acceptodds
Under review as a conference paper at ICLR 2027

Causality-Inspired Spatio-Temporal Attention for Forecasting on Evolving Graphs

Abstract

Urban forecasting is particularly critical when infrastructure changes, as road-network and communication entities may be added, removed, or disrupted, altering graph topology and data distributions. Most spatio-temporal forecasters assume fixed node sets and connectivity. Newly activated entities lack sufficient long-term observations, while frequent retraining may be impractical for time-sensitive traffic control and emergency recovery. We therefore study multi-step forecasting on evolving graphs and propose DyCSTA, a spatio-temporal forecasting framework for multi-step prediction on evolving graphs. DyCSTA retains usable history for persistent nodes, models newly arriving nodes from their current observations, reconstructs dependencies on the current graph, and uses latent context to assist prediction. Its predictive representation split separates forecast-relevant dynamics from context-related variation, while a model-defined replacement branch regularizes context conditioning during training. Experiments on traffic, communication, and other evolving systems evaluate changing node sets, new-node forecasting, and all prediction horizons. DyCSTA achieves its strongest results on the two flow datasets and remains competitive in the other domains.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatio-Temporal Forecasting, Evolving Graphs, Dynamic Graph 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_gtfUOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies multi-step spatio-temporal forecasting when nodes may arrive or disappear and graph topology changes across stages, and proposes DyCSTA. The method uses a dynamic node-evolution module to exploit historical information from persistent nodes while handling newly arriving nodes from their current observations. It then performs temporal and spatial modeling on the current graph, decomposes representations into forecast-relevant and contextual components, and conditions prediction on latent environment prototypes with a context-replacement training objective. Experiments span traffic, communication, air-quality, and energy datasets, with analyses on overall forecasting, new-node prediction, forecast horizons, component ablations, and representation separation. The strongest gains are reported on the two flow datasets, while performance remains competitive in the other domains.

Critical strengths

S1. The paper goes beyond standard forecasting under a fixed node set by jointly considering node arrivals, departures, and connectivity changes. Deciding what historical information should be retained for persistent nodes, how to forecast newly arriving nodes with limited history, and how to rebuild dependencies after topology changes are practically relevant problems in evolving infrastructures.

S2. The experiments go beyond aggregate MAE/RMSE/MAPE and include analyses of existing versus newly arriving nodes, different forecast horizons, and forecast-relevant versus contextual representations. In particular, the new-node and long-horizon analyses directly relate to the paper’s motivation of topology evolution and limited-history forecasting.

Critical weaknesses

W1. The paper states that persistent-node stability is measured using Wasserstein distance between empirical distributions from adjacent stages, with stable nodes prioritized for the evolution pool and high-activity new nodes filling the remaining slots. However, the released ENERGY-Wind code directly sets all current nodes as top_nodes and keep_nodes, without the described Wasserstein-based ranking or activity-based selection. This makes it unclear whether the central dynamic-evolution mechanism was actually used to obtain the reported results, directly affecting soundness, reproducibility, and claim-evidence consistency.

W2. The paper motivates evolving-graph forecasting partly by arguing that frequent retraining may be impractical, but reports no comparison of training time, inference latency, peak memory, parameter count, or scaling behavior. Since DyCSTA combines GRU-based evolution, temporal attention, graph propagation, and latent-context inference, its computational overhead is not obviously negligible. This weakens the practical deployment motivation.

Questions for authors

Q1. Were the Table 1 results actually obtained using the Wasserstein-based stability selection described in the paper? If so, could the authors explain why this mechanism is absent from the released ENERGY-Wind implementation and provide the corresponding implementation? If not, please clarify which experiments used a simplified evolution mechanism.

Q2. Could the authors provide efficiency comparisons with the main dynamic/continual baselines, including some combination of training time, inference latency, peak GPU memory, and parameter count, together with at least one scaling analysis over graph size?

Q3. How should the term “causality-inspired” be interpreted in this work? The current matched-probe analysis mainly shows that CC and TT have different predictive utility, but does not establish that they correspond to causal versus contextual factors.

