acceptodds
Under review as a conference paper at ICLR 2027

ERST: Learning Interaction from Entity Replacement in Spatiotemporal Forecasting

Abstract

Spatiotemporal forecasting requires modeling interactions among multiple entities whose values change over time. Existing approaches obtain interaction information from entity representations, either by modeling relations among entities or by combining multiple entities into one representation. However, these approaches mainly use the relations or the resulting representation itself and do not directly measure how a representation formed from all entities changes with respect to one entity. We study this change by replacing the representation of one entity with a reference. We propose Entity Replacement for Spatiotemporal Forecasting (ERST), which replaces one entity representation while keeping the other entities, normalization, and function unchanged. ERST applies the same function before and after the replacement and uses the resulting difference as the response for that entity. The reference is constructed from the history of the entity and information from the other entities without directly using the current representation being replaced. The response therefore measures how the representation formed from all entities changes when one entity is replaced. We evaluate ERST on 12 datasets across five domains. ERST improves MAE in 23 of 24 combinations of forecasting architecture and dataset and matches the baseline in the remaining case.

Primary area:
learning on time series and dynamical systems
Keywords:
spatiotemporal forecasting, time series forecasting, interaction modeling, representation learning, entity replacement

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_UyqZOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ERST, which constructs entity-specific interaction features from changes in an aggregate representation after replacing one entity’s representation with a learned reference. These features update entity representations for forecasting. Experiments cover 12 datasets across five domains, multiple forecasting architectures, component ablations, and efficiency evaluations.

Critical strengths

S1. Comprehensive empirical evaluation. The experiments cover 12 datasets across five spatiotemporal domains, including different system scales and forecasting horizons. The evaluation across four additional forecasting architectures is particularly valuable: ERST improves MAE in 23 of the 24 architecture–dataset combinations and matches the baseline in the remaining case. This provides substantial evidence of its practical utility beyond a single backbone or dataset.

S2. Extensive and informative ablations. The paper investigates the response construction, reference inputs, replacement versus removal, response modulation and gating, and response assignment. The additional five-seed comparisons strengthen the evidence for selected design choices. These experiments go considerably beyond reporting overall forecasting accuracy and help establish which components contribute to performance.

S3. A concrete and computationally practical formulation. The replacement response is mathematically specified, and the shared-mean update avoids explicitly reconstructing the full entity set for every replacement. The resulting mechanism provides a practical way to generate entity-specific features, with supporting evaluations on large systems.

S4. I read the offered code, which is reasonable.

Critical weaknesses

W1. The abstract and introduction emphasize that existing approaches do not measure changes under entity replacement, but do not clearly identify the forecasting limitation this omission creates. They should explain what prediction-relevant information is missing and distinguish ERST as a reusable interaction module from the complete forecasting model built around it.

W2. The evidence does not fully substantiate the interpretation of the response as interaction information. Figure 1 shows that similar raw means can conceal entity-level differences, but does not establish that learned aggregation loses these distinctions. Nor does it explain why replacing a hidden representation with the learned reference captures meaningful interactions. The ablations support the construction’s predictive utility, but leave its advantage over a more direct context-dependent feature insufficiently isolated.

Questions for authors

Q1. What concrete forecasting limitation motivates the replacement response? Could the authors clarify the intended mechanism linking entity-specific replacement sensitivity to improved prediction, and explicitly distinguish the reusable interaction module from the full forecasting architecture?

Q2. How are the scalar values in Figure 1(b–c) obtained and normalized from the vector representations and responses? Which case or cases are used, and how do these plots connect to the observed differences in Figure 1(a)?

Q3. What makes the learned reference a useful comparison point beyond its empirical performance?

Q4. ERST appears to be a plug-and-play interaction module applicable to different forecasting models. Could the authors clarify whether the ERST model in Table 1 uses an existing backbone or a custom architecture? Additionally, given the competitive performance of GWNet, MTGNN, and TWIST in Table 1, why were these models not included in Figure 4? Evaluating ERST with these backbones would help assess whether it provides complementary benefits to their existing interaction mechanisms.

Q5. Attention-based architectures are relatively underrepresented in the evaluation. Could the authors include representative models such as GMAN, PDFormer, and STAEformer, particularly given the competitive performance reported for some of these models in prior evaluations? These comparisons would help assess ERST against attention-based interaction modeling.

