acceptodds
Under review as a conference paper at ICLR 2027

Seeing the Unseen Nodes: Zero-Shot and Few-Shot Traffic Flow Prediction via Semantic Prototype-based Learning

Abstract

Traffic flow forecasting is a fundamental task in intelligent transportation systems. However, most existing methods assume a fixed set of sensors and operate under a transductive setting, limiting their ability to handle continuously expanding traffic networks. While previous studies have explored traffic expansion forecasting, they typically rely on sufficient observations from newly deployed sensors. In contrast, we study a more challenging setting in which new sensors come with only a few or even no historical observations, referred to as few-shot and zero-shot Traffic Expansion Flow Forecasting (TEFF). To tackle this problem, we propose a Semantic-driven Prototype-based Spatio-Temporal framework (SemProtoST). Specifically, we first introduce a metadata-driven node representation to enable inductive generalization to unseen nodes. We then develop dual spatio-temporal pattern prototypes, where node metadata and temporal features query shared prototypes via a cross-attention mechanism to model traffic patterns. Furthermore, we incorporate a semantic consistency constraint to enhance generalization to unseen nodes. To support few-shot adaptation, we design a knowledge distillation strategy that preserves performance on seen nodes while enabling effective adaptation to unseen nodes. Extensive experiments on four real-world datasets demonstrate that SemProtoST achieves average improvements of 3.58% and 14.22% on unseen nodes in few-shot and zero-shot settings, respectively. Our code is available at https://anonymous.4open.science/r/SemProtoST.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Traffic Flow Forecasting, Traffic Expansion, Intelligent Transportation System

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

This paper focuses on the task of traffic expansion flow forecasting, focusing on the more challenging few-shot and zero-shot Settings. For this task, the authors propose SemProtoST. This method uses sensor metadata to construct representations that can be generalized to unseen nodes, and models traffic patterns through dual spatio-temporal pattern prototypes and cross-attention query shared prototypes, and enhances the generalization ability to unseen nodes with semantic consistency constraints. In the few-shot adaptation stage, knowledge distillation is used to alleviate the forgetting of existing nodes when adapting new nodes. Compared with multi-class prediction and expansion methods on four traffic datasets including ESD, EGBA, EGLA and ECA, the average performance of the proposed method in unseen node is improved under few-shot and zero-shot Settings.

Critical strengths
  1. This paper focuses on few-shot and zero-shot Traffic Expansion Flow Forecasting tasks, points out the forgetting problem in fixed node ID representation and few-shot adaptation, and performs targeted metadata representation and knowledge distillation to solve it. It has promoted the development of this field to a certain extent.
  2. This paper conducts a large number of experiments on four real-world datasets to compare with a variety of baselines, covering fewshot and zero-shot Settings, and exploring different traffic expansion scenarios, with a relatively comprehensive experimental design.
Critical weaknesses
  1. This paper only reports the fixed loss-weight settings, but lacks corresponding experiments on the parameter sensitivity of the three key loss-weight hyperparameters, λortho, λsem and λdistill. This results in insufficient basis for weight selection and an inability to reflect the impact of different values on the model performance.
  2. Although the paper emphasizes its application scenario of rapid prediction after the new sensor is put into use, it did not provide specific comparisons with other representative baselines in terms of efficiency indicators such as actual inference delay, throughput, and memory usage to intuitively demonstrate the efficiency advantage.
Questions for authors
  1. Could the authors conduct experiments on the sensitivity of the three Loss weight hyperparameters with respect to the supplementary parameters of the dataset to illustrate the impact of different values?
  2. Could the authors provide efficiency experiments comparing the actual reasoning delay, throughput, and memory usage of SemProtoST with representative baselines on the corresponding datasets?
  3. The paper reports the average improvement results of Overall metrics and under the settings of few-shot and zero-shot learning, but does not clearly state the aggregation formula of Overall, nor does it clarify what exactly the baseline, metrics, and average method of data sets being compared for the average improvement are. Could the authors explain this and clarify it in the paper?
  4. The value of λortho is 1000, which is significantly different from the other weight scales. Could the author explain the reason for this?
Reviewer_G4t6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SemProtoST, designed for few-shot and zero-shot scenarios in traffic flow prediction. Traditional methods struggle with node generalization, spatiotemporal pattern inference, and balancing plasticity and stability. Therefore, the paper proposes three improvements: a metadata-driven node representation, a dual spatiotemporal pattern prototypes, and a knowledge distillation-based few-shot adaptation strategy. Experiments on four LargeST datasets evaluate performance in few-shot and zero-shot scenarios.

Critical strengths
  1. SemProtoST identifies unseen nodes by matching known patterns based on their attributes, which is conceptually intuitive and clearly motivated. Experiments verifies an average relative improvement of 14.22% in the zero-shot scenario.
  2. The architecture does not require an explicit graph structure, its computational complexity grows linearly with the number of nodes N, rather than the O(N2)O(N^2) commonly seen in traditional traffic prediction models. And the number of parameters is independent of N, resulting in lower computational overhead and good scalability.
Critical weaknesses
  1. There is no method in experiments that is truly designed for zero-shot scenarios. In fact, whether it is general time series or traffic flow prediction, there are relevant works in zero-shot scenarios, so the unique advantages of the model in zero-shot scenarios cannot be fully demonstrated.
  2. The current ablation study includes four variants: w/o Cat, w/o Num, w/o SCC, and w/o KD, but none of them removes the prototype itself. Moreover, orthogonal loss is the most weighted item in the total loss, and it does not appear in the ablation list. This cannot support a comprehensive analysis of each module of SemProtoST.
Questions for authors
  1. What are the advantages of SemProtoST compared to existing zero-shot methods? Can the authors discuss at least one model designed for zero-shot prediction in experiments?
  2. Can the authors incorporate more ablation variants into the comparison to more precisely illustrate the contributions of different modules? For instance, replacing the spatial/temporal prototype library with zero vectors, or direct MLP concatenation, while keeping the remaining structure and tensor dimensions unchanged.
  3. According to the code provided in the paper, the semantic consistency loss computes cosine similarity on torch.cat([meta_cat, meta_num]) (STEncoder.py:164). It seems that similarity is dominated by the lexicographic Fwy label index rather than metadata semantics. Can the authors explain it?
6 comments
AuthorsOpenReview ↗

Response to Reviewer G4t6

Dear Reviewer G4t6,

We sincerely thank the reviewer for the insightful comments and constructive suggestions. Below, we provide a point-by-point response to each concern.


Response to W1/Q1:

We have added MoGERNN [1], an inductive traffic predictor supporting newly added sensors without retraining, and UniST [2], a universal spatio-temporal model with zero-shot generalization across scenarios. We evaluate both methods on our zero-shot TEFF task.

Zero-shot TEFF results.

DatasetModelUnseen MAE / RMSