acceptodds
Under review as a conference paper at ICLR 2027

SDF-SP: Semantic Dual-Order Spatiotemporal Forecasting with Simplicial Projection

Abstract

Spatiotemporal forecasting must capture temporal dynamics and cross-node interactions under changing contextual conditions. Contextual semantics and historical trajectories provide complementary information, but incorporating node-specific contextual effects and combining divergent historical responses remain challenging. We propose SDF-SP, a framework that integrates semantically conditioned dual-order modelling with retrieval-based forecast correction. Its Dual-Order Spatiotemporal Encoder processes temporal-first and spatial-first branches in parallel, using context–node semantic relevance to modulate temporal features and spatial propagation. A context-conditioned gate fuses the branch representations, which a frozen pretrained language-model backbone uses to generate a base forecast. Retrieval-Augmented Simplicial Residual Projection then refines this forecast using retrieved historical trajectories. Guided by the base forecast, global and node-level projections model joint cross-node responses and individual node trajectories, respectively. Local simplicial constraints restrict interpolation to connected historical responses, and the projected forecasts are adaptively combined according to their deviations from the base forecast. Experiments on six real-world datasets spanning tourism, transportation and energy demonstrate strong forecasting performance across these domains. The code is publicly available at https://anonymous.4open.science/r/SDF-SP/.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatiotemporal forecasting, Time series forecasting, Semantic conditioning, Retrieval-augmented learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CUTgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SDF-SP, a framework combining semantic-guided dual-order spatiotemporal modelling with retrieval-based forecast correction. The Dual-Order Spatiotemporal Encoder processes temporal-first and spatial-first branches in parallel, using contextual semantics to modulate temporal features and spatial propagation.A frozen pretrained GPT-2 backbone, followed by trainable prediction heads, produces the base forecast. The Retrieval-Augmented Simplicial Residual Projection module further corrects the forecast by combining retrieved historical responses through local simplicial projection at global and node levels. Experiments on six real-world datasets spanning tourism, transportation, and energy demonstrate strong forecasting performance, achieving the best results on 10 of 12 MAE/RMSE comparisons against 15 baselines.

Critical strengths
  1. The paper addresses meaningful spatiotemporal forecasting problems by incorporating contextual semantics into temporal and spatial modelling and using historical responses to refine predictions.
  2. The method achieves strong performance across six datasets and 15 baselines, with ablation studies evaluating its major components.
  3. The retrieval correction module can be applied to existing forecasters without retraining, and the paper provides relatively detailed mathematical formulations and implementation procedures.
Critical weaknesses
  1. The distinct contributions of dual-order modelling and semantic modulation could be better substantiated, given that related architectures and conditioning mechanisms have been explored in prior studies. The benefits of individual components are not sufficiently isolated. (Q1–Q2)
  2. The motivation for local simplicial projection could be strengthened. Although its mathematical formulation is provided, the specific advantage of simplicial constraints over simpler sparse interpolation remains insufficiently demonstrated. (Q3–Q4)
  3. The reliability of base-guided projection and adaptive fusion requires further validation, particularly when base forecasts are inaccurate. The sensitivity of retrieval correction and the performance degradation observed when applying it to some external forecasters also deserve further investigation. (Q5–Q6)
  4. The independent contribution of LLM-generated semantics could be more clearly demonstrated, and the consistency of contextual information available to different baselines requires clarification. (Q7)
Questions for authors

Q1. What distinguishes the proposed dual-order encoder from existing approaches? Previous studies have explored parallel TS/ST processing. Could the authors clarify the additional methodological contribution of the proposed semantic conditioning mechanism?

Q2. Can the contributions of different components be isolated? The two branches differ in computation order, semantic modulation, and graph normalisation. Why does ST-only underperform the variant without dual-order encoding in Table 2? Also, given node-specific contextual effects, why is sample-level rather than node-level gating used?

Q3. What is the advantage of simplicial projection over simpler sparse interpolation? Could the authors compare simplicial projection with cardinality-matched sparse convex projection without connectivity constraints, while keeping the global/node-wise projection and fusion settings unchanged?