4 comments
AuthorsOpenReview ↗

Response to Reviewer UyqZ: Motivation, Learned Reference, and Controlled Evaluations

We thank the reviewer for the detailed comments. We clarified the motivation, Figure 1, and the learned reference, and added controlled comparisons.

W1/Q1 — Motivation and forecasting mechanism. An aggregated representation F(H) combines information from all entities, but its value alone does not explicitly measure how that representation changes when one entity is replaced. This is the change ERST explicitly measures; we do not claim that existing graph or attention methods lose entity information. For entity i, ERST constructs a reference and measures the difference between the outputs of the same aggregation function before and after replacing only that entity's representation. Other entities, normalization, and the function remain fixed (Eq. 7). The response reflects both the change in that entity and the current aggregate context (Eq. 10). It is not a prediction or a causal effect. ERST modulates the response using entity, time, and layer information, then uses a gate to control the resulting update to the existing entity representation. This provides a separate response for each entity, which is used to update that entity's representation before prediction. The decoder predicts from the updated representations (Section 3.3). Removing the response increases PEMS03 MAE from 14.28 to 16.23 and T-Drive MAE from 16.56 to 24.01 (Table 3). The reusable ERST interaction module consists of response construction and update, without an encoder or decoder. Tables 1–2 evaluate a custom forecasting model containing this module; Figure 4 and Appendix I.2 evaluate the module added to other architectures (Section 3.1).

W2/Q2 — Figure 1 and direct-context control. Figure 1(a) plots observed PEMS08 flow for Cases A and B, not the output of learned aggregation. The cases have nearly equal means (296.48 and 296.66), but differ across entities; the six marked entities have the largest observed differences. This motivates examining entity-level changes, but does not show that learned aggregation loses information. Panels (b–c) use Case B at the first interaction layer of a trained ERST model. In (b), each response L2 norm is divided by the mean response norm (dashed line = 1); entities are ordered by their observed A–B differences. In (c), each response is replaced with the mean response vector, and the update term is recomputed in the same model. The plotted value is the L2 change in the update divided by the mean original update norm. Case A is not used as the learned reference in panels (b–c). Neither panel reports prediction error (Appendix I.1, Eqs. 62–63).

To compare the replacement response with a direct contextual feature, the Context Feature variant uses c_i = ρ(φ(b_i)) instead of the replacement response. It uses the same learned reference inputs, modulation, update, decoder, and training settings (Appendix H.1). The following are MAE results under the same dataset protocols:

Datasetw/o ResponseContext FeatureERST
PEMS0316.2314.3214.28
PEMS0816.1814.5414.08
KnowAir15.0014.2113.98
SDWPF98.9898.4993.08
Milan206.80223.24201.52
T-Drive24.0118.8216.56

ERST has lower MAE than Context Feature on all ten datasets in Tables 14–16. The margin is small on PEMS03 and larger on some other datasets. These comparisons support the predictive value of the replacement response beyond the tested Context Feature variant; they do not establish a causal interpretation (Section 4.3 and Appendix H.1).

Q3 — Why use a learned reference? The reference for entity i is constructed from its history, information from other entities, and time and layer embeddings, without directly using the current representation being replaced (Eqs. 2–3). It therefore provides a context-dependent comparison point rather than a fixed value. If the reference equals the current representation, the replacement produces zero response. ERST keeps the other entities, normalization, and aggregation function unchanged, unlike removing an entity (Eq. 7; Appendix B.2).

Under the local approximation in Eq. 10, the response depends on both the difference between the current and reference representations and the sensitivity of the aggregation function at the current mean. The reference is learned through the forecasting objective, rather than representing a unique or causal alternative state.

The five-seed PEMS08 results support its construction: ERST achieves MAE 14.074, compared with 14.228 for ψ = 0, 14.410 without history, and 14.418 without information from other entities. All three differences are significant in paired two-sided tests (p < 0.001). See Section 3.2, Appendix B.2–B.3, and Appendix H.1, Table 17.

Q4 — ERST module across forecasting architectures. Tables 1–2 evaluate the complete ERST forecasting model, which uses a custom temporal encoder, entity transformations, ERST interaction layers, and a prediction decoder. The ERST interaction module consists of response construction and update, without the encoder or decoder (Section 3.1). Figure 4 evaluates the module with LSTM, STID, iTransformer, and PatchTST. Appendix I.2 extends this evaluation to GWNet, MTGNN, and TWIST, which already contain interaction mechanisms.

