acceptodds
Under review as a conference paper at ICLR 2027

More Sensors Only One Field: Rethinking Continual Spatio-Temporal Forecasting

Abstract

Continual spatio-temporal forecasting supports traffic management and environmental monitoring under evolving dynamics and expanding sensor networks. However, conventional graph-based continual learning methods tie forecasting representations to the current sensor layout, so adding sensors can alter how previously learned spatial relationships are represented. Our key insight is that sensor expansion changes the evidence available about a process without necessarily changing the dynamics to be learned. Motivated by this insight, we propose STFO (Spatio-Temporal Field Operator), which parameterizes forecasting knowledge as a shared field-evolution operator while handling changing sensor layouts through observation and query interfaces. STFO lifts irregular sensor histories onto a fixed latent grid through normalized coordinate-based aggregation. This common computational domain enables reuse of learned spatial maps across observation sets without sensor-specific parameters. To accommodate process drift within this shared representation, a spectral descriptor summarizes variation across spatial scales and conditions Fourier propagation and attention, enabling the operator's responses to adjust to the current spatial regime. Coordinate-based decoding reads the evolved field at observed sensor locations and combines spatial corrections with local-history predictions. Experiments on PEMS-Stream, CA-Stream, and AIR-Stream demonstrate state-of-the-art average forecasting performance. Specifically, STFO -Large reduces average MAE relative to DOL by 8.4% on PEMS-Stream and 4.7% on CA-Stream. Our code is available at https://anonymous.4open.science/r/STFO2027/.

Primary area:
learning on time series and dynamical systems
Keywords:
Continual Learning, Neural Operators, Spatio-Temporal Forecasting

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_R3n9OpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper introduces STFO for continual spatial-temporal forecasting under expanding sensor networks. By mapping irregular sensor observations onto a fixed latent grid, the authors summarize the spatial-frequency characteristics of the current field to capture the current spatial regime. The method is is evaluated on three datasets across multiple horizons and hidden widths, demonstrating improvements in most settings.

Critical strengths
  1. The paper claims a meaning and interesting insights in continual spatial-temporal forecasting.
  2. The SPECTRAL REGIME ENCODER may provide a useful mechanism on changing spatial patterns.
  3. The experiment results are competitive.
  4. Theoretical consistency analysis to support their motivation.
Critical weaknesses
  1. The novelty might be limited.
  2. The ablation study shows slight performance differences after removing individual component. Since the final prediction combines field and local-history contributions, it remains unclear how much of the overall improvements is attributable to the proposed field modeling components and what truly contributes to the improvements.
Questions for authors

The improvements appear considerably better on traffic forecasting. Could the authors discuss why this may be the case and whether this indicates sensitivity to the dataset domains.

4 comments
AuthorsOpenReview ↗

Response to Reviewer R3n9

We thank the reviewer for the careful review and constructive suggestions. Below, we clarify the contribution, quantify the roles of the two prediction branches, and discuss performance across domains.

w1. Novelty and contribution

Our core contribution is a mechanism for reusing and adapting spatial forecasting knowledge as sensor networks expand. The key insight is that changes in the observation layout need not entail equally large changes in the underlying spatial dynamics. CFE and CQD provide coordinate-based observation and query interfaces to a shared latent field, allowing the same parameterization to operate across sensor layouts; SRE adjusts the operator response according to the current spectral state. Together with our consistency analysis, these designs form a continual forecasting framework that supports knowledge reuse across layouts and adaptation to changing spatial patterns.

These mechanisms are described in Sections 3.1–3.5, with the consistency analysis provided in Appendix C.

w2. Contributions of the field and local-history branches

To isolate the two branches, we independently retrained field-only and local-history-only variants over all periods of PEMS-Stream and AIR-Stream. We report three-seed mean ± standard deviation of MAE, after averaging each run over all periods and 12 prediction horizons.