Q4. How much does each simplicial component contribute? Could the authors compare vertex-only, vertex-plus-edge, and full simplicial projection, and report the frequency of each selected projection type?

Q5. What justifies the preference for smaller corrections to the base forecast? Could the authors clarify the motivation for this conservative fusion rule and compare it with simpler alternatives, such as uniform averaging, particularly when the base forecast is inaccurate?

Q6. How sensitive is retrieval correction to its heuristic parameters and historical-response coverage? Could the authors evaluate the sensitivity to retrieval size K = 6, neighbourhood size k = 3, and radius factor γ = 2, and explain why retrieval correction increases RMSE for 9 of the 15 baselines on SDWPF in Table 3?

Q7. What additional benefits do LLM-generated semantics provide? Could the authors compare LLM-generated embeddings with direct encoding of the same structured contextual variables and clarify whether the baselines receive equivalent external information?

8 comments
AuthorsOpenReview ↗

Response for W1 & Q1: Order-specific semantic modulation in the dual-order encoder

Thank you for asking how SDF-SP differs from earlier dual-order encoders. SDF-SP extends the parallel dual-order structure by using the external context of each sample to modulate temporal features and spatial propagation within the two branches. In the temporal-first (TS) branch, the model scales each node’s temporal features according to how its description matches the context and where the observation falls in the historical window. It then propagates these features over the fixed graph. In the spatial-first (ST) branch, the model matches the same context to the nodes at both ends of each existing edge and adjusts its weight, changing how much each neighbour contributes during aggregation.

To examine how this modulation affects the base forecast, we added an eight-step comparison before retrieval correction in Section 4.3.2 and Appendix Table 10. All three variants retain the dual-order structure and the context-conditioned gate. On Beijing Scenic, Beijing Metro and SDWPF, concatenating the node and context embeddings at the input reduced both MAE and RMSE relative to the unmodulated variant. Using the same semantic information to modulate temporal features and edge weights within the branches reduced both errors further.

Semantic conditioning in eight-step base forecasts (Table 10, MAE / RMSE)

Semantic injectionBeijing ScenicBeijing MetroSDWPF
Without modulation688.49 / 1487.1764.37 / 167.70370.49 / 487.39
Input concatenation532.51 / 1110.8359.37 / 154.50328.87 / 454.69
Order-specific modulation (SDF-SP)501.75 / 1065.5856.04 / 140.46324.03 / 426.35
AuthorsOpenReview ↗

Response for W1 & Q2: Branch ablations and the choice of fusion gate

Thank you for asking about the performance of the ST-only variant and for suggesting a comparison between sample-level and node-level gating.

Historical updates in the ST-only variant

The w/o Dual-order variant passes the embedded history directly to the temporal and spatial compression layers. ST only adds an update to the history before compression. This update comes entirely from the ST path, which aggregates neighbours before extracting temporal features. The full encoder also has a TS path that extracts each node’s temporal features before spatial aggregation. Its gate assigns more weight to TS in every test window on Beijing Scenic and Beijing Metro. On SDWPF, the gate gives more weight to ST in a minority of windows.

On SDWPF, ST only raised base MAE in more windows than it lowered it relative to w/o Dual-order. The average increase also exceeded the average decrease, yielding a higher overall MAE. On Beijing Metro, ST only achieved lower base MAE, but w/o Dual-order benefited more from retrieval correction and achieved lower MAE after correction. A spatial-first update can therefore help in individual windows, yet relying on it alone to update every window can yield higher overall error than encoding the history directly. The full encoder achieves lower base MAE and RMSE than either single branch on all three datasets in Table 2. This supports using ST to provide neighbourhood information alongside the temporal features that TS extracts from each node’s own history.

Rationale for sharing fusion weights across nodes

SDF-SP receives one external context representation per forecasting window. Before fusion, TS uses the match between the context and each node to modulate that node’s temporal features at each historical time step. ST adjusts edge weights to control how much each neighbour contributes to the target node. Because both branches already capture node-specific responses to the same context, we use the window context to set one pair of TS/ST weights. Each node applies that pair to its own two branch representations. We have clarified the rationale for sharing these weights in the paragraph following Eq. (4) in Section 3.4 of the revised manuscript. To test whether the mixing weights should also vary across nodes, we trained a node-level variant that computes a separate pair of weights from the window context and each node’s semantic representation. We also added Table 15 in Appendix E.7 to report the eight-step forecast errors for both gates before retrieval correction. The sample-level gate yielded lower base MAE and RMSE on all three datasets.