4 comments

Responses to Question 1: We apologize for the confusion caused by our accidental release of the full‑node‑pooling version in the shared files. And, the corrected code version has been uploaded. For the main experiments on PEMS‑Stream, Air‑Stream and Energy‑Stream, we adopt Wasserstein‑based stable node selection. For JiNan, since the number of active nodes is small, all active nodes at each stage are used. We have complemented the above remarks in Section 5.2 of the revised manuscript.

Responses to Question 2: According to this comment, we evaluated the prediction performance and resource overhead of the six approaches on PEMS. DyCSTA handles changes in node sets and connectivity through forward state evolution and current‑graph encoding, without additional gradient‑based retraining.. (Detailed experimental settings and discussions are provided in Table 8, Appendix A.10)

MethodParameter countMAE ↓RMSE ↓Batch=1 median / p95 (ms) ↓Batch=32 throughput (windows/s) ↑PyTorch peak allocated / reserved (MiB) ↓Device‑memory increase (MiB) ↓
DyCSTA44303714.6025.0015.92 / 16.13260.275096.25 / 1185823473
EAC860619.2431.701.69 / 2.0324470.4434.30 / 701900
STBP73381616.6427.805.69 / 6.325396.02264.69 / 7702393
DySubTree6408417.6229.8516.65 / 18.9261.9756.21 / 1565505
ScaDyG10542020.9235.0517.86 / 18.45741.281549.62 / 21483143
DynaSTy14930516.3927.8165.99 / 116.89153.321322.52 / 140823691

We further conduct experiments on EAC and STBP after amortizing stage‑wise costs, and the results are shown in Table 9, Appendix A.10.

In order to implement scaling analysis, we have further complemented the following experiment (part of the table is shown below; the full version and related discussion are provided in Table 10, Appendix A.11). Throughout the experiments, the Batch=1 latency of DyCSTA remains approximately 16 ms, indicating that the single‑window inference latency stays stable as nodes and edges grow stage by stage.

Graph scale2011→2017
Active nodes655 → 715 → 786 → 822 → 834 → 850 → 871
Edges3154 → 3858 → 4632 → 5072 → 5188 → 5382 → 5576
Method2011201220132014201520162017
Batch=1 forward latency (ms)
DyCSTA15.9215.8415.8415.8815.9515.7415.92
EAC1.201.231.191.171.181.191.17
STBP3.693.623.693.623.633.583.59
DySubTree13.0813.3815.2914.8215.8616.3616.65
ScaDyG17.9917.5518.0417.6917.8717.7417.86
DynaSTy47.6946.8062.9062.3562.9165.6565.99

Responses to Question 3: Thank you for the reviewer’s comments. First, we agree that the matched‑probe analysis does not establish that CC and TT recover causal and contextual factors of the underlying data‑generating process. Actually, we initially planned to follow our prior line of reasoning from our prior work and build the model based on SCM causal learning. However, preliminary experiments yielded unsatisfactory results. We therefore retain the valuable idea of decomposing target‑related predictive information from disturbance caused by environmental changes, and adopt an alternative technical route to realize this idea under the setting of dynamically changing nodes. Correspondingly, we revise “Causality” in the title to “Causality‑Inspired”, to indicate that this work is inspired by causal thinking rather than adopting rigorous causal modeling methods. We recognize that several statements in the manuscript still overemphasize causal properties and may lead to misunderstanding. We have carefully examined the entire paper and rephrased potentially misleading descriptions.

Thank you for the clarification and the updated code. Could you provide a quantitative comparison of the Energy-Stream results obtained using the original full-node-pooling implementation and the corrected Wasserstein-based selection? This would help clarify whether the implementation difference affects the reported results.

Thank you for the reviewer’s comments.

To provide a quantitative comparison, we have implemented the following experiment, where Energy‑Stream results were obtained by employing the full‑node‑pooling implementation and the Wasserstein‑selection protocol, respectively.