Table 20 reports the MAE of each original architecture and its ERST module variant:

DatasetGWNetMTGNNTWIST
PEMS0314.61 → 14.1814.91 → 14.6615.11 → 14.91
PEMS0419.35 → 18.2119.17 → 18.5418.51 → 18.05
PEMS0721.22 → 19.8521.02 → 19.9619.70 → 19.40
PEMS0815.06 → 14.1815.06 → 14.5314.21 → 13.96
KnowAir14.89 → 13.8514.85 → 13.9919.92 → 13.85
SDWPF96.88 → 92.6494.19 → 92.83140.47 → 92.64

Each entry shows Original → + ERST Module; lower MAE is better. Adding the module reduces MAE for all three architectures on all six datasets. Across the seven architectures in Table 20, MAE decreases in 41 of 42 comparisons, with one tie. These results evaluate the benefit of adding the ERST interaction module to forecasting architectures that already model interactions. See Section 4.4, Figure 4, and Appendix I.2, Table 20.

Q5 — Attention-based baselines. Table 1 and Appendix F include GMAN, PDFormer, and STAEformer as attention-based forecasting baselines. They are evaluated under the same dataset protocols as the complete ERST forecasting model. Table 1 reports the following MAE:

DatasetGMANPDFormerSTAEformerERST
PEMS0314.6215.0115.3014.28
PEMS0720.0120.3219.5319.51
KnowAir14.7515.1814.8213.98
SDWPF113.84102.89104.0593.08
NYCTaxi-In13.7513.5113.7413.43
T-Drive-In18.1117.9017.2616.56

ERST obtains lower MAE than all three attention baselines on these six datasets. The margins vary by dataset and model; for example, ERST and STAEformer differ by 0.02 MAE on PEMS07. ERST does not have the lowest MAE against these baselines on every dataset; on PEMS08, STAEformer and PDFormer obtain lower MAE (13.62 and 13.67, respectively, versus 14.08 for ERST; Appendix F, Table 8). See Section 4.1, Table 1, and Appendix F for the evaluation protocols and additional results.

Reviewer_UyqZOpenReview ↗

Increase my score

Thanks for authors hardworking and careful reply. All of my concerns are addressed. Learning entity's bias compared with normal value via learning a response is interesting and make sense. I will improve my score to 4. Good luck!

We sincerely thank the reviewer for taking the time to carefully reconsider our submission and update the evaluation. Your thoughtful comments and constructive suggestions have been very helpful in improving the clarity and presentation of our work. We greatly appreciate the opportunity to address your concerns and further strengthen the manuscript through the discussion and revision process. Thank you again for your valuable feedback, careful consideration, and continued engagement with our work.

Reviewer_oqy2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a model for spatiotemporal forecasting. The main idea is to learn interactions among entities by replacing one entity at a time to observe the response from the system. The paper shows that such interactions help improve the forecasting accuracy. With a comprehensive evaluation, the authors showed that ERST improves MAE in 23 of 24 combinations of forecasting architecture and dataset and matches the baseline in the remaining case.

Critical strengths

S1. The main contribution of the paper is a mechanism to discover interactions among entities. ERST provides a simple but effective approach to identify interactions that works well in practice.

S2. The ablation section of the paper is well organized, showing the origin of ERST’s performance.

S3. The update module discussed in Figure 3 is perhaps the most interesting component of the proposed model. It works well, as also demonstrated in the ablation studies.

S4. Figure 4 shows that the core method applies to other forecasting models as well.

Critical weaknesses

W1. While the writing of the methods and results is great, the introduction could be improved. A complex figure on the first page of the paper needs to be explained adequately. For example, I was not sure what “Aggregation” means in that example. Is it a specific paper or just naïve aggregation that nobody uses in practice? What are those values on the axis in Figure 1(a)? What do you mean by a response in Figure 1(b)? I needed to read the method section and then come back to the introduction to understand those figures.

W2. Some graph-based methods found that the interaction among entities should be sparse because it is unlikely that all entities influence a target entity. Sparsity can often help us understand who is influencing more, giving some notion of explainability.

W3. The main forecasting horizons presented in this paper are small compared to other papers. For example, FaST (KDD ‘26) forecasts the next {48,96,192,672} horizons based on observations from the 96 historical time steps for the LargeST benchmark. By contrast, this paper only predicted the next 12 time steps, which is significantly simpler than long-horizon forecasting. The appendix only shows results for up to 60 forecasting steps for two datasets, but this still falls short of the current evaluation standard used in other papers.