Sample-level and node-level gates before retrieval correction (Appendix E.7, Table 15, MAE / RMSE)

DatasetSample-level gateNode-level gate
Beijing Scenic501.75 / 1065.58522.78 / 1093.47
Beijing Metro56.04 / 140.4659.59 / 157.02
SDWPF324.03 / 426.35338.13 / 453.22
AuthorsOpenReview ↗

Response for W2 & Q3: Local connectivity at matched sparsity

We thank the reviewer for suggesting a test of whether local connectivity improves sparse interpolation. We added a sparse convex projection baseline without a connectivity constraint. Both methods use the same base forecasts, retrieved historical responses, global and node-level projection settings, and fusion rule. For each test window, we matched the number of active vertices separately for the global projection and for each node. Local simplicial projection requires the selected vertices to be pairwise connected. The sparse baseline can choose any subset of that size from the same retrieved responses.

With the vertex count matched, local simplicial projection achieved lower eight-step MAE and RMSE on all three datasets. These results suggest that restricting which historical responses can be interpolated together improves forecast accuracy beyond the benefit of sparsity alone. We describe the comparison in Section 4.3.3. Figure 5(a) shows the MAE reductions relative to the base forecast, and Appendix Table 13 reports the MAE and RMSE for each dataset.

Eight-step forecasts with the active vertex count matched (Appendix E.5, Table 13, MAE / RMSE)

DatasetSparse convex projectionLocal simplicial projection
Beijing Scenic378.88 / 918.98370.03 / 901.63
Beijing Metro43.19 / 138.0840.97 / 128.79
SDWPF310.40 / 429.73297.67 / 407.59
Reviewer_XAKGOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SDF-SP, a two-stage spatiotemporal forecasting framework. The first stage is a semantically conditioned Dual-Order Spatiotemporal Encoder. Context–node semantic relevance, derived from LLM-prepared node descriptions and daily context descriptions, modulates temporal features in the temporal-first branch and spatial edge weights in the spatial-first branch. A context-conditioned gate fuses the two branch representations, and a frozen pretrained language-model backbone generates the base forecast. The second stage is Retrieval-Augmented Simplicial Residual Projection. Historical residual responses retrieved from a training-only memory form local simplicial supports, onto which the base forecast is projected through base-guided global and node-level projections. The two projected forecasts are then adaptively combined according to their deviations from the base forecast to produce the final prediction.

Critical strengths
  1. The work is well engineered. The code and complete implementation details are provided, which benefits reproducibility and comparison in future work.
  2. Formulating retrieval-based correction as projection onto local simplicial supports has a clear geometric motivation. The ablation design is also clean, and its results support this design well.
  3. The retrieval correction module requires no retraining and can be applied directly to the outputs of any forecaster. It makes a practical contribution to post-hoc forecast correction in time-series forecasting.
Critical weaknesses
  1. The first contribution of the paper is the semantically conditioned Dual-Order Spatiotemporal Encoder. However, Table 2 shows that the base forecast of the full dual-order encoder, before retrieval correction, is not competitive: the MAE is 501.75 on Beijing Scenic (vs. 380.90 for TimeCMA in Table 1), 56.04 on Beijing Metro (vs. 42.22 for TimeMixer), and 324.03 on SDWPF (vs. 310.09 for DLinear).
  2. Appendix Table 11 shows that when the same retrieval module is applied to the baselines, some of them outperform the full SDF-SP. These results seem to suggest that the performance advantage mainly comes from the retrieval stage. Do the results then support the claim that "Ablation studies further support the contributions of dual-order encoding and retrieval-based correction"?
  3. Is the source of the retrieval gains adequately delimited by simple baselines? Section 4.3.3 compares similarity-weighted aggregation, full convex-hull projection and local simplicial projection, and reports that local simplicial projection yields the lowest errors. However, these comparisons are all based on the same retrieved cases, and each retrieval key explicitly includes known future calendar features (Section 3.5), while the datasets exhibit strong daily and weekly periodicity.
  4. The paper does not compare with simple baselines such as a seasonal naive forecast (e.g., the same time in the previous week), the historical mean of the same time slot, or retrieval using calendar features only. Moreover, on SDWPF, which is less periodic, the retrieval correction reduces RMSE for only 6 of the 15 baseline forecasters (Table 3), and similarity-weighted aggregation increases MAE by 2.6% relative to the base forecast (Figure 4). Does this suggest that the retrieval gains may largely come from calendar matching in periodic data? The contribution of local simplicial projection itself may need to be further clarified.
Questions for authors
  1. For the scenic datasets, which day's context description does each forecasting example use? The example in Appendix B includes the weather and rainfall of the same day, while the forecast origin falls within that day's observation hours. Are the weather and rainfall of that day fully available at the forecast origin? How is the context handled when the historical window spans two days?
  2. In Section 4.3.2, are the results of the semantic injection comparison ("Without modulation", "Input concatenation" and "Order-specific modulation") computed on the base forecast before retrieval correction, or on the final forecast after correction? If the latter, could the authors report the comparison on the base forecast, so as to reflect the effect of semantic modulation itself more directly?
8 comments
AuthorsOpenReview ↗