DatasetLocal-history-onlyField-onlyFull
PEMS-Stream17.07 ± 0.0011.41 ± 0.0711.30 ± 0.03
AIR-Stream22.00 ± 0.0120.78 ± 0.0720.06 ± 0.03

Additionally, field-only outperforms local-history-only on both datasets and approaches Full on PEMS, indicating that field modeling provides the main predictive capability, with local history offering complementary gains. Full reduces MAE relative to local-history-only by 33.80% and 8.82% on PEMS and AIR, respectively, and further improves over field-only by 0.99% and 3.42%.

Removing a single field component leaves the remaining field components active, so the small gaps in the original ablations reflect individual components’ incremental effects rather than the contribution of the entire field branch.

We have added these independently retrained branch ablations and their analysis in Section 4.3 and Figure 3(d).

AuthorsOpenReview ↗

Response to Reviewer R3n9

q1. Performance across domains

1. Additional non-traffic datasets

To assess performance beyond traffic, we added temperature forecasting on USRCRN (2009–2014, six periods) and soil-moisture forecasting on CTP-SMTMN (2010–2016, seven periods), using observed station changes. These supplementary experiments compare STFO-L with three selected baselines: STRAP, DOL and FNO. We report three-seed mean ± sample standard deviation after equally averaging over all periods and 12 prediction horizons, evaluated on all current stations. The units are °C for USRCRN and 10−3 m3/m310^{-3}\,\mathrm{m}^3/\mathrm{m}^3 for CTP-SMTMN.

DatasetMethodMAE ↓RMSE ↓
USRCRNSTFO-L2.64 ± 0.033.45 ± 0.04
USRCRNSTRAP3.26 ± 0.164.25 ± 0.18
USRCRNDOL2.72 ± 0.043.54 ± 0.06
USRCRNFNO2.87 ± 0.153.70 ± 0.15
CTP-SMTMNSTFO-L6.00 ± 0.1310.13 ± 0.20
CTP-SMTMNSTRAP8.07 ± 1.4713.40 ± 1.66
CTP-SMTMNDOL6.60 ± 0.1011.63 ± 0.27
CTP-SMTMNFNO6.88 ± 0.4511.78 ± 0.50

STFO-L outperforms all compared baselines on both additional datasets, with substantial gains on CTP-SMTMN (9.09% lower MAE and 12.90% lower RMSE than the strongest baseline), demonstrating strong performance beyond traffic forecasting.

We have added these supplementary experiments in Appendix L (Table 15), with corresponding descriptions in Sections 4.1 and 4.2.

2. Understanding the smaller gains on AIR-Stream

To examine possible factors behind the smaller gains on AIR, we compare sensor growth and mean Pearson correlations over fixed neighboring sensor pairs from the first-period geographic 8-nearest-neighbor graph, restricted to sensors present in both the first and last periods. Correlations are computed using each period’s training observations and averaged over pairs with valid correlations in both periods. Sensor counts and correlations below are shown as first → last; growth is the net increase in sensor count.

DatasetSensors (first → last)Growth (%)Neighbor corr. (first → last)
PEMS-Stream655 → 87133.00.83 → 0.82
CA-Stream480 → 1698253.80.88 → 0.86
AIR-Stream1087 → 120210.60.68 → 0.74
USRCRN15 → 72380.00.86 → 0.76
CTP-SMTMN37 → 5754.10.77 → 0.92

AIR’s characteristics may make the challenges addressed by our method less pronounced. Its sensor count increases by only 10.6%, the smallest growth among the five datasets, suggesting less demand for adaptation to new observation layouts. Its neighboring-sensor correlations are also the lowest in both periods (0.68 → 0.74), which may limit the gains from sharing predictive information across locations. Consequently, decoupling layout adaptation from the underlying forecasting dynamics and reusing predictive structure through a shared field may yield smaller incremental gains on AIR. These observations suggest that the magnitude of the gains may vary with network expansion and spatial dependence across domains.

We have added this analysis in Appendix M (Table 16) and summarized the findings in Section 4.2.

We thank you again for the constructive feedback, which has helped us improve the clarity of the paper and strengthen the supporting evidence. All revisions are highlighted in red in the newly uploaded PDF for your review.