Questions for authors

Q1. If possible, please report long-horizon {48, 96, 192, 672} results for the LargeST benchmark. Comparison with just one recent baseline (e.g., Fast (KDD ‘26)) should be sufficient. For example, the FaST paper (https://arxiv.org/pdf/2601.05174) reported a MAE value of 20 for the CA dataset with a 96-step forecasting horizon. However, this paper reports a MAE of 23.95 for the same method and dataset for a 12-step forecasting horizon (Table 2). Please explain these discrepancies as well.

Q2. Do you observe any sparsity in the interactions?

Q3. It is not clear how tau_t is obtained in Equation 1. The previous paragraph mentioned a scalar for every time step. How was the D-dimensional vector obtained?

3 comments
AuthorsOpenReview ↗

Response to Reviewer oqy2: Clarifications and Additional Experimental Evaluations

We thank the reviewer for the detailed questions and suggestions. We clarified Figure 1 and added an analysis of the predictive contribution of entity-specific responses.

W1 — Meaning of aggregation and Figure 1. "Aggregation" refers to combining entity representations into a representation F(H), not to a particular forecasting baseline. Figure 1(a) compares observed PEMS08 traffic flow for the same entities in two training observations: Case A on the x-axis and Case B on the y-axis. The plotted values are observed flows, not learned representations or outputs of F. The cases were selected to have similar mean flows (296.48 and 296.66) but different values across entities; the six marked entities have the largest observed differences. This illustrates that similar observed means can accompany different entity values. It does not establish information loss in a learned aggregation model.

In ERST, the response for entity i is the difference between the outputs of the same aggregation function before and after replacing that entity's representation with a learned reference; all other entities, normalization, and the function remain unchanged (Eq. 7). Panels (b–c) use Case B and the first interaction layer of a trained ERST model. Panel (b) plots the L2 norm of each response divided by the mean response norm (dashed line = 1); entities are ordered by their observed differences between Cases A and B. Panel (c) replaces each response with the mean response vector in the same trained model, then recomputes the update term. The plotted value is the L2 difference between the original and recomputed updates, divided by the mean original update norm. Panel (c) measures an update difference, not forecasting error. We revised the Introduction and Figure 1 caption; the calculations are given in Appendix I.1, Eqs. (62)–(63).

W2/Q2 — Sparsity, entity participation, and response contribution. As clarified in Section 2.3 and Appendix I.5, ERST defines interactions through entity-specific changes in an aggregated representation, rather than a matrix of pairwise edge weights. Thus, ERST does not directly estimate sparse graph connections. We instead examine whether entity-level changes are concentrated, how restricted participation affects forecasting, and whether individual responses differ in predictive impact.

First, Table 5 shows that the top 10% of entities account for 26.75%–35.31% of standardized observed changes across 11 datasets, and 80.52% on CHIBike-In. These results describe concentration in observed data, not learned interaction sparsity.

Second, Appendix H.2 evaluates interaction depths L=2–9 under randomly sampled entity participation ratios of 10%, 30%, 50%, 75%, and 100%. Under 10% participation, the lowest MAE occurs at L=5 for PEMS08 (14.33) and T-Drive (18.07), and L=4 for SDWPF (94.20), compared with 14.43, 19.79, and 96.71 at L=2. However, full participation at L=2 achieves lower MAE on all three datasets (14.08, 16.56, and 93.08). This shows that the preferred depth depends on participation, without implying that sparse pairwise relations are learned.

Third, Appendix I.5 directly examines predictive sensitivity to response removal. We use fixed, trained two-layer ERST models and 24 evenly spaced test windows each from T-Drive and PEMS08. At inference, selected first-layer responses are set to zero before response projection, while model parameters remain fixed and the second layer operates normally. The Top 10% condition ranks entities by the change in their first-layer update term when the response is set to zero (Eq. 64), without using future targets or forecasting errors. It removes 103 responses on T-Drive and 17 on PEMS08. Random 10% removal uses the same counts and three repetitions. All conditions use the same checkpoint and test windows within each dataset.

Condition (MAE)T-DrivePEMS08
Full ERST19.7213.60
Random 10% removal20.7914.13
Top 10% removal26.0114.21
All first-layer responses removed26.1422.75

On T-Drive, Top 10% removal increases MAE by 6.29, compared with 1.07 for random removal. On PEMS08, the corresponding increases are approximately 0.61 and 0.52. Removing all first-layer responses increases MAE by 6.41 and 9.15. These results indicate that the effect of response removal varies across entity selections and datasets. The values are diagnostic MAEs over 24 test windows, not full-test benchmark scores.

