{"id":"c5e7ec76-ce95-424b-b05f-a69a3cb8212b","arxiv_id":"2412.02942","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"STDCformer adds learned spatial and temporal confounder weights and cross-time attention to a transformer, reporting small accuracy gains in crowd flow prediction on two New York taxi datasets.","lead":"This paper proposes STDCformer, a transformer for crowd flow prediction that learns separate spatial and temporal confounder weights and uses cross-time attention between past and future. The authors report small accuracy gains on two New York taxi datasets and a larger gain in zero-shot transfer to a different borough.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 2 is not the standard backdoor adjustment even when S and T are the only confounders; the paper provides no identification argument, so the central de-confounding causal claim is unsupported.","rationale":"The paper's central claim, as stated in the abstract and Section 3.1.1, is that the proposed strategy learns a de-confounded representation space and estimates the causal effect of past on future via a spatial-temporal backdoor adjustment. The load-bearing step is therefore identification: the paper must show that the quantity being optimized is actually P(Y|do(X)). It does not. Equation 2 is not the standard backdoor formula; it is a sum of two terms involving two different confounders, not a stratification over their joint values, and it is not well-defined notationally. Equation 6 further replaces the terms with sigmoid-weighted attention outputs, which is a standard gated-fusion operation without any proof that it corresponds to the backdoor-adjusted estimand. This is more fundamental than the concern about unobserved confounders: even in a fully observed toy model where S and T are the only confounders, Eq. 2 would generally not equal the true interventional effect. The empirical sections (Tables 3 and 4, ablation in Section 4.5) provide some evidence that the model performs reasonably as a predictor and that the auxiliary inputs and gating help, but these results do not rescue the causal claim because the causal estimand is never derived. The absence of released code and datasets (contribution 4 promises datasets but gives no access link) further hampers independent verification. I therefore agree with the reader's REJECT verdict, while acknowledging that a revised manuscript that reframes STDCformer as a gated-fusion transformer with auxiliary feature conditioning, without causal vocabulary, could be a legitimate engineering contribution. My concern overlaps with the reader's weakest assumption but is more basic: the causal formula itself is invalid, not merely vulnerable to unobserved confounding.","tokens_in":25916,"tokens_out":5026,"duration_ms":50706,"concrete_test":"Analytically instantiate the causal graph in Fig. 6 with binary S and T as the only confounders via structural equations such as X = S + T + ε and Y = X + S + T + δ. Compute the true interventional mean E[Y|do(X=x)] from the structural equations. Then compute the expression implied by Eq. 2 after estimating P(Y|X, CS) and P(Y|X, CT) from the observational joint distribution, and compare the two quantities. If they differ for any parameter setting, Eq. 2 is not a valid backdoor adjustment, which refutes the claim that STDCformer estimates the de-confounded causal effect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline contribution is the Spatial-Temporal Backdoor Adjustment strategy (Section 3.1.1) that purportedly estimates P(Y|do(X)) and learns a de-confounded representation space. The standard backdoor adjustment formula is P(Y|do(X)) = Σ_c P(Y|X, C=c) P(C=c) (or the corresponding integral for continuous C). Equation 2 writes instead P(Y|do(X)) = P(Y|X, CS) P(CS) + P(Y|X, CT) P(CT). This is not a special case of the standard formula: CS and CT are two distinct confounders, not a single categorical variable, and the expression omits conditioning on the joint C = (CS, CT) while summing marginal terms rather than strata. As written, 'P(Y|X, C=CS)' is also undefined because it sets a random vector C equal to a random variable CS. The implementation in Eq. 6 replaces these terms by sigmoid-weighted spatial and temporal attention outputs, i.e., a standard gated fusion: H = P(CS)·STR'_S + (1−P(CS))·STR'_T. No theorem is stated to show that the estimand in Eq. 2 equals the interventional distribution for the graph in Fig. 6, nor that the learned weights equal P(CS) and P(CT). Thus, even granting the paper's assumption that S and T are the only confounders and that the auxiliary features fully represent them, the causal identification step does not go through. The architecture is a gated-fusion transformer with auxiliary feature conditioning; its causal interpretation is not derived. This directly threatens the abstract's claim of 'estimating the de-confounding causal effect of historical data on future data.'","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"STDCformer proposes a transformer architecture for crowd flow prediction in which the past-to-future mapping is decomposed into encoding, cross-time mapping, and