Thank you for the detailed response and for addressing my concerns. I appreciate the authors’ efforts and the additional experiments. I will maintain my original score. I wish the authors all the best with their work.

Thank you for your follow-up and for taking the time to consider our response and additional experiments. We appreciate your thoughtful feedback, which has helped us improve the paper.

Reviewer_o9JxOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies spatio-temporal forecasting as sensor networks expand. STFO maps a variable sensor set to a fixed latent grid, evolves the field with a spectrally conditioned Fourier/attention operator, and warm-starts the same parameters across periods. The method is evaluated on three expanding traffic and air-quality streams against thirteen baselines, together with consistency analysis and component studies.

Critical strengths

The expanding-sensor forecasting setting is practically relevant, and the paper evaluates it across both traffic and air-quality data streams.

Critical weaknesses

CFE/CQD use normalized Gaussian scatter/gather to map variable sensor sets to and from a fixed grid—directly yielding sensor-count-independent parameter shapes—while related field and variable-input interfaces already appear in [a]–[e]. The paper compares neither the structurally close GINO pipeline nor the same-protocol STBP method. The claimed novelty of the sensor-independent field representation, and its advantage over the nearest alternatives, are therefore not established.

References

[a] Spatio-Temporal Field Neural Networks for Air Quality Inference. IJCAI 2024.

[b] Spatiotemporal Implicit Neural Representation as a Generalized Traffic Data Learner. Transportation Research Part C 2024.

[c] Continuous PDE Dynamics Forecasting with Implicit Neural Representations. ICLR 2023.

[d] Operator Learning with Neural Fields: Tackling PDEs on General Geometries. NeurIPS 2023.

[e] Variable-Input Deep Operator Networks. arXiv 2022.

[f] Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting. ICLR 2018.

Questions for authors
  • How do STFO's CFE/CQD differ from GINO's irregular-to-grid interface and the coordinate-based field representations of ST-FNN and ST-INR [a], [b]?
  • CFE and CQD both use normalized Gaussian averaging. Is this essentially kernel interpolation, and how well does it preserve local information relative to other variable-input encoders?
  • How does SRE distinguish “spatial-dynamics drift” from covariate or regime shift in the observed field?
  • Table 2 evaluates only current-period sensors. What evidence demonstrates retention rather than current-period adaptation?
  • Are STFO and DOL/ST-TTC compared with the same label access, update timing, number of passes and optimization budget?
  • Figure 1(a) counts edges introduced by newly added sensors as graph reconfiguration, although these edges are necessarily new whenever the node set grows. How does this statistic demonstrate disruption of the neighbourhoods or aggregation relationships among existing sensors?
  • Even if the graph changes in Fig. 1(a) are accepted, what evidence shows that they impair forecasting or cross-period knowledge reuse in graph-based models?
  • How does the current evaluation establish STFO's advantage over GINO on model structure, STBP under the same protocol, and STGCN/DCRNN [f] as classical graph baselines?
7 comments
AuthorsOpenReview ↗

Response to Reviewer o9Jx

We thank the reviewer for the careful assessment and specific suggestions. Below, we clarify STFO’s relationship to prior work and the scope of its claims, and provide additional evidence on local spatial information, knowledge reuse, and predictive performance.

q1. Contribution and relationship to prior field methods

Thank you for highlighting these related methods. STFO focuses on multi-step forecasting as new sensors are added to the network. GINO solves PDEs on different geometries, ST-FNN infers air quality at unobserved locations, and ST-INR represents and reconstructs traffic data. Our key insight is that expanding the observation layout need not entail equally large changes in the underlying predictive dynamics. Accordingly, we design CFE/CQD to encode changing observations into a shared field and read forecasts by coordinates, supporting the reuse of forecasting knowledge across layouts. The distinctions in field representations and interfaces are summarized below.

GINO [g]: Both map between point sets and grids, but parameterize their interfaces differently. GINO transfers features through learned graph-integral kernels; CFE/CQD uses normalized Gaussian aggregation and readout with separately learnable bandwidths, applied to history encodings and predicted field features, respectively, explicitly separating shared feature mappings from spatial weighting.