Response for W1: The dual-order encoder before and after retrieval correction

Thank you for drawing attention to the accuracy of the base forecasts.

Accuracy before correction

We agree that, before retrieval correction, SDF-SP has a higher MAE than TimeCMA on Beijing Scenic, TimeMixer on Beijing Metro, and DLinear on SDWPF. To assess the contribution of dual-order encoding to the base forecast, we compare the full encoder with its two single-branch variants in the Base columns of Table 2. The temporal-first branch (TS only) is the stronger single branch on all three datasets. Using both branches reduces MAE from 545.67 to 501.75 on Beijing Scenic, from 58.24 to 56.04 on Beijing Metro, and from 349.21 to 324.03 on SDWPF. These improvements are already present before retrieval correction.

Base forecasts and the external comparator identified for each dataset (Tables 1 and 2, MAE)

DatasetExternal baseline (MAE)TS onlyDual-order
Beijing ScenicTimeCMA (380.90)545.67501.75
Beijing MetroTimeMixer (42.22)58.2456.04
SDWPFDLinear (310.09)349.21324.03

Rankings when every forecaster receives the same correction

Section 4.3.4 examines how the full SDF-SP model compares with other forecasters when they also receive the same retrieval correction. We applied the correction module to the forecasts of all 16 baselines while keeping their trained parameters fixed. Some corrected baselines achieved lower errors on individual metrics. We ranked the 17 methods by corrected error for each metric (MAE and RMSE) on each of the three datasets, assigning tied methods their mean rank. We then averaged the six ranks for each method. SDF-SP ranked first in three comparisons, fourth in one and fifth in two. Its mean rank of 2.83 placed it first overall, 2.92 places ahead of STEMLP (5.75). TimeMixer ranked third (6.33). Table 3 reports the relative change in error for each baseline, and Appendix E.6 gives the absolute errors before and after correction (Table 14 in the revised manuscript).

Rankings after correction. Each dataset column gives the MAE rank / RMSE rank, with 1 best.

Overall rankForecasterMean rankBeijing ScenicBeijing MetroSDWPF
1SDF-SP2.835 / 14 / 51 / 1
2STEMLP5.756 / 7.51 / 210 / 8
3TimeMixer6.3316 / 123 / 32 / 2
4EAD-STN6.5010 / 96 / 65 / 3
5DPSTGC7.503 / 38 / 1111 / 9
6ST-MambaSync7.927 / 7.52 / 115 / 15
7RAFT8.1711 / 119 / 77 / 4
8TimeBridge9.002 / 410 / 1014 / 14
9TS-RAG9.1713 / 157 / 86 / 6
10SARAF9.3314 / 1411 / 48 / 5
11CALF9.674 / 514 / 159 / 11
12DLinear9.8317 / 175 / 94 / 7
13TimeCMA10.831 / 215 / 1417 / 16
14PFRP11.009 / 613 / 1313 / 12
15Time-LLM12.338 / 1017 / 1712 / 10
16GPT4TS13.1715 / 1616 / 163 / 13
17iTransformer13.6712 / 1312 / 1216 / 17
AuthorsOpenReview ↗

Response for W2: Contributions of the encoder and retrieval correction

Thank you for raising this point. We agree that retrieval correction contributes substantially to forecast accuracy and that some baselines outperform SDF-SP on individual datasets after receiving the same correction.