decoding. The paper's headline contribution is a 'Spatial-Temporal Backdoor Adjustment' strategy that is claimed to estimate the causal effect P(Future | do(Past)) by categorizing confounders into temporal and spatial types, and to learn a de-confounded representation space. The architecture adds spatial and temporal confounder embeddings, a spatial-temporal de-confounded fusion gating, and a cross-time attention module. The authors report state-of-the-art or competitive IID results on two new NYC taxi datasets and better zero-shot OOD transfer than two transformer baselines, along with ablations, hyperparameter analyses, and qualitative interpretations of learned weights.","tokens_in":26236,"tokens_out":3172,"duration_ms":30145,"significance":"If the causal claims were valid, the paper would offer a principled way to bring backdoor adjustment into spatiotemporal forecasting and would contribute two new publicly described benchmarks with auxiliary covariates. The empirical scaffolding is substantial: two real-world datasets, a Transformer architecture, ablations, and a zero-shot transfer experiment are all presented. However, the central causal derivation is not sound. Equation (2) is not the standard backdoor adjustment, and Equation (6) implements what is effectively a learned gated fusion rather than a causal estimator. No identification conditions are stated, and the interpretation of the learned weights in Section 4.6.3 is circular. The empirical gains are modest and not consistent across all metrics, so the paper's contribution rests on the causal framing, which is not supported.","major_comments":[{"comment":"The proposed 'spatial-temporal backdoor adjustment' is not the standard backdoor adjustment formula. For a confounder set C, the backdoor adjustment is P(Y|do(X)) = Σ_c P(Y|X, C=c)P(C=c) when C is discrete. Equation (2) instead writes P(Y|do(X)) = P(Y|X, C=C_S)P(C=C_S) + P(Y|X, C=C_T)P(C=C_T). Even if S and T are the only confounders, the adjustment should condition on the joint value (S=s, T=t) and sum or integrate over the joint strata; adding two marginal terms with the notation 'C=C_S' is not a special case of the formula. The paper provides no identification theorem showing that the right-hand side equals the interventional distribution for the graph in Figure 6. The central claim of the abstract, that the model 'estimates the de-confounding causal effect of historical data on future data,' is therefore unsupported.","section":"Section 3.1.1, Eq. (2)"},{"comment":"The implemented fusion in Eq. (6) is a sigmoid-gated convex combination of spatial and temporal attention outputs: H = P(C_S)·STR'_S + P(C_T)·STR'_T, with P(C_T)=1−P(C_S). No argument is given that the sigmoid of a summed embedding equals P(C=C_S), nor that the attention outputs STR'_S and STR'_T equal the conditional outcome terms P(Y|X,C=C_S) and P(Y|X,C=C_T). The text asserts this correspondence without derivation. In practice this is standard gated fusion of two representation streams conditioned on auxiliary features; calling it a causal effect estimator is not justified.","section":"Section 3.3.2, Eq. (6)"},{"comment":"The interpretation of learned P(C_S) weights as evidence of successful de-confounding is circular. The weights are deterministic functions of the learned confounder embeddings through the Sigmoid in Eq. (6) and are trained end-to-end to minimize MAE. The manuscript then uses these same weights to conclude that the model compensates for 'the confounding bias in the original data' and that 'the de-confounding process adjusts this bias.' No external ground truth, no causal validation experiment, and no comparison of the learned weights to known confounding distributions are provided. The zero-shot OOD result is the only independent evidence, but it does not by itself establish that the adjustment identifies a causal effect rather than improving shift robustness through feature conditioning.","section":"Section 4.6.3"},{"comment":"The assumed causal graph is not defended. The paper states in Section 3.1 that confounders are often unobservable, yet the method rests on the assumption that the measured auxiliary features (Table 2) exactly represent the complete set of confounders and that no unobserved confounders exist. No argument is given that the backdoor criterion is satisfied for the graph in Figure 6, nor that the proxy features (e.g., POI counts, crime statistics) adequately capture the proposed constructs. As written, the model's causal interpretation collapses if unobserved confounders exist or if the proxies are incomplete, and the manuscript does not address this threat.","section":"Section 3.1.1 and Figure 6"}],"minor_comments":[{"comment":"In Eq. (3), the temporal confounder is written as C_T = Conv_{s→d}(T), but T ∈ R^{1×t}; the subscript appears to be a typo and should be Conv_{t→d}(T).","section":"Section 3.3.1, Eq. (3)"},{"comment":"The dimension of the attention map A_MAP is stated as R^{n×T_p×T_f}; with Q from the future STE and K from the past STE, the attention matrix should have shape R^{T_f×n×T_p} (or a flattened variant). Please correct the dimension