ST-FNN [a]: Both use continuous fields, but differ in the represented quantity and query mechanism. ST-FNN learns the spatiotemporal gradient field of concentration and estimates target states through path integration and neighbor weighting; CFE constructs a latent feature field driven by historical observations, and CQD reads future outputs from the predicted field. The two approaches use gradient-integration inference and latent-field encoding/readout, respectively.

ST-INR [b]: Both use coordinate queries, but their field representations are conditioned differently. ST-INR fits traffic states as a coordinate function; our latent field is generated by CFE from the current observation history, and CQD queries this observation-conditioned representation. Predictions are jointly determined by the current observations and query coordinates.

The CFE and CQD formulations are provided in Sections 3.1 and 3.4, respectively.

[g] Zongyi Li et al. Geometry-Informed Neural Operator for Large-Scale 3D PDEs. NeurIPS 2023.

q2. Gaussian aggregation and preservation of local spatial information

CFE/CQD use normalized Gaussian kernel smoothing on learned features: CFE aggregates history encodings, while CQD reads the evolved latent field. Their learnable bandwidths control the spatial range of aggregation. Since smoothing may weaken local distinctions, we tested whether CFE can still support accurate prediction of differences between neighboring sensors.

For this comparison, we used independently trained models with CFE, the GINO input encoder, or a VIDON encoder [e] on AIR-Stream. All use the same pseudo-coordinates, latent grid, and downstream SRE/DFO/CQD architecture, with the local-history branch disabled. Each model learns its own parameters end to end. In addition to MAE, we measure neighborhood-difference MAE, the average error in predicted differences between neighboring sensors, ∣(y^i−y^j)−(yi−yj)∣\left|(\widehat{y}_i-\widehat{y}_j)-(y_i-y_j)\right|.

Neighbor pairs come from the original graph, with self-loops and duplicate undirected pairs removed. We report three-seed mean ± sample standard deviation. Both errors average all four periods and all 12 prediction horizons; CQD bandwidths are taken from the final period, 2019, and expressed in units of 10⁻³ of the normalized coordinate scale.

EncoderMAE ↓Neighbor-diff. MAE ↓CQD σ (×10⁻³)
CFE20.78 ± 0.0716.46 ± 0.106.08 ± 0.12
GINO input25.13 ± 0.1418.70 ± 0.0632.23 ± 1.93
VIDON input29.44 ± 0.1618.07 ± 0.0744.50 ± 6.14

Under this common downstream setup, CFE achieves lower overall MAE and reduces neighborhood-difference MAE by 12.00% and 8.95% relative to the GINO and VIDON input encoders, respectively. The latter improvement is particularly relevant to the smoothing concern: the forecasts predict differences between neighboring sensors more accurately, alongside improving pointwise prediction. CQD also learns a narrower bandwidth with CFE, indicating more localized readout from the latent field. Together, these results support CFE’s ability to retain local spatial information useful for forecasting in this setting.

We have added this encoder comparison and the corresponding spatial visualization in Appendix K (Table 14 and Figure 11).

[e] Michael Prasthofer, Tim De Ryck, and Siddhartha Mishra. Variable-Input Deep Operator Networks. arXiv:2205.11404, 2022.

AuthorsOpenReview ↗

Response to Reviewer o9Jx

q3. SRE and the interpretation of spatial-dynamics drift

We thank the reviewer for this insightful comment. We agree that SRE’s spectral descriptors can also change under covariate shifts or regime changes, and therefore cannot uniquely identify changes in the underlying dynamics. Our intended interpretation is correlational: SRE uses the observed spectral state as a predictive cue to condition the forecasting operator, without claiming causal identification of dynamics drift. Its role is to help prediction adapt to the current field. We have made this limitation explicit in the Conclusion (Section 5).

q4. Evidence for cross-period knowledge reuse

Thank you for raising this point. Our emphasis in this expanding-sensor setting is on how historical knowledge supports subsequent model updates, rather than on how much historical knowledge the updated model retains.