Pooling implementationMetricH3H6H12Avg.Mean training / epoch (s)Mean inference / epoch (s)
Original full‑node poolingMAE0.961.312.151.44104.817.59
RMSE1.331.762.731.89
MAPE (%)9.6514.1327.0417.04
Wasserstein selectionMAE0.921.181.791.26102.667.18
RMSE1.311.672.411.76
MAPE (%)10.8913.2317.4113.51

The Wasserstein selection gives lower MAE and RMSE at all reported horizons and lower average MAPE, with the largest gains at the 12‑step horizon. It also has the shorter recorded mean training and inference times. The original full‑node implementation has a slightly lower MAPE at H3, while the Wasserstein‑selection protocol is better for the longer horizons and the average score.

Thanks for your rebuttal. My questions have been addressed, and I’ll raise my score.

Reviewer_VFvNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes DyCSTA for multi-step forecasting on evolving graphs. It transfers persistent-node states, models new nodes from current observations, and uses latent context prototypes for prediction. Results are strong on PEMS-Stream and JiNan.

Critical strengths
  1. The paper jointly considers node arrival, node departure, connectivity changes, and multi-step forecasting. The distinction between retained states for persistent nodes and current-window initialization for arriving nodes is well motivated.

  2. DyCSTA consistently outperforms the listed baselines on PEMS-Stream and JiNan across MAE, RMSE, and MAPE, demonstrating its effectiveness on evolving flow graphs.

  3. The paper includes component ablations, separate evaluation of existing and new nodes, horizon-wise comparisons, parameter sensitivity, and matched-probe analyses. These experiments provide more insight than a single aggregate benchmark table.

Critical weaknesses
  1. The stability-aware node selection may use future test observations. Equation (3) ranks persistent nodes using empirical distributions �0��0� and �1��1�. Appendix A.1 states that these distributions are built from valid observations in adjacent stages, while the final snapshot of JiNan, Air-Stream, and Energy-Stream is entirely reserved for testing. The paper does not state that $(p_v^t)$ is computed only from observations available before each forecast. If the complete test-stage distribution is used, the dynamic-evolution module has access to future target information. Because this selection is central to DyCSTA, the evaluation cannot be interpreted reliably until the construction is clarified or rerun causally.

  2. The causality-inspired decomposition is not convincingly validated. The replacement branch is model-defined rather than derived from an intervention or an identified causal model. More importantly, the factual path uses the tanh gate in Equation (12), whereas Section 4.5 states that the replacement path uses a sigmoid gate. The two predictions therefore differ in both the context code and the gating function, so a larger replacement error cannot be attributed specifically to context replacement. The matched-probe result also mainly confirms that the branch trained for forecasting is more predictive than the branch assigned a contextual role; it does not establish that the two representations capture causal and contextual factors.

Questions for authors
  1. Which exact timestamps are used to construct �0��0� in Equation (3) for the held-out test snapshot? Can the authors report results when it is computed strictly from the causal prefix available before each prediction?

  2. What happens if the factual and replacement paths use the same gating function? A matched-gate ablation would isolate the effect of replacing the context code from the tanh-versus-sigmoid difference.

  3. New nodes are evaluated after a complete 12-step input window is available. How does performance change when a node has only 1, 3, or 6 observations after activation?

  4. For each baseline, what training or parameter-update operations are allowed when moving to a new graph stage? A stage-wise protocol and computational comparison would clarify whether all methods are evaluated under the same no-retraining constraint.

4 comments

Responses to Question 1: In the experiments, we are careful not to leverage future test observations. Equation (3) uses training snapshots only: the first 7 days for PEMS‑Stream, the first 3 months for Air‑Stream, and the first 30 days (4,320 ten‑minute observations) for Energy‑Stream. The final training‑selected pool is reused in validation and testing. Since the JiNan dataset has a small number of nodes, we directly adopt all active nodes at each stage without applying Equation (3). Thus, node selection is based strictly on the prefix available before prediction, satisfying the causal‑prefix computation requirement.