statement.","section":"Section 3.4, Eq. (7)"},{"comment":"Table 4 reports MAPE, but the metric is not defined in the text and the baseline comparisons in Table 3 use only MAE and RMSE; please define MAPE and state how it is computed for aggregate flow.","section":"Section 4.4.2, Table 4"},{"comment":"Several in-text citations appear to be mismatched with the reference list. For example, the citation [23] in Section 2.2.1 for self-supervised deconfounding does not match the listed 'STG4Traffic' reference, and [30] is cited as node2vec in the text but appears as MTGNN in the list. A careful renumbering and cross-check of all citations is needed.","section":"References"},{"comment":"The manuscript contains frequent typos and formatting artifacts, including 'di fferent', 'wights', 'f uture', and inconsistent spacing in equations and citations. A thorough proofread is recommended before resubmission.","section":"Figures and text"},{"comment":"The claimed state-of-the-art result is not uniform: PDFormer achieves a lower MAE than STDCformer for inflow on MHT (14.64 vs. 14.81). The text acknowledges this candidly, but the contribution statement in Section 1 should be made consistent with the full set of results.","section":"Section 4.4.1, Table 3"}],"recommendation":"reject","confidential_remarks":"The paper has useful empirical components, particularly the new datasets and the zero-shot transfer experiment, but the central causal contribution is not defensible as written. The error in Eq. (2) is not a local slip; it undermines the abstract's claim of estimating a causal effect. If the authors were to remove the causal framing and present STDCformer as an empirically motivated gated-fusion transformer with auxiliary feature conditioning, the paper might be resubmitted as a solid applied contribution, but that would be a substantially different manuscript. I would advise the editor that the current version should not be accepted, and that a revision path would require either a correct identification argument with explicit assumptions and external validation of the de-confounding claim, or a complete reframing away from causal effect estimation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful core here is a cleanly engineered transformer for crowd flow: an encoder-decoder with an explicit past-to-future mapping module, spatial and temporal attention, a learned gate to fuse them, and cross-time attention between future and past STE embeddings. Ablations show each piece matters, and the zero-shot MHT-to-BKL transfer is a genuine OOD check. The two new datasets with auxiliary covariates are also a real resource.\n\nThe problem is the packaging. The headline contribution is a 'Spatial-Temporal Backdoor Adjustment' that estimates P(Y|do(X)), but Eq. 2 is not the backdoor adjustment formula. The standard formula sums over strata of a single confounder C: Σ_c P(Y|X,c)P(c). Eq. 2 writes P(Y|X,C_S)P(C_S) + P(Y|X,C_T)P(C_T) with C = {S,T} as a vector; that is not a special case, and 'C = C_S' is type-incoherent. No identification theorem is stated for the graph in Fig. 6. The implementation in Eq. 6 replaces those terms with a sigmoid-weighted sum of spatial and temporal attention outputs, so the de-confounding story reduces to a learned gated fusion, a variant of what GMAN and PDFormer already do. The reading of learned weights in Section 4.6.3 as evidence of successful de-confounding is post hoc.\n\nThis matters because the causal claim is load-bearing for the title, abstract, and contributions. If the paper were reframed as a gated-fusion transformer with auxiliary feature conditioning, it would be a modest but valid engineering contribution: IID gains over PDFormer are 0.09 MAE on MHT and 0.04 on BKL, no error bars, and the zero-shot gains are larger. No code or data are released, which compounds the evidential weakness.\n\nIs it worth a referee? Yes. The architecture is reproducible in principle, the ablations are sensible, and the gap between causal language and gated attention is exactly what a referee should force the authors to confront. If they cannot provide an identification argument, the causal vocabulary should go. That is a major revision, not a desk reject. I would send it to review with the expectation that the causal claims be either derived or removed.\n\nNot something I would cite in its current form.","headline":"A solid gated-fusion transformer for crowd flow wrapped in an unsupported causal backdoor story; refereeing should force the causal claims to be either derived or removed.","tokens_in":26849,"tokens_out":2081,"would_cite":false,"duration_ms":19652,"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":"The paper argues that crowd flow prediction should be modeled as the causal effect of the past on the future, and builds a transformer that de-confounds spatial and temporal attributes to estimate that effect.","keywords":["crowd flow prediction","spatial-temporal prediction","causal de-confounding","backdoor adjustment","cross-time attention","spatial-temporal transformer","out-of-distribution generalization","spatial-temporal graph networks"],"falsifier":"On a spatial-temporal