Table 2 evaluates existing and newly introduced sensors on AIR-Stream’s current-period test data. Its zero-shot setting directly applies the preceding-period weights without current-period training, providing evidence of immediate knowledge transfer. We further examine whether these weights improve subsequent forecasting and reduce the training needed to adapt.

We compare inherited initialization with training from scratch across all 12 expansion transitions: six on PEMS-Stream, three on CA-Stream, and three on AIR-Stream. The table reports current-period test MAE on existing sensors and the epochs required to reach a common validation target.

DatasetTransitionMAE: inherited ↓MAE: scratch ↓Epochs: inherited ↓Epochs: scratch ↓
PEMS2011→201211.055211.01843080
PEMS2012→201311.143511.18371769
PEMS2013→201411.504211.68981260
PEMS2014→201511.524211.9247378
PEMS2015→201610.893411.20421071
PEMS2016→201712.592513.1304267
CA0→116.480616.510626120
CA1→214.157814.359616157
CA2→314.585115.21304116
AIR2016→201721.025021.176608
AIR2017→201818.022818.125708
AIR2018→201917.626218.058207

Inherited initialization achieves lower existing-sensor MAE in 11 of 12 transitions and reaches the common validation target in fewer epochs in all 12 transitions. On AIR-Stream, the inherited weights meet this target before any adaptation. These results show that knowledge learned in previous periods provides a useful starting point for model updates, helping improve current-period predictions and reducing the training needed for adaptation.

We have added inherited-versus-scratch learning curves for each stream’s first and last expansion transitions in Section 4.5 (Figure 7).

q5. Label access and optimization budgets

Thank you for raising this point. We include DOL and ST-TTC to compare STFO with methods that continue adapting during testing using newly observed labels.

(1) Label access and update timing. All three methods share the same chronological data splits and use training labels for model fitting and validation labels for model selection. STFO updates on the current-period training split and keeps its parameters fixed throughout testing, using test labels only for evaluation. DOL additionally uses validation samples to initialize its replay buffer for online adaptation.

During testing, DOL and ST-TTC process windows chronologically and use a past window's labels for updates only after all 12 target observations have arrived. DOL updates its local adaptation modules before the current forecast during awake phases. ST-TTC keeps its EAC backbone fixed and updates its calibrator after the current forecast, affecting subsequent predictions.

(2) Number of passes. All three methods use a maximum of 200 epochs per period for offline training, with validation-based early stopping determining the actual number of passes. During online adaptation, DOL can reuse historical windows through replay, whereas ST-TTC performs one calibration update for each window that becomes eligible before the test stream ends. STFO performs no test-time parameter updates.

(3) Optimization budget. The offline training batch size is 64 for all three methods. The methods therefore share the same offline epoch cap and batch size. The actual number of optimizer steps depends on early stopping and any additional online updates.

STFO remains competitive while keeping its parameters fixed during each test period, even though the online baselines can continue adapting to newly observed labels.

AuthorsOpenReview ↗

Response to Reviewer o9Jx

q6. Graph growth and changes in existing sensors’ neighborhoods

Thank you for pointing this out. In Fig. 1(a), the red component counts outgoing edges from newly added sensors, while the yellow component counts neighbor replacements at existing sensors. The yellow component directly captures changes in existing sensors' neighborhoods, but edges introduced by new sensors can also affect how existing sensors aggregate information.

First, in models using graph symmetrization or bidirectional diffusion, newly introduced edges can change the information sources of existing sensors and the normalization weights of retained edges.

Second, when an existing sensor selects a new sensor as a neighbor, the new sensor's connections can further change the existing sensor's multi-hop receptive field. For example, suppose an existing sensor A replaces one of its previous neighbors, D, with a new sensor C, which is a neighbor replacement counted by the yellow component. If C selects another existing sensor B, the connection C→B is counted by the red component. The selection relations A→C and C→B allow information to flow as B→C→A, since each sensor aggregates information from its selected neighbors. If B was previously outside A's two-hop neighborhood, it now enters A's receptive field after two rounds of neighbor aggregation.

q7. Impact of network expansion on graph-based forecasting and cross-period reuse

We examine two complementary aspects: how network expansion affects graph-based predictors before any parameter updates, and how inherited parameters help subsequent adaptation.