Table 2 lets us examine each component by removing it from the full model. With the dual-order encoder retained, disabling retrieval increases MAE by 131.72 on Beijing Scenic, 15.07 on Beijing Metro, and 26.36 on SDWPF. With retrieval retained, removing the dual-order module increases MAE by 20.83, 1.39, and 19.34 on the same datasets. Removing retrieval incurs the larger MAE increase in all three comparisons, while removing dual-order encoding also worsens the final forecasts on all three datasets. The ablations thus support a contribution from both components, with the size of each contribution varying by dataset.

Eight-step MAE with each component removed in turn (Table 2)

Retrieval correctionDual-order encoderBeijing ScenicBeijing MetroSDWPF
OffRetained501.7556.04324.03
OnRemoved390.8642.36317.01
OnRetained370.0340.97297.67
AuthorsOpenReview ↗

Response for W3 & W4: Calendar matching and the contribution of local simplicial projection

Thank you for pointing out that our original comparison held the retrieved examples fixed. This isolated the effect of the correction rule. The benefit of calendar matching alone was still an open question.

Weekly baselines and calendar-only retrieval

We added the three suggested controls for eight-step forecasts on Beijing Scenic, Beijing Metro and SDWPF (Appendix E.9, Table 18). The first two use the observation at the same node and time one week earlier, and the training-set mean for that node, weekday and time of day, respectively. SDF-SP achieved lower MAE and RMSE than both baselines on all three datasets. Thus, neither repeating the previous week’s observation nor averaging past observations from the same weekly slot reproduced the accuracy of the full model.

The third control uses only calendar features known at the forecast origin to retrieve the same number of training windows as SDF-SP on each dataset. It uses the same frozen base forecaster, local simplicial projections and fusion rule as SDF-SP. On Beijing Metro, calendar-only retrieval reduced MAE and RMSE from 56.04 and 140.46 for the base forecast (Table 2) to 46.55 and 133.19. Calendar matching thus helps on this dataset. The full retrieval key also includes each node’s recent level, variability and changes over time, bringing the errors down to 40.97 and 128.79. On SDWPF, calendar-only retrieval increased MAE and RMSE from 324.03 and 426.35 for the base forecast to 330.40 and 462.07, while the full key reduced them to 297.67 and 407.59. The full retrieval configuration, which includes recent time-series features, yielded lower MAE and RMSE than calendar-only retrieval on all three datasets. Calendar matching may therefore account for some of the improvement, but these results do not establish that retrieval gains arise primarily from calendar matching on periodic datasets.

Weekly baselines and calendar-only retrieval (Table 18, MAE / RMSE)

MethodBeijing ScenicBeijing MetroSDWPF
Weekly seasonal naïve558.87 / 1389.0859.55 / 135.53446.82 / 572.99
Same-slot historical mean1691.78 / 3340.3965.08 / 149.59370.92 / 463.19
Calendar-only retrieval + local simplicial projection398.00 / 926.9946.55 / 133.19330.40 / 462.07
SDF-SP370.03 / 901.6340.97 / 128.79297.67 / 407.59