Together, these analyses distinguish observed change concentration, partial response-update participation, and entity-specific predictive contribution from learned pairwise graph sparsity. See Table 5, Appendix H.2, Figure 7, Table 19, and Appendix I.5, Eq. (64), Table 22.

Q3 — Time embedding in Eq. (1). The scalar mentioned before Eq. (1) is the projected observation at each time step, x̄_(τ,i), not the time embedding. The sequence of these scalar observations is projected to the D-dimensional history vector q_i. Independently, the last time step of the input window determines a time-of-day slot, whose learned row is selected from E_tod ∈ R^(M×D). This row is the D-dimensional embedding τ_t, where M is the number of time slots per day. Equation (1) concatenates q_i and τ_t before applying g_temp and layer normalization. Therefore, τ_t is obtained by embedding lookup, not by expanding a scalar observation. See Section 3.1, Eq. (1).

W3/Q1 — LargeST protocol and long-horizon forecasting. Our main LargeST evaluation follows the established protocol of the original LargeST benchmark (Liu et al., 2023b). The benchmark uses 12 historical steps to predict the next 12 steps, with traffic observations at 15-minute intervals and a chronological 6:2:2 train/validation/test split. The original LargeST paper evaluates multiple forecasting baselines under this protocol (Section 5.1). Accordingly, our Table 2 evaluates the listed baselines and ERST under the same 12→12 setting on LargeST-CA and LargeST-GLA. This maintains consistency with the original benchmark and allows comparisons under a common protocol. The 12→12 setting was therefore selected to follow the established LargeST evaluation, not as a substitute for long-horizon forecasting.

Additional evaluation under the FaST protocol. To address the reviewer's request, we evaluated ERST on LargeST-CA using FaST's data and evaluation protocol (Zhao et al., 2026). We used the 2019 California traffic-flow data (8,600 sensors, 15-minute intervals), chronological sliding-window samples with a 6:2:2 train/validation/test split, and 96 input steps with prediction horizons of 12, 48, 96, 192, and 672 steps. Normalization statistics were computed from the training data only, and MAE was evaluated after inverse normalization, excluding zero-valued targets. ERST retained its original architecture and training procedure while matching FaST's data and evaluation settings.

FaST results are taken from Table 6 of its paper, while ERST results are obtained from our additional rebuttal experiments under these matched forecasting settings.

LargeST-CA MAE (96 input steps)124896192672
FaST15.7819.5720.9022.8725.35
ERST15.4519.5520.9023.5727.18

Under these common forecasting settings, ERST achieves a lower reported MAE at 12 steps (0.33) and a nearly identical MAE at 48 steps (0.02 lower). Both methods report the same MAE of 20.90 at 96 steps to two decimal places. At 192 and 672 steps, FaST achieves lower MAE by 0.70 and 1.83, respectively. Thus, ERST remains competitive at the shorter and intermediate horizons, while FaST performs better at the two longest horizons. These results provide the requested long-horizon comparison without claiming consistent superiority across prediction lengths.

Clarification of the FaST results. The reviewer notes the difference between FaST MAE 23.95 in our Table 2 and MAE 20.90 reported in FaST's paper. These values come from different evaluation protocols. Our Table 2 follows the original LargeST 12→12 setting, whereas FaST's reported 20.90 corresponds to 96→96. At 15-minute intervals, the input windows cover 3 hours and 24 hours, respectively, and the prediction horizons also differ. Longer observation windows can improve forecasting accuracy by providing more historical context, as demonstrated by PatchTST (Nie et al., 2023, Appendix A.4.2). Conversely, longer prediction horizons can make forecasting more challenging, as illustrated by FaST's results under a fixed 96-step input (Zhao et al., 2026, Table 6). For example, FaST reports MAE 15.78 for 96→12 and 20.90 for 96→96. However, these effects are model- and dataset-dependent, and the difference between 23.95 and 20.90 cannot be attributed to either factor alone. Thus, the two scores should not be directly compared as evidence of model performance under the same conditions.