(1) Impact of network expansion on forecasting in graph-based models. To directly address the effect of network expansion on graph-based forecasting, we evaluate STGCN, DCRNN, EVOLVE and STBP with their parameters inherited from the preceding period held fixed. For each transition, we compare inference on the old and expanded networks using identical current-period test windows and normalization, and evaluate the same existing sensors over all 12 forecasting horizons. The old-network setting retains the previous sensor set and its aggregation structure, using current observations from those sensors. The expanded setting incorporates the current sensor set and the corresponding aggregation structure, including observations from newly added sensors.

DatasetTransitionSTGCNDCRNNEVOLVESTBP
PEMS2011→2012+9.89%+4.95%+5.03%+4.15%
PEMS2012→2013+5.65%+3.07%+3.01%+2.05%
PEMS2013→2014+4.81%+3.32%+2.01%+2.14%
PEMS2014→2015+0.27%+0.25%+0.01%+0.03%
PEMS2015→2016+1.81%+2.29%+1.01%+1.15%
PEMS2016→2017+0.72%+0.70%+0.09%+0.03%
AIR2016→2017+1.31%+1.07%+1.05%+1.07%
AIR2017→2018+1.49%+1.02%+1.04%+0.72%
AIR2018→2019+0.22%+0.93%+0.51%+0.22%

Based on the supplied relative changes, all four graph-based forecasting models show higher MAE after expansion in every transition on both PEMS-Stream and AIR-Stream. The reported increases range from 0.01% to 9.89% on PEMS-Stream and from 0.22% to 1.49% on AIR-Stream, with larger changes generally occurring in the earlier PEMS transitions. This pattern is consistent with network expansion impairing the direct reuse of learned graph-based predictors: keeping inherited parameters fixed does not guarantee that their predictive performance on existing sensors is preserved when the sensor set and aggregation context change.

(2) Reuse efficiency of graph-based models after network expansion. We have added learning curves in Section 4.5 (Figure 7) comparing inherited initialization with training from scratch for STGCN, DCRNN and STRAP, with FNO and STFO included for comparison. The figure covers the first and last transitions of PEMS-Stream, CA-Stream and AIR-Stream and plots existing-sensor test MAE against adaptation epochs. Within each method, both settings update all parameters using the same current-period data and optimization settings.

The benefits of inherited initialization vary across graph-based models and transitions. For example, in the first AIR-Stream transition, inherited STGCN and DCRNN have higher MAE than their from-scratch counterparts at epoch five, while STRAP shows similar errors under the two settings. Thus, reusing graph-model parameters does not consistently translate into lower error during early adaptation.

In comparison, inherited STFO achieves lower MAE than its own from-scratch counterpart at epoch five in all six illustrated transitions, with particularly clear advantages on PEMS-Stream and CA-Stream. This shows that STFO can use historical parameters to reach low prediction errors after only a few adaptation epochs, supporting efficient model updates as the sensor network expands.

Reviewer_XhL3OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper considers spatio-temporal forecasting when sensors are added over time and the underlying spatial dynamics may also drift. The proposed model, STFO, aggregates sensor histories onto a fixed latent grid using their coordinates, propagates information with a spectrally conditioned Fourier operator and attention, and produces forecasts by querying coordinates, with an additional local-history branch. The authors prove sensor-sampling consistency and grid consistency under stated assumptions, and bound the error from coordinate reassignment separately. Experiments on PEMS-Stream, CA-Stream and AIR-Stream compare STFO with several baselines on accuracy, adaptation to new sensors, and computational cost.