Projection with identical retrieved windows

In the revised Table 3, 7 of the 16 external forecasters achieved lower RMSE after correction on SDWPF, so the benefit varies across forecasters. For SDF-SP itself, Section 4.3.3 holds the retrieved windows fixed. On SDWPF, similarity-weighted aggregation increased MAE from 324.03 for the base forecast to 332.46 (a 2.6% increase), while local simplicial projection reduced it to 297.67. To test the role of local connectivity, we added a sparse convex projection that uses the same number of vertices per projection but can select any subset of that size. Removing the connectivity constraint gave an MAE of 310.40 on SDWPF and higher MAE and RMSE than local simplicial projection on all three datasets (Figure 5(a) and Table 13). A second new control held the base forecasts, retrieved windows and fusion rule fixed. Allowing interpolation along edges reduced both errors on all three datasets compared with selecting a single vertex. Allowing interpolation over triangles reduced both errors further (Figure 6 and Table 16). These controls show that even when the retrieved windows are identical, local simplicial projection further reduces MAE and RMSE on all three datasets by restricting which historical responses can be combined and interpolating between neighbouring responses to fit the base forecast residual.

Correction rules applied to the same retrieved windows (Appendix E.5, Table 13, MAE / RMSE)

Correction ruleBeijing ScenicBeijing MetroSDWPF
Similarity-weighted aggregation394.31 / 998.7350.72 / 133.36332.46 / 454.41
Full convex-hull projection382.51 / 929.2542.00 / 131.78309.29 / 425.42
Sparse convex projection378.88 / 918.9843.19 / 138.08310.40 / 429.73
Local simplicial projection370.03 / 901.6340.97 / 128.79297.67 / 407.59
Reviewer_cv8uOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SDF-SP, a two-stage spatiotemporal forecasting framework. A semantically conditioned dual-order encoder combines temporal-first and spatial-first processing, while a retrieval-based residual correction module uses global and node-level simplicial projections followed by adaptive fusion. Experiments on six datasets and comparisons with 15 baselines report promising accuracy and some transfer of the correction module to existing forecasters. The training-only retrieval memory and chronological splits are appropriate, but the evidence does not yet cleanly establish which components drive the gains or whether the comparisons are input-matched.

Critical strengths
  • The work combines semantic conditioning, dual-order spatiotemporal modeling, and structured retrieval-based residual correction in a coherent and potentially useful framework. The distinction between global and node-level projections is technically distinctive.
  • The evaluation spans six datasets, multiple forecasting families, and several component studies. The chronological split and training-only retrieval memory are appropriate safeguards against direct leakage.
Critical weaknesses
  • Major Issue 1 — The central algorithm is ambiguous, and component isolation is incomplete. Equation (10), its surrounding explanation, and Algorithm 1 appear to implement different fusion rules: Eq. (10) appears to favor the larger deviation, whereas Algorithm 1 favors the smaller deviation. This directly changes the final forecast and must be resolved. In addition, Table 2’s “w/o Dual-order” variant “directly tokenizes the history,” so it changes encoder capacity and residual-history processing rather than isolating temporal-first versus spatial-first computation. The current evidence therefore does not establish the claimed mechanism.
  • Major Issue 2 — The comparison is uncontrolled and reliability analysis is insufficient. Section 4.1 and Table 1 do not establish that external baselines receive the same calendar, weather, event, node, and semantic inputs as SDF-SP. The reported means over three random seeds also lack standard deviations, confidence intervals, or significance analysis. These issues make it difficult to attribute the gains to the proposed architecture rather than additional information or run-to-run variation.

Minor Issues:

  • The semantic preprocessing pipeline is not fully specified across all six datasets, including prompts, fields, generation settings, and deterministic preprocessing (Appendix B and the experimental setup).
  • The practical cost, memory size, and latency of retrieval correction are not reported, despite exhaustive retrieval and projection for large-node datasets (Appendix D.5).
  • The plug-and-play claim should be qualified because Table 3 shows mixed effects across models, metrics, and datasets.
  • Wide tables are difficult to verify in the supplied rendering; this is secondary and may partly be a rendering artifact.
Questions for authors
  • Clarification that could change the assessment: Which fusion rule was actually used to produce the reported results: the formula in Eq. (10) or Algorithm 1? Please provide the exact implemented expression and, if possible, results for both rules.
  • For every baseline in Table 1 and every model in the plug-in experiment, which numerical, graph, calendar, weather, event, node-text, and future-known inputs are available?
  • Are semantic descriptions generated deterministically and only from information available at each forecast origin? Please provide the dataset-by-dataset availability protocol.
  • Can the authors report per-seed results or uncertainty estimates for the main comparisons and key ablations?
10 comments
AuthorsOpenReview ↗

Response for W1 & Q1: Projection fusion and the scope of the encoder ablation

Thank you for asking us to clarify how the global and node-level projections are weighted. We rechecked Eq. (10), the accompanying text and Algorithm 1 in the original submission and confirmed that all three specify the same fusion rule.