The additional experiments follow FaST's forecasting evaluation setting, allowing comparisons at each requested prediction horizon. The original LargeST results in Table 2 remain comparisons under the established 12→12 benchmark protocol. Our earlier long-horizon evaluations on PEMS04 and PEMS08 (Appendix G, Tables 12–13) provide additional results but do not replace the LargeST evaluation requested by the reviewer. See Section 4.1, Table 2, Appendix G, LargeST (Liu et al., 2023b, Section 5.1), and FaST (Zhao et al., 2026, Table 6).

I thank the authors for their rebuttal. I appreciate the response related to Q2 and Q3, as they clarified my understanding. However, the additional results comparing FaST and ERST showed that some of the benefits shown in the results section might have originated from the specific experimental protocols. I thank the authors for their tremendous effort. I will keep my current score.

Reviewer_fc1COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a new method for modeling interactions among entities for spatiotemporal forecasting. Here, interactions for a particular entity are defined as the change in representation when the entity's representation is replaced by a reference. Rather than using relations between entity representations or aggregation of representations, their method ERST considers the change in response when an entity's representation is replaced by a reference. The ERST interaction is integrated into the forecasting representation. The authors evaluate their method on several datasets covering tasks including urban mobility, air quality and wind power prediction. The authors compare their model against several spatiotemporal forecasting methods and find that ERST outperforms the other methods

Critical strengths

The paper seems to present a promising method, looking whether using responses to changing a given entity's representation with a reference provides information beyond the initial representation itself, and the results seem promising. The authors also present some interesting ablation studies have been considered.

Critical weaknesses

The paper is unequal in clarity in different sections. I would recommend the authors add a bit more context on interactions and different definitions to make it a bit more accessible and clearer. The structure of the paper is clear and one can broadly follow the paper, but some parts would benefit being more precise . For example, in 2.3 it's not really clear how "responses" for a given entity would normally be defined, how one actually integrates investigating different contributions of different entities in other methods. Generally, some parts lack explanation of the intuition or the rationale behind certain design choices (in particular how to integrate the response (3.3) and the choice of datasets and models compared against (see questions)) This makes it difficult to understand why the authors chose these particular datasets and models to evaluate their modelson, when there are more widely used benchmarks for spatiotemporal forecasting. This ius not a problem in itself but it would require some minimal justification of the focus. The vagueness of the language used in my review reflects what can be understood from the paper.

I would have liked to see additional interpretation of results too. For example, when the authors find that as the number of interaction layers increase, the performance decreases, what is their hypothesis for that?

Questions for authors
  • Could the authors explain the idea behind integrating the response into the forecasting representation in the way described in 3.3?
  • Can the authors elaborate on their choice of baselines and evaluation datasets? This is not entirely a criticism, the authors use datasets spanning different applications and look into a variety of models. But I was wondering why there was no overlap with methods and datasets of established benchmarks like PredBench for example (PredBench: Benchmarking Spatio-Temporal Prediction across Diverse Disciplines)
  • That being said why have the authors not used methods that were the ones presented in the dataset papers as comparisons? ()e.g. PCDCNet for KnowAir) I may be missing something but it makes it a bit difficult to assess the benchmark if the methods that were developed specifically for certain datasets were not included.
  • Can the authors explain further why they ran the response shuffle experiment? Does it not tell more about the relationships between different entities than whether responses are interchangeable? I don't understnad why in the first place, one would suspect that the responses are not associated with the entities they were computed for...

Suggestions: figure captions could be a bit more complete (e.g. Fig 1 ) to improve paper readability. I found that the accompanying in text description also did not fully enable the reader to understand figure 1

5 comments
AuthorsOpenReview ↗

Response to Reviewer fc1c: Clarifications and Additional Analyses

We thank the reviewer for the comments. We revised the explanations of interaction, response integration, dataset selection, and the evaluation protocols.

Critical Weaknesses. Section 2.3 now distinguishes the operations being compared. Graph models propagate information through defined or learned relations, while attention assigns weights to and combines entity representations. Aggregation forms a representation from multiple entities. These approaches can preserve entity information; ERST does not claim otherwise. ERST measures the change in the aggregated representation before and after replacing one entity representation with a reference, while keeping the other entities, normalization, and aggregation function fixed (Eq. 7). The response is not a prediction or a causal measure.