Critical strengths

S1. The motivation is clear. Adding sensors can change the graph adjacency a great deal while the underlying spatial relationships stay largely similar. STFO handles this by encoding and decoding every sensor set through coordinates on a fixed latent grid, so one set of parameters is shared across periods and no sensor-specific parameters are needed. The spectrally conditioned operator, which combines two computation paths, is intended to absorb changes in the dynamics.

S2. The results are quite strong. STFO-Large is best at every horizon and metric on both traffic streams. The paper also evaluates a third stream, new-sensor adaptation, and computational cost. The theoretical results (sampling consistency, grid consistency, reassignment error) come with explicit assumptions, which makes their scope easy to assess.

Critical weaknesses

W1. Appendix D says the coordinate assignment is recomputed from the current sensor set in every period. An existing sensor can therefore end up at a different location in the latent domain even though its identity and history have not changed. Appendix C.4 bounds the prediction error in terms of coordinate displacement, but the paper never reports how large the displacements are in practice, and it does not compare against a variant that keeps existing sensors’ coordinates fixed. As a result I cannot tell how much of the previously learned spatial representation survives an expansion step, which is what the approach is meant to enable.

W2. The ablations in Figure 3 come from a single run, so small gaps between variants are hard to interpret. SRE also combines the spectral centroid, decay slope and band-energy ratios into one descriptor, and none of the three is tested individually. It is unclear whether all of them are needed. This does not call the overall architecture into question, but the specific SRE design is not yet well supported.

Questions for authors

Q1. After an expansion step, how far do existing sensors move under the Appendix D assignment? Could you report displacement statistics and compare against keeping existing coordinates fixed, at one representative stage? If the displacements are small, or fixing them makes little difference, I would be much less concerned about this.

Q2. Are the Figure 3 results consistent across random seeds? On one data stream, could you also remove the spectral centroid, decay slope and band-energy ratios from SRE one at a time? That would show whether each statistic contributes anything.

5 comments
AuthorsOpenReview ↗

Response to Reviewer XhL3

We thank the reviewer for the careful review and constructive suggestions. Below, we address your two concerns with coordinate-reassignment experiments and ablations.

w1. Coordinate displacement and the fixed-coordinate comparison

1. Displacement of existing sensors

We match existing sensors by ID across consecutive periods and measure their Euclidean displacement in the normalized domain [0,1]2[0,1]^2. The table pools sensor–transition pairs across all expansion steps.

DatasetMeanMedianP95Max.
PEMS-Stream0.037410.033850.112940.34483
CA-Stream0.162980.131900.397090.64165
AIR-Stream0.017790.014700.050440.26227

The mean displacements on PEMS and AIR are relatively small, corresponding to 3.74% and 1.78% of the domain side length, respectively. CA exhibits appreciably larger displacement. A plausible contributing factor is its stronger expansion: at the first transition, its sensor count increases by 44.0%, compared with 9.2% for PEMS and 6.2% for AIR, and its coordinate-assignment grid changes from 22×2222\times22 to 27×2727\times27.

2. Comparison with fixing existing coordinates

To assess the impact of coordinate reassignment, we added a comparison that keeps existing sensors’ coordinates fixed at each dataset’s first expansion. Reassignment uses current-period coordinates, while Fixed-old retains old sensors’ previous-period coordinates; new-sensor coordinates are identical. Both variants start from the same previous-period checkpoint: Before is evaluated without current-period training, and After follows current-period fine-tuning. MAE covers all current sensors and is averaged across 12 horizons and three seeds.

MAE at the first expansion (↓)

DatasetStageReassignmentFixed-old
PEMS-StreamBefore14.2513.20
PEMS-StreamAfter10.9811.07
CA-StreamBefore24.9020.12
CA-StreamAfter13.8914.66
AIR-StreamBefore21.2721.26
AIR-StreamAfter20.7620.77

Bold indicates the lower MAE in each row.

Before fine-tuning, fixing existing coordinates improves MAE on PEMS and CA, while the two variants are nearly identical on AIR. Thus, reassignment can introduce an immediate transfer cost, consistent with the preceding-period weights having been learned under the previous coordinate layout. After fine-tuning, reassignment achieves lower mean MAE on all three datasets: reductions of 0.80%, 5.24%, and 0.07% relative to Fixed-old on PEMS, CA, and AIR, respectively. The benefit is most pronounced on CA, whereas the AIR difference is negligible.