Fusion weights in Eq. (10)

In the reported SDF-SP forecasts, we used this rule to give greater weight to the projection closer to the base forecast. For a given window and node, let dgd_g and dnd_n be the mean absolute deviations of the global and node-level projections from the base forecast, averaged over the eight forecast steps. Equation (10) defines the weights as

\begin{cases} \dfrac{d_n}{d_g+d_n}, & d_g+d_n > 0,\\\\ \dfrac{1}{2}, & d_g+d_n = 0, \end{cases} \qquad w_n=1-w_g.$$ The final forecast is $\widehat{\mathbf y}=w_g\widehat{\mathbf y}_g+w_n\widehat{\mathbf y}_n$. The numerator of the global weight is the *node-level* deviation $d_n$. Thus, if $d_g < d_n$, then $w_g > w_n$. When both deviations are zero, each projection receives a weight of $1/2$. In Section 4.3.3, we compare weighting in favour of the smaller deviation, weighting in favour of the larger deviation and equal weighting (0.5 for each projection). **Fusion rules on the full test sets (MAE / RMSE)** | Dataset | Equal weights | Favour smaller deviation | Favour larger deviation | | :--- | ---: | ---: | ---: | | Beijing Scenic | 373.03 / 922.74 | **370.03 / 901.63** | 378.78 / 946.39 | | Beijing Metro | 41.71 / 135.85 | **40.97 / 128.79** | 42.21 / 136.36 | | SDWPF | 313.33 / 437.67 | **297.67 / 407.59** | 315.24 / 440.60 | Favouring the smaller deviation gave the lowest MAE and RMSE on all three datasets. The same rule also gave the lowest errors on the 10% of windows with the highest base forecast MAE in each dataset. Figure 7 in Section 4.3.3 shows the error changes relative to equal weighting for both sets of windows. Table 13 in Appendix E.5 gives the absolute errors. ### **What the encoder ablation measures** Our description of the Table 2 ablation in the original submission was inaccurate. The phrase “directly tokenizes the history” omitted the embedding step before compression. We have revised Section 4.3.1 to clarify that the *w/o Dual-order* variant first embeds the numerical history and then passes the embeddings directly to the original temporal and spatial compression layers. This comparison holds the capacity of the input embedding and both compression layers fixed and measures the effect of removing the dual-order encoder as a whole.
AuthorsOpenReview ↗

Response for W2 & Q2: Inputs available to each baseline

Thank you for asking which inputs the forecasters in Table 1 share with SDF-SP.

Inputs of the original baselines

All 15 original baselines use the same 24-step numerical history as SDF-SP to forecast the next eight steps. Some also share spatial relationships, calendar features or access to historical windows and their observed continuations from the same training period. The four language-model baselines derive their inputs from the same numerical observations. The table gives the inputs for each forecaster.

Original baselinesInputs and how the model uses them
DLinear, TimeMixer, iTransformer, TimeBridgeForecast directly from the numerical history.
DPSTGCUses spatial relationships through its own graph representation.
ST-MambaSyncUses time of day and day of week from the timestamps.
STEMLPUses its own spatial graph representation and calendar features derived from the timestamps.
RAFT, TS-RAG, PFRPEach builds its own retrieval memory from training windows and their observed continuations.
SARAFUses the same training-period records and also uses timestamps to retrieve windows.
GPT4TSPasses patches of the numerical history to a pretrained model.
CALFAligns the historical observations with precomputed word embeddings.
Time-LLMPrompts describe the dataset and forecasting task and include statistics of the input window.
TimeCMAConstructs prompts from the historical observations and their timestamps.

None of the 15 original baselines receives both the node descriptions and the window-specific external context used by SDF-SP. These comparisons alone therefore cannot separate the benefit of the architecture from that of the additional inputs.