Responses to Question 2: According to this comment, we trained tanh/tanh and sigmoid/sigmoid matched‑gate variants on PEMS‑Stream. The proposed tanh/sigmoid combination obtains the lowest MAE and RMSE, lower mean training and validation‑inference time. (Tables 14, Appendix A.13)

Gate setting (factual / replacement)MAE ↓RMSE ↓MAPE (%) ↓Mean training time (s) ↓Mean validation‑inference time (s) ↓
tanh / sigmoid (ours)13.9724.4116.49750.6146.78
tanh / tanh14.0024.6616.82772.0248.37
sigmoid / sigmoid13.9924.5316.47795.3464.39

Responses to Question 3: According to this comment, we added a fixed‑target paired evaluation and an arrival protocol (Detailed experimental settings and discussions are provided in Tables 12 and 13, Appendix A.12). The arrival protocol compares DyCSTA with persistence after 1, 3, 6, or 12 observations. DyCSTA has lower MAE and RMSE than persistence at every observation length, confirming short‑history new‑node adaptation.

Observations after activationDyCSTA MAEDyCSTA RMSEPersistence MAEPersistence RMSE
13.354.794.816.20
33.014.464.926.80
63.244.854.836.68
122.843.924.256.20

DyCSTA has lower MAE and RMSE than persistence at every observation length, confirming short‑history new‑node adaptation.

Responses to Question 4: According to this comment, we supplement a comparison table describing the operations of all methods upon entering a new‑graph stage (Due to character limits, only part of the tables are shown below; the full version and related discussion are provided in Table 6 and Table 7, Appendix A.10)

MethodParameter update at new graph stagePermitted forward‑state or graph operationsUses future test targets
DyCSTANo parameter updateCurrent graph, training‑selected node pool, input‑driven propagation weights and cachesNo
STBPUpdate pattern bankCurrent‑graph preparation and pattern‑bank expansionNo
DySubTreeNo parameter updateCurrent graph and recurrent‑memory updatesNo

And, we have further provided a stage‑wise computational comparison as follows. It should be noted that DyCSTA enables multi‑step prediction on the graph without stage‑wise parameter updates.

MethodGraph preparation time (s)Median batch=32 forward latency (ms)Windows/sPeak device‑memory increase (MiB)
DyCSTA0.03121.12264.207588
STBP0.024.936491.661874
DySubTree0.06532.5360.091094

Thank you for the detailed response, which addresses most of my concerns. However, several key issues remain unresolved.

The additional gate ablation does not isolate whether the replacement-path performance difference is specifically caused by replacing the context code. Without an intervention, an identifiable causal model, or evidence of independence or identifiability, the two representations still cannot be interpreted as capturing “causal” and “contextual” factors, respectively.

For Question 3, DyCSTA is compared only with the persistence baseline, rather than with the main baselines or other methods capable of handling variable-length histories. The results therefore do not sufficiently establish DyCSTA’s relative advantage in adapting to new nodes with short histories.

For Question 4, compared with STBP, DyCSTA has approximately 24.6× higher inference latency, 24.6× lower throughput, and 4.0× higher peak memory usage. These results reveal substantial computational and deployment costs. The absence of stage-wise parameter updates does not necessarily imply lower overall deployment cost, and this trade-off should be discussed directly.

Given that these key concerns remain unresolved, I maintain my original score.

Thank you for the reviewer’s comments.

To address comment 1: We formulate two research questions: (1) Can the replacement path serve as an auxiliary training mechanism and improve model performance? (2) Do the context code used by the replacement path carry useful contextual information?