Section 3.3 explains why ERST retains the entity representation and uses the response to construct a gated residual update, rather than replacing the entity representation with the response. Modulation uses entity, time, and layer information; gating controls the update while retaining the existing representation. Section 4.1 and Appendix E clarify the choice of datasets, forecasting protocols, and baselines. For depth, each ERST layer computes responses from entity representations updated by preceding layers. Increasing depth repeats the response computation and update operations on these updated representations. Xiao et al. (2025) reported less stable training with three layers than with two, while Shi et al. (2026) found that two attention layers performed best on PEMS04 and PEMS08. Figure 7 shows that increasing depth does not consistently reduce MAE. Under full entity participation, L = 2 achieves the lowest MAE on PEMS08 and SDWPF and matches the lowest on T-Drive, supporting its selection as the default (Appendix H.2). We also revised Figure 1 to distinguish observed flow, response norms, and update differences (Appendix I.1).

Q1. The response measures a change in the aggregated representation, rather than the full state of an entity. Replacing the entity representation with this response could discard information about the entity's existing state. ERST therefore retains the entity representation and uses the response to construct a residual update. Modulation conditions the response on entity, time, and layer information. The candidate update is computed from the entity representation and modulated response, while the gate controls its contribution. This design allows response information to influence forecasting without replacing the existing representation. Table 3 reports:

DatasetERSTw/o Modulation and Gating
PEMS0314.2815.67
SDWPF93.0895.39
T-Drive16.5622.66

Removing both operations increases MAE in all three cases. This ablation tests their combined role, not their separate effects. See Section 3.3, Eqs. (11)–(14), Table 3, and Appendix C.

Q2. We evaluate ERST on 12 datasets across five domains and different spatial structures, system scales, and protocols. PeMS and LargeST contain sensor-based traffic data; NYCTaxi, CHIBike, and T-Drive follow the PDFormer grid protocol; KnowAir, SDWPF, and Milan cover other domains. Non-mobility datasets use 12→12 forecasting and a 6:2:2 split; mobility datasets use 6→1 and a 7:1:2 split.

We select recurrent, MLP, graph, attention, and recent spatiotemporal baselines to compare modeling mechanisms under each dataset's protocol.

PredBench-related comparison. PredBench (Wang et al., 2024) uses different datasets and forecasting protocols, so its published scores cannot be directly compared with ours. We additionally evaluated SimVPv2 (Tan et al., 2025), a video-oriented successor to the SimVP family evaluated in PredBench, on three grid-based mobility datasets using our 6→1 protocol.

Dataset (Inflow)ModelMAE ↓RMSE ↓MAPE ↓
NYCTaxiSimVPv214.1124.6014.91
NYCTaxiERST13.4322.6912.97
CHIBikeSimVPv24.025.7630.30
CHIBikeERST3.525.3624.35
T-DriveSimVPv217.4631.1315.40
T-DriveERST16.5629.0513.44

ERST achieves lower errors on all nine metrics. SimVPv2 expects regular 2D grids; applying it to irregular PeMS and LargeST sensor networks would require spatial transformations. We therefore restrict this comparison to grid data. This is one representative comparison, not a claim of superiority over all PredBench methods. We will incorporate these results and implementation details into the revised manuscript within the rebuttal period.

References:

Wang et al., PredBench: Benchmarking Spatio-Temporal Prediction across Diverse Disciplines, ECCV 2024, doi:10.1007/978-3-031-73636-0_17.

Tan et al., SimVPv2: Towards Simple Yet Powerful Spatiotemporal Predictive Learning, IEEE Transactions on Multimedia, 2025, doi:10.1109/TMM.2025.3543051.

Q3. We agree that comparisons with task-specific models are valuable. However, PCDCNet (Wang et al., 2025) was evaluated using a different dataset, forecasting horizon, and input protocol:

SettingERST on KnowAirPCDCNet
Entities184 cities355 monitoring stations
Interval3 hours1 hour
PredictionPM2.5, 12→12PM2.5 and O3, 72 hours
Available dataKnowAir observations and weather attributesPollutants, meteorology, emissions

PCDCNet uses meteorological variables and emission inventories, including future-step inputs, to model pollutant transport and formation (Eq. 1 and Algorithm 1). Our KnowAir protocol does not provide the forecast-step meteorological and emission inputs required by its published formulation. Therefore, directly comparing its reported results with ours would be inappropriate. We have not evaluated PCDCNet under our protocol and do not claim superiority over it.

Additional AirFormer evaluation. To further address the reviewer's concern regarding domain-specific baselines, we additionally evaluated AirFormer (Liang et al., 2023), a model specifically developed for air quality forecasting, on KnowAir under our 12→12 forecasting protocol with a 6:2:2 split.