Comparison with matched external information

To compare forecasters with access to the same external information, we added EAD-STN to Table 1. Shen et al. (2027) developed EAD-STN (event-aware dynamic spatiotemporal network) to forecast visitor flows at tourist attractions. It uses language-model embeddings to represent external events and models relationships among attractions through spatial and semantic graphs. On each dataset, EAD-STN and SDF-SP use the same 24-step histories, static spatial graph, and window-specific context vectors, including the calendar, weather, or event information used for that dataset. EAD-STN constructs its semantic graph from the same node-description embeddings. In this comparison, SDF-SP achieves lower eight-step MAE and RMSE on all six datasets. Appendix A specifies the time periods covered by each dataset’s context, including the observed weather for the forecast date used in the retrospective scenic evaluations.

Eight-step forecasts with matched external information (Table 1, MAE / RMSE)

DatasetEAD-STNSDF-SP
Beijing Scenic514.23 / 1285.32370.03 / 901.63
Shanghai Scenic536.67 / 1033.29423.74 / 784.15
Beijing Metro42.93 / 134.7340.97 / 128.79
Beijing Bus43.10 / 72.6439.75 / 65.62
BDG236.85 / 104.4914.62 / 37.99
SDWPF316.44 / 441.66297.67 / 407.59

Reference

Shen, X., Liu, S., & Zhang, Y. (2027). Scenic spot crowd flow prediction: Combining large language model semantic understanding and spatio-temporal network. Tourism Management, 119, 105548. DOI.

Inputs in the plug-in experiment

In the plug-in experiment in Section 4.3.4, all 16 baselines retain their trained checkpoints and their original inputs. The correction module takes each baseline forecast as its projection target. It retrieves historical responses from the training set using statistics of the numerical history and calendar features known for the eight forecast timestamps. The before-and-after comparison for each baseline thus measures the effect of adding this correction to its forecasts.

AuthorsOpenReview ↗

Response for W3 & Q3: Semantic preprocessing and information available at forecast time

We appreciate the reviewer’s request to clarify how we generate semantic descriptions and what information is available at forecast time for each dataset.

Alignment of the inputs with the forecast window

We revised Appendix A to explain how each dataset’s context is aligned with the forecast window and which periods its inputs cover. Beijing Scenic and Shanghai Scenic use the daily record for the date of the first forecast step, with search indices drawn from the previous day and earlier dates. These retrospective experiments use observed weather and rainfall for that date. Operational forecasts would require weather forecasts available at the forecast origin. For Beijing Metro and Beijing Bus, the first forecast step determines the half-hourly calendar and announced-event record. Weather features refer to the preceding full hour. BDG2 and SDWPF summarise weather over the 24 hours preceding the first forecast hour.

DatasetNode profileContext and timing
Beijing Scenic and Shanghai ScenicAttraction features, rating and visitor groups.City record for the date of the first forecast step. The calendar is known in advance. The two search indices use data through the previous day. Events are recorded by date. Weather and rainfall are full-day observations in the retrospective evaluation.
Beijing Metro and Beijing BusStation lines or bus routes and training-period passenger flows.The half-hour of the first forecast step determines the calendar and event record. Events must have been announced by the forecast origin. Weather refers to the preceding complete hour.
BDG2Building use and daily and weekly demand patterns from the training period.Calendar features cover the eight forecast hours. Weather summaries use the 24 hours preceding the first forecast hour.
SDWPFTurbine coordinates, elevation, position in the farm and distance to the nearest turbine.Wind speed, wind direction, temperature and humidity are summarised over the 24 hours preceding the first forecast hour.

Text generation and reuse across training runs

Appendix B.4 now documents the inputs to both stages for all six datasets. Qwen3-8B first writes node descriptions from the node profiles and context descriptions from the records for each forecast window. Qwen3-Embedding-8B then encodes each description as a 4,096-dimensional vector. Figures 8 and 9 illustrate the node and daily context prompts for Beijing Scenic, and Appendix B.4 gives their generation settings. The daily scenic descriptions use sampling, so regenerating them can change their wording. All forecasting runs reuse the same precomputed vectors, giving each window the same semantic input across training seeds.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.