For Question (1), we design a controlled experiment by constructing a factual‑only variant. This variant retains the factual branch but removes the replacement branch during training. For Question (2), we retain the complete operations of the replacement branch during inference but replace its input context code with uninformative white noise. All variants use the same JiNan data split, input and output lengths, training budget, and optimization settings. The results show that DyCSTA outperforms FACTUAL‑ONLY on all three metrics, indicating that the replacement path provides valuable auxiliary supervision and improves the forecasting performance of the factual path. DyCSTA also outperforms the random‑noise replacement variant, showing that the benefit of the replacement path depends on informative context encodings rather than merely on adding an extra branch or introducing random perturbations. Therefore, the replacement path helps the factual path learn better predictive representations, while its context code carries valuable contextual information.

MethodFactual PathReplacement PathReplacement‑Path Context EncodingMAE ↓RMSE ↓MAPE ↓
DyCSTARetainedRetainedLearned context encoding221.31347.0729.34%
FACTUAL‑ONLYRetainedRemoved—244.21374.9333.66%
RANDOM‑NOISE‑REPLACEMENTRetainedRetainedStandard Gaussian white noise244.24371.0434.33%

To address comment 2: To provide a direct relative comparison for short‑history new‑node adaptation, we added ScaDyG, DynaSTy, DySubTree, STRAP, COIN‑GNN, EAC, and STBP under the same PEMS‑Stream arrival protocol.

The experiment uses the 21 nodes that are newly activated in the 2017 test stage relative to the preceding graph stage. When a new node has received 1, 3, 6, or 12 observations, respectively, the method makes one 12‑step forecast immediately. The input length and forecast length are both 12; unavailable early positions are filled with the first available observation. Persistence repeats the latest available observation for 12 steps. The columns 1, 3, 6, and 12 denote the number of observations available after node activation.

The results show that DyCSTA achieves the lowest MAE and RMSE among all listed methods at each of the four history lengths.DyCSTA also remains below the Persistence baseline at every observation length. This extended comparison directly supports DyCSTA's relative advantage in using short activation histories to produce a 12‑step forecast for newly arrived nodes.

MethodMetric13612
DyCSTAMAE3.353.013.242.84
RMSE4.794.464.853.92
PersistenceMAE4.814.924.834.25
RMSE6.206.806.686.20
ScaDyGMAE4.364.454.743.92
RMSE5.586.196.465.76
DynaSTyMAE17.7121.4917.253.06
RMSE21.2625.9821.304.24
DySubTreeMAE15.1815.8815.7716.70
RMSE16.2816.9316.9717.71
STRAPMAE17.5416.8915.0720.12
RMSE19.5419.2017.2720.77
COIN‑GNNMAE17.3319.2716.8319.57
RMSE18.5620.0517.7920.08
EACMAE13.4412.5912.4411.05
RMSE19.2118.2617.7515.23
STBPMAE7.337.487.717.69
RMSE8.739.019.078.79

To address comment 3: The aforementioned STBP is a continual‑learning model. While it enjoys fast single‑inference latency, it requires continuous parameter updates during operation. To comprehensively evaluate deployment costs, we devise the following experiments. Train on data from 2011 to 2015 and test over 8,905 windows from 2017 to compare the prediction performance and costs. Although STBP achieves fast single‑inference latency, it requires the adaptation procedure for the 2016 stage, which incurs substantially higher overall overhead. Meanwhile, its prediction performance is inferior to DyCSTA.

MethodMAERMSE2016 stage adaptation2017 graph/pool preparationBatch=1 median (ms)Per-window cost (ms)
DyCSTA14.6025.00No parameter update1.95 s15.9216.14
STBP16.6427.80715.15 s; 99 epochs0.04 s5.6986.00

Thank you for the additional response, which resolves many of my concerns. The new experiments show that the replacement branch and learned context code improve prediction. They do not, however, justify interpreting the two representations as causal and contextual factors. Removing the branch changes the training setup, while replacing its input with Gaussian noise merely shows that the model benefits from informative input. If the paper retains its causal interpretation, direct supporting evidence is still needed.

The short-history evaluation also remains ambiguous. Forecasts are made immediately after 1, 3, 6, and 12 observations, so the four columns correspond to different target periods. Their differences therefore cannot be attributed solely to history length. It is also unclear whether every baseline was trained using the same short-prefix construction and padding strategy. Otherwise, methods that encounter repeated-value padding only at test time may be unfairly disadvantaged.