dataset where the true causal effect is known (for example, a synthetic process with an intervention that sets past flow), compare STDCformer’s estimated $P(Y|\\operatorname{do}(X))$ with the ground-truth intervention distribution; if the estimate is no closer than an associative baseline, the de-confounding claim collapses. A cheaper test would shuffle the auxiliary confounder features while keeping the flow data identical: if the weights are causal, prediction error on shifted test distributions should grow, whereas a pure attention gate would be insensitive.","tokens_in":25628,"feed_emoji":"🚕","tokens_out":8114,"duration_ms":67357,"temperature":0.7,"pith_summary":"The paper proposes to view crowd flow prediction not as a direct learned map from past to future observations, but as three linked processes — encoding, cross-time mapping, and decoding — and argues that the key is learning a representation space where the past-to-future mapping is causal rather than merely correlational. It claims that each spatial-temporal token’s spatial attributes (where) and temporal attributes (when) act as confounders that bias ordinary training, and introduces a spatial-temporal backdoor adjustment that reweights spatial and temporal representations during fusion. The resulting model, STDCformer, is reported to beat STGNN and ST-transformer baselines on two New York City taxi-based crowd flow datasets and to generalize zero-shot from Manhattan to Brooklyn. If correct, this suggests that explicitly modeling confounders in urban flow data improves both prediction accuracy and out-of-distribution robustness.","feed_headline":"De-confounded transformer beats prior models on crowd flow","feed_subtitle":"Treats 'when' and 'where' as confounders, reweights them, and transfers zero-shot from Manhattan to Brooklyn.","key_machinery":"The central object is the spatial-temporal backdoor adjustment identity, $P(Y|\\operatorname{do}(X)) = P(Y|X, C=C_S)\\,P(C=C_S) + P(Y|X, C=C_T)\\,P(C=C_T)$, where the confounders are the spatial attributes (‘where’) and temporal attributes (‘when’) of each spatial-temporal token. The identity is made operative in the Spatial-Temporal De-Confounded Fusion module, which computes scalar weights $P(C_S)=\\operatorname{Sigmoid}(C_S+C_T)$, $P(C_T)=1-P(C_S)$ from learned confounder embeddings and uses them to gate the sum of the spatial-attention and temporal-attention representations. This gated fusion is what simulates intervention. The Spatial-Temporal Embedding (STE) and Cross-Time Attention are the complementary machinery that let the model query the past-to-future relationship using those same confounder representations.","core_discovery":"The paper’s central claim is that the right objective for spatial-temporal prediction is the de-confounded causal effect $P(Y|\\operatorname{do}(X))$ rather than the associative fit $P(Y|X)$. It argues that each spatial-temporal token carries background attributes — spatial attributes $S$ such as location, functionality, and safety, and temporal attributes $T$ such as time of day, holiday, and weather — that influence both historical and future observations, making them confounders of the past-to-future mapping. To remove their bias, the paper proposes a spatial-temporal backdoor adjustment that partitions the confounders into $C=\\{C_S, C_T\\}$ and implements the adjustment in a fusion module as $H = P(C_S)\\cdot STR'_S + P(C_T)\\cdot STR'_T$ with $P(C_S)=\\operatorname{Sigmoid}(C_S+C_T)$ and $P(C_T)=1-P(C_S)$. This defines the Spatial-Temporal De-Confounded (STDC) representation space. Cross-time attention between future and past spatial-temporal embeddings then answers how to map from past to future in that space. Empirically the paper reports state-of-the-art MAE and RMSE on two New York City taxi-derived crowd flow datasets and stronger zero-shot transfer from Manhattan to Brooklyn than transformer and STGNN baselines.","pith_inferences":["The soft Sigmoid gating is not a stratified backdoor estimator: it collapses the confounder distribution into a single scalar trade-off between spatial and temporal weights, so the causal interpretation rests on how faithfully the auxiliary features represent the true confounders.","A straightforward next test is to compare the de-confounding weights against inverse-propensity reweighting computed from known confounders on a synthetic spatial-temporal dataset, where the true intervention effect is computable.","Because the paper’s zero-shot OOD test moves between two boroughs of the same city with the same taxi infrastructure, a stronger test would transfer across cities or sensor modalities; if the advantage disappears there, the de-confounding benefit may be task-specific.","The same backdoor-adjustment skeleton might transfer to traffic, epidemic, or air-quality forecasting by swapping the auxiliary confounder features, which is a testable extension the paper does not run."],"forward_implications":["If the causal graph is correct, training with the backdoor adjustment should reduce spurious correlations between past and future flow, which is the stated reason for the reported gains in IID accuracy.","The