These results clarify the trade-off: maintaining existing coordinates favors immediate continuity, while adapting the assignment to the expanded layout can improve performance after fine-tuning, especially under the larger expansion in CA. They provide empirical context for Appendix C.4 and support adapting the shared representation across layouts.

We have added these coordinate-displacement statistics and the fixed-coordinate comparison in Appendix J (Table 13 and Figure 10).

AuthorsOpenReview ↗

Response to Reviewer XhL3

w2. Multi-seed ablations and SRE component ablations

1. Multi-seed ablations

To assess the stability of the ablation conclusions, we repeated the component ablations in Figure 3(a,b) on PEMS-Stream and AIR-Stream with three random seeds. We report mean ± standard deviation across seeds after averaging each run over all periods and 12 prediction horizons.

DatasetVariantMAE ↓MAPE (%) ↓
PEMS-StreamFull11.30 ± 0.0315.43 ± 0.09
PEMS-Streamw/o SRE11.36 ± 0.0315.55 ± 0.13
PEMS-Streamw/o DFO-S11.38 ± 0.0215.60 ± 0.03
PEMS-Streamw/o DFO-A11.54 ± 0.0415.84 ± 0.12
PEMS-Streamw/o CFE-C11.84 ± 0.0215.96 ± 0.03
AIR-StreamFull20.06 ± 0.0328.36 ± 0.39
AIR-Streamw/o SRE20.11 ± 0.0428.55 ± 0.27
AIR-Streamw/o DFO-S20.10 ± 0.0928.52 ± 0.26
AIR-Streamw/o DFO-A20.17 ± 0.0928.55 ± 0.22
AIR-Streamw/o CFE-C20.20 ± 0.0728.58 ± 0.18

Across both datasets, the full model achieves the lowest mean MAE and MAPE, and removing any component increases both mean errors.

We have updated the component ablations in Section 4.3 and Figure 3(a,b) to report three-seed means and standard deviations.

2. SRE component ablations

To assess each SRE statistic’s contribution, we removed the spectral centroid, clipped decay slope, and band-energy ratios one at a time on AIR-Stream. We report three-seed mean ± standard deviation using the same averaging protocol.

VariantMAE ↓MAPE (%) ↓
Full20.06 ± 0.0328.36 ± 0.39
w/o centroid α\alpha20.08 ± 0.0228.48 ± 0.18
w/o slope β~\tilde{\beta}20.09 ± 0.0228.46 ± 0.21
w/o band ratios η1:B\eta_{1:B}20.10 ± 0.0328.53 ± 0.17
w/o SRE20.11 ± 0.0428.55 ± 0.27

Removing each statistic raises mean MAE and MAPE, supporting complementary contributions from frequency location (centroid), spectral decay (slope), and energy distribution across scales (band ratios). Removing band ratios has the largest effect among the individual ablations.

We have added these SRE statistic ablations and their analysis in Section 4.3 and Figure 3(c).

Thank you again for these suggestions, which helped us clarify the practical consequences of coordinate reassignment and strengthen the empirical support for the SRE design. All revisions are highlighted in red in the newly uploaded PDF for your review.

Thank you again for your time and thoughtful review. As the discussion deadline approaches, we would be grateful if you could consider our response when your schedule permits. We would be happy to address any remaining questions.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Request for assistance with reviewer follow-up

Dear Area Chair,

Thank you for your time and effort in handling our submission.

We have provided detailed responses to the reviewers’ comments, but have not yet received follow-up feedback from some reviewers. We would greatly appreciate it if you could kindly encourage them to review our responses and share any remaining concerns.

We would welcome the opportunity to clarify any outstanding points during the discussion period.

Thank you very much for your help.

Best regards,
The Authors

Dear Area Chair,

Thank you very much for your time and support. We sincerely appreciate your help in facilitating the discussion.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.