ModelKnowAir MAE ↓RMSE ↓
AirFormer14.2122.59
ERST13.9820.50

ERST achieves lower MAE and RMSE than AirFormer under this evaluation. This additional comparison strengthens our evaluation against a domain-specific air quality model, although it does not replace a direct PCDCNet comparison. We will incorporate these results and the corresponding experimental details into the revised manuscript within the rebuttal period.

References:

Wang et al., PCDCNet: A Surrogate Model for Air Quality Forecasting with Physical-Chemical Dynamics and Constraints, arXiv:2505.19842, 2025.

Liang et al., AirFormer: Predicting Nationwide Air Quality in China with Transformers, AAAI 2023, doi:10.1609/aaai.v37i12.26676.

Q4. Response Shuffle tests whether forecasting depends on assigning each response to the entity for which it was computed. Responses are entity-specific by construction; the experiment does not test whether this definition holds. At inference, we keep the trained model parameters fixed, reassign the computed response vectors across entities, and recompute the updates and forecasts. This changes response assignment while preserving the response values and model parameters. Examples from Table 21 are:

Dataset / ArchitectureOriginal ArchitectureERST ModuleERST + Response Shuffle
PEMS07 / LSTM30.3120.4526.93
PEMS08 / STID15.6514.2816.04
KnowAir / iTransformer15.3115.0415.28
SDWPF / PatchTST96.7295.1696.76

Relative to the corresponding ERST Module model, Response Shuffle increases MAE in all 24 comparisons across four architectures and six datasets. Some shuffled models still outperform the original architectures, so the relevant control is the same trained model with and without shuffled responses. These results show that forecasting performance depends on the correspondence between entities and their assigned responses, rather than response values alone. See Section 4.5, Figure 5, and Appendix I.3, Table 21.

Figure 1. We revised the figure caption and text. Panel (a) shows two PEMS08 training observations selected for similar mean flows (296.48 and 296.66) but different entity values: Case A (x-axis) and Case B (y-axis). The plotted values are observed flows, not model outputs. The six marked entities have the largest observed differences. Similar observed means do not establish information loss in learned aggregation. Panels (b)–(c) use Case B and the first ERST interaction layer. Panel (b) plots each response L2 norm divided by the mean response norm (dashed line = 1). Panel (c) replaces each response with the mean response vector in the same trained model and plots the L2 change in the update divided by the mean original update norm. The plot concerns the update, not prediction error. See Figure 1 and Appendix I.1, Eqs. (62)–(63).

Update on Q2 and Q3: Additional Baseline Evaluations

We sincerely thank the reviewer for the constructive suggestions regarding baseline selection.

Our original evaluation prioritized established forecasting methods under the corresponding dataset protocols, rather than exhaustive comparisons with video prediction or domain-specific architectures. Moreover, the published PCDCNet setting differs from our KnowAir protocol in its dataset, prediction horizon, and required future-step inputs.

To further address the reviewer's concerns, we conducted additional evaluations with SimVPv2 on three regular-grid mobility datasets and AirFormer on KnowAir, using our respective forecasting protocols.

We have updated our Q2 and Q3 responses with the detailed results and will incorporate them into the revised manuscript within the rebuttal period. We appreciate the reviewer's suggestions, which helped strengthen our experimental evaluation.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Manuscript Revision and Responses to Reviewers

We thank all reviewers for their constructive feedback.

We have uploaded the revised manuscript and posted detailed responses to each reviewer. The revision includes clarifications of the motivation, response mechanism, learned reference, and evaluation protocols, along with additional controlled experiments and forecasting backbone comparisons.

We have also clarified the scope of our contributions and provided further explanations to address the reviewers' questions and concerns.

We appreciate the reviewers' time and consideration.

Reviewer_UyqZOpenReview ↗

Highlight the revised parts

Hi,

Thanks for your careful reply. To able us better locate the updated results in the paper, could you please using a different color, such as blue, to highlight those revision parts? Much appreciated.

AuthorsOpenReview ↗

Revised Manuscript with Highlighted Changes

Thank you very much for your kind follow-up and helpful suggestion.

Following your request, we have carefully highlighted the revised portions of the manuscript in blue to make the changes easier to identify and review.

We have also made a concerted effort to incorporate your constructive feedback throughout the revision, including clarifications of the motivation, the ERST interaction mechanism, and the learned reference, as well as additional controlled experiments and evaluations across different forecasting architectures.

We sincerely appreciate your careful reading, valuable suggestions, and continued consideration of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.