adaptive graph is assembled from global and local node embeddings, one per hidden dimension, so the dynamic kernel can gate and adjust spatial features during convolution.","core_discovery":"On its own terms, the paper's discovery is that the expensive parts of a spatio-temporal synchronous graph convolutional network can be replaced without giving up accuracy: the graph aggregation is factored into two low-rank steps, and the convolution kernel is made time-aware by splitting it into a static adaptive embedding and a dynamic adaptive embedding whose combination forms a spatio-temporal synchronous kernel $\\Psi$. The static part models stable temporal dependencies and feature mapping; the dynamic part, derived from the input through a $1\\times1$ convolution and a softmax, adjusts the kernel and the adaptive graph over time. The authors argue that this lets one graph-convolution layer capture static and dynamic spatial correlation, static and dynamic temporal correlation, and the heterogeneity among them. They claim that this design outperforms ten prior models in three error metrics on four real-world traffic datasets while reducing training time and GPU memory several-fold relative to the spatio-temporal synchronous baselines.","pith_inferences":["Editorial inference: the low-rank aggregation pattern (project to $n$ dimensions, aggregate, project back) is not specific to traffic data, so it could be dropped into other learned-adjacency graph convolutions where $N$ is large, such as recommender systems or molecular graphs.","Editorial observation: the paper's own Table IV shows STWave with lower average MAE (19.60 vs 19.67) and MAPE (8.21 vs 8.30) on PEMS07, so the sentence claiming that FasterSTS leads all baselines in all three metrics on all four datasets is not supported by the reported numbers; the supportable claim is that FasterSTS is competitive with STWave and better in RMSE.","Editorial inference: the static-plus-dynamic kernel factorization suggests a separable design principle, use a low-cost static kernel to anchor stable patterns and a data-dependent kernel to modulate them, which could be tested on other spatio-temporal tasks such as energy load or crowd-flow forecasting."],"forward_implications":["If the reported numbers survive a shared evaluation protocol, FasterSTS would be a spatio-temporal synchronous model that matches attention-based and wavelet-based transformers on the PEMS datasets without using attention at all.","The reduction from $O(N^2)$ to $O(nN)$ would let the same architecture scale to city-scale networks with tens of thousands of sensors on a single GPU.","Because the model uses only embeddings, convolutions, and residual normalizations, its forward pass is simpler to compile for edge devices and fast enough to sit inside reinforcement-learning loops.","The ablation study indicates that the fast graph operation is mostly a speed device: removing it leaves MAE essentially unchanged, so the accuracy gains are attributed to the synchronous kernel and the adaptive graphs rather than to the low-rank factorization."],"supporting_citations":[{"why":"Defines STSGCN, the spatio-temporal synchronous framework whose fused local spatio-temporal graph the paper aims to make faster and higher-quality.","marker":"[23]"},{"why":"Provides STFGNN, a synchronous baseline whose spatio-temporal fusion graph construction the paper identifies as high-complexity and limited.","marker":"[24]"},{"why":"Provides STGODE, the ODE-based synchronous baseline used for accuracy comparison and the main complexity contrast in Table VI.","marker":"[33]"},{"why":"Introduces graph wavenet, source of the adaptive-graph idea and a baseline whose static adaptive graph is compared with the model's dynamic adaptive graphs.","marker":"[18]"},{"why":"The hybrid time-varying graph neural network whose adaptive embedding construction inspires the static part of the spatio-temporal synchronous kernel.","marker":"[12]"},{"why":"The Bi-STAT transformer baseline whose strong average errors define the state-of-the-art level the paper claims to exceed.","marker":"[35]"},{"why":"The STWave baseline that decomposes traffic into trend and event channels; its results are the closest competitor in Tables III and IV.","marker":"[38]"},{"why":"Supplies the first-order Chebyshev approximation used as the base graph-convolution operation that the paper reforms into the time-aware kernel.","marker":"[34]"}],"fun_headline_variants":["FasterSTS: low-rank graph cuts traffic forecast cost","Speedy spatio-temporal graph net for traffic flow","Low-rank kernel speeds traffic forecasting without accuracy loss","FasterSTS: efficient spatio-temporal sync graph conv for traffic","Time-aware splits make graph traffic forecast faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every baseline number in Tables III and IV was produced under exactly the same protocol as FasterSTS, the same 6:2:2 data split, the same 12-step history, the same normalization, and the same evaluation code, because the paper takes those baseline numbers from earlier papers rather than rerunning the models in its own environment.","fun_headline_variants_meta":{"raw":{"variants":["FasterSTS: low-rank graph cuts traffic forecast cost","Speedy spatio-temporal graph net for traffic flow","Low-rank kernel speeds traffic forecasting without accuracy loss","FasterSTS: efficient spatio-temporal sync graph conv for traffic","Time-aware splits make graph traffic forecast faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000568,"raw_usage":{"total_tokens":2619,"prompt_tokens":807,"completion_tokens":1812,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":423,"completion_tokens_details":{"reasoning_tokens":1735}},"tokens_in":423,"tokens_out":1812,"duration_ms":12717,"temperature":1.0,"reasoning_tokens":1735,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:43:11.397351+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the ten baselines on the same machine, with the same data splits, normalization, and evaluation script used for FasterSTS, and recompute Tables III and IV; if any baseline beats FasterSTS's reported average MAE, MAPE, or RMSE on any dataset, the 'leads all baselines' claim is refuted. A cheaper check is already available in the paper: in Table IV, the STWave row has lower average MAE and MAPE than FasterSTS on PEMS07.","supporting_citations":[{"cited_title":"Spatial- temporal synchronous graph convolutional networks: A new framework for spatial-temporal network data fore- casting,","cited_arxiv_id":null,"evidence_quote":"Defines STSGCN, the spatio-temporal synchronous framework whose fused local spatio-temporal graph the paper aims to make faster and higher-quality."},{"cited_title":"Spatial-temporal fusion graph neural networks for traffic flow forecasting,","cited_arxiv_id":null,"evidence_quote":"Provides STFGNN, a synchronous baseline whose spatio-temporal fusion graph construction the paper identifies as high-complexity and limited."},{"cited_title":"Spatial- temporal graph ode networks for traffic flow forecasting,","cited_arxiv_id":null,"evidence_quote":"Provides STGODE, the ODE-based synchronous baseline used for accuracy comparison and the main complexity contrast in Table VI."},{"cited_title":"Bidirectional spatial-temporal adaptive transformer for urban traffic flow forecasting,","cited_arxiv_id":null,"evidence_quote":"The Bi-STAT transformer baseline whose strong average errors define the state-of-the-art level the paper claims to exceed."},{"cited_title":"When spatio-temporal meet wavelets: Dis- entangled traffic forecasting via efficient spectral graph attention networks,","cited_arxiv_id":null,"evidence_quote":"The STWave baseline that decomposes traffic into trend and event channels; its results are the closest competitor in Tables III and IV."}],"review_version":1}