The efficiency conclusion depends on the chosen workload of 8,905 windows. STBP incurs a one-time adaptation cost, but its per-window latency is 5.69 ms, compared with 15.92 ms for DyCSTA. Over a sufficiently long stream, this lower inference cost can offset the initial adaptation overhead. The response also does not address the previously reported fourfold gap in peak memory usage. The current evidence therefore does not establish a general deployment-efficiency advantage for DyCSTA.

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

This paper presents DyCSTA, a method of spatio-temporal forecasting based on dynamic graph models. The method consists of three main components: dynamic node evolution for capturing the history information of graph evolution for persistent and newly arriving nodes; spatio-temporal encoder combining temporal convolution attention and direction-aware spatial attention; latent context inference through learned context prototypes. The method is evaluated on datasets such as PEMS-Stream road traffic, JiNan wireless traffic, Air-Stream air quality, and Energy-Stream wind power. Also the model components are validated through ablation study.

Critical strengths
  1. The proposed method seems straightforward to implement and reproduce.

  2. The method is extensively evaluated with a number of up-to-date baseline methods.

  3. The method achieves overall good performance against recent baselines.

Critical weaknesses
  1. Technical Novelty. The issue lies in the decomposition method of forecast and context representations based on the prototypical approach. In Sec. 4.4, the authors propose the design of representation decomposition based on prototypes, which is claimed to be key contribution. However, the idea appears to be heavily inspired by CaST [1]. CaST proposes to disentangle (decompose) entity (forecast) feature and environment (context) feature. CaST introduces the concept of temporal environments as the source of unobservable confounding factors. CaST also introduces Environmental Codebook based on vector quantization, which is conceptually similar to prototypical approach (this paper even uses the same notation e1,e2,…,eke_1,e_2,\dots,e_k as CaST for learnable vectors). Even in Figure 2, the latent context method is labelled as Latent Environment Inference, using similar term as CaST, however this should be Latent Context Inference in this work to be consistent. The term environment in CaST is based on explicit modeling of causal relations of variables modeled through SCM. This work does not explicitly model such causal graphs, but the term environment from CaST is still used in Figure 2. Such an inconsistent use of the term hints the influence from CaST. Overall, the key decomposition approach proposed in this paper is similar in principle to that in CaST and thus has limited technical novelty. In addition, the design of spatio-temporal encoder with separate spatial and temporal attention is not new, e.g., the use of temporal convolutional attention with causal masks is well-known [2]. Also, it would be good if there are more details on how CT,CS,TT,TS\mathbf{C}_T,\mathbf{C}_S,\mathbf{T}_T,\mathbf{T}_S are computed.

  2. Obscure motivation. One of the central concept of this paper is the term causality. The authors heavily use causality as their main design principle (and contribution), and the term is also in the title "Causality-inspired..". In the Introduction and Related Work, the authors parallel their work with previous work on causality in graph modeling. The causality mentioned in the key related work is from Structural Causal Model (SCM) [3] [4], a graphical model for causal relation among variables. However, there is no formulation or discussion of SCM in the paper (except related work section). The only other place the term causality appears is in the discussion of the temporal attention encoder: the causality there refers to either casual temporal convolution or causal self-attention. The lack of the proper discussion or introduction of key conceptual models leads to unclear motivation, which makes the paper confusing to read.

[1] Xia, Yutong, et al. "Deciphering spatio-temporal graph forecasting: A causal lens and treatment." Advances in Neural Information Processing Systems 36 (2023): 37068-37088.

[2] Li, Shiyang, et al. "Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting." Advances in neural information processing systems 32 (2019).

[3] Judea Pearl et al. 2000. Models, reasoning and inference. Cambridge, UK: Cambridge University Press 19 (2000).

[4] Yongduo Sui, Xiang Wang, Jiancan Wu, Min Lin, Xiangnan He, and Tat-Seng Chua. 2022. Causal attention for interpretable and generalizable graph classification. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1696–1705.