learned confounder weights carry interpretable physical meaning: the model raises the spatial weight for low-traffic, park-like zones and lowers it for busy mixed-function districts, so the weights can be inspected to explain predictions.","Cross-time attention exposes how each region maps past to future: short-horizon matches attend to the most recent past, while longer horizons shift attention to historically similar segments, giving a per-region explanation of the mapping.","Zero-shot transfer from Manhattan to Brooklyn is reported to work better than transformer baselines, suggesting that the de-confounded representation retains usable structure across spatial distributions.","The encode–map–decode decomposition frames spatial-temporal prediction as representation learning plus causal mapping, a skeleton the paper claims can be instantiated beyond crowd flow."],"supporting_citations":[{"why":"Supplies the causal-inference and backdoor-adjustment framework used to define the de-confounded causal effect $P(Y|\\operatorname{do}(X))$.","marker":"[49]"},{"why":"Motivates the problem of spatio-temporal shifts and deconfounding that the paper’s backdoor strategy addresses.","marker":"[23]"},{"why":"Representative prior work that learns confounder representations from the raw observation time series, which the paper contrasts with its auxiliary-information-based confounders.","marker":"[24]"},{"why":"Prior method (STEVE) that handles finite confounder categories and layers; the paper positions its temporal/spatial partition against this.","marker":"[25]"},{"why":"Prior method that uses POI and regional attributes as confounders with a predefined number of categories, which the paper avoids by learning distributions adaptively.","marker":"[26]"},{"why":"Strongest transformer baseline (PDFormer); the paper must beat it to support its SOTA claim in the IID and zero-shot experiments.","marker":"[18]"},{"why":"Representative STGNN baseline (Graph WaveNet) used in the comparison tables, supporting the empirical claim against graph-based models.","marker":"[13]"}],"fun_headline_variants":["Causal de-confounding lifts crowd flow accuracy","De-confounded transformer beats priors on crowd flow","Backdoor adjustment yields stronger crowd flow forecasts","Unbiased crowd flow mapping with causal transformer","STDCformer: de-confounded spatial-temporal prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The causal story assumes that the chosen spatial and temporal auxiliary features are the only confounders of the past-to-future relationship, so the learned weights estimate the true intervention effect rather than merely acting as data-dependent attention gates.","fun_headline_variants_meta":{"raw":{"variants":["Causal de-confounding lifts crowd flow accuracy","De-confounded transformer beats priors on crowd flow","Backdoor adjustment yields stronger crowd flow forecasts","Unbiased crowd flow mapping with causal transformer","STDCformer: de-confounded spatial-temporal prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000353,"raw_usage":{"total_tokens":2015,"prompt_tokens":1133,"completion_tokens":882,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":749,"completion_tokens_details":{"reasoning_tokens":811}},"tokens_in":749,"tokens_out":882,"duration_ms":7590,"temperature":1.0,"reasoning_tokens":811,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:55:58.472818+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a spatial-temporal dataset where the true causal effect is known (for example, a synthetic process with an intervention that sets past flow), compare STDCformer’s estimated $P(Y|\\operatorname{do}(X))$ with the ground-truth intervention distribution; if the estimate is no closer than an associative baseline, the de-confounding claim collapses. A cheaper test would shuffle the auxiliary confounder features while keeping the flow data identical: if the weights are causal, prediction error on shifted test distributions should grow, whereas a pure attention gate would be insensitive.","supporting_citations":[{"cited_title":"Pearl, D","cited_arxiv_id":null,"evidence_quote":"Supplies the causal-inference and backdoor-adjustment framework used to define the de-confounded causal effect $P(Y|\\operatorname{do}(X))$."},{"cited_title":"Seeing the Unseen: Learning Basis Confounder Representations for Robust Traffic Prediction","cited_arxiv_id":"2311.12472","evidence_quote":"Motivates the problem of spatio-temporal shifts and deconfounding that the paper’s backdoor strategy addresses."},{"cited_title":"Rahmani, A","cited_arxiv_id":null,"evidence_quote":"Prior method (STEVE) that handles finite confounder categories and layers; the paper positions its temporal/spatial partition against this."},{"cited_title":"Jiang, C","cited_arxiv_id":null,"evidence_quote":"Strongest transformer baseline (PDFormer); the paper must beat it to support its SOTA claim in the IID and zero-shot experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Representative STGNN baseline (Graph WaveNet) used in the comparison tables, supporting the empirical claim against graph-based models."}],"review_version":1}