Questions for authors

Please see weaknesses. Also please check the paper formats for citation: the references are not hyperlinked.

2 comments

Responses to Weakness 1: Thank you for the reviewer’s comments. CaST is an excellent work that provides valuable insights, from which we have drawn important inspiration. The idea of separating target‑prediction information from environment‑related variation in our paper is partly inspired by CaST and other related works. However, the proposed model (DyCSTA) mainly focuses on spatiotemporal forecasting with both the node set and connectivity change across stages, addressing changing sampling nodes in practical applications. The model therefore faces changes in observations, spatial dependencies, prediction targets, and the historical information available for different nodes. The research problem we address is substantively distinct from prior works such as CaST.

The concrete modeling challenges brought by the above research problem lie in: For persistent nodes, the model needs to identify which historical states are still worth retaining, so as to avoid transferring outdated information directly to the changed graph. For newly added nodes, the model cannot rely on their states from earlier stages and must be modeled using only currently available observations. When nodes leave or edges change, the model also needs to reorganize information propagation based on the current active nodes and their connections. Simply introducing representation decomposition or an environment codebook cannot automatically resolve these challenges.

Our technical design is built around these requirements: the dynamic node evolution module enables cross‑stage utilization of historical states; persistent nodes and new nodes adopt distinct historical processing pipelines before they feed into the unified current‑graph encoding; the spatiotemporal encoder reconstructs prediction dependencies on the current graph. On this basis, the context representation is used to condition the prediction representation. Empirically, DyCSTA achieves the best performance in eight of the twelve dataset–metric combinations and ranks second in the remaining four, measured by averages over all 12 forecasting horizons. Compared with the strongest baseline for each comparison, it reduces average MAE by 6.87% on PEMS‑Stream and 7.83% on JiNan, with reductions of 10.10% and 10.68%, respectively, at the 12th forecasting step.

According to this comment, we have revised the related statements in manuscript. We have replaced “Latent Environment Inference” with “latent context”, revised the Introduction, Related Work, and contribution statements, specify the computation of CT,CS,TT,TSC_T,C_S,T_T,T_S, and cited convolutional temporal attention and other existing components as basic modules.

Responses to Weakness 2: In our earlier research, we used an SCM‑based causal‑learning method for multi‑base‑station communication forecasting; due to anonymity, we cannot provide detailed bibliographic information. However, this work relies on an idealized assumption that the number of base stations within the specific region remains constant. In real‑world scenarios, base stations may be dynamically added or removed. Our present study is developed to accommodate such real dynamic changes, which constitutes the core motivation of this paper.

In terms of method design, we initially planned to follow our prior line of reasoning and build the model based on SCM causal learning. However, preliminary experiments yielded unsatisfactory results. We therefore retain the valuable idea of decomposing target‑related predictive information from disturbance caused by environmental changes, and adopt an alternative technical route to realize this idea under the setting of dynamically changing nodes. Correspondingly, we revise “Causality” in the title to “Causality‑Inspired”, to indicate that this work is inspired by causal thinking rather than adopting rigorous causal modeling methods.

We recognize that several statements in the manuscript still overemphasize causal properties and may lead to misunderstanding. We have carefully examined the entire paper and rephrased all potentially misleading descriptions in the revised version.

Responses to Question: Thank you for pointing out the reference‑link issue. We repaired in‑text links to the reference list according to the ICLR 2027 template and added DOI/arXiv links where available.

Thank you for the responses. While the anonymity issue is understandable, it is more important (and up to authors) to make the paper self-contained, especially regarding the formulation of the key model. In addition, the issue on novelty is not convincingly resolved; while the authors argue architectural novelty lies in handling dynamic graphs, such the coverage of dynamic issue is mostly concentrated in a single model component (4.2). The rest of model components (4.3 and on) do not make particular treatment of node dynamics while relying heavily on CaST. Thus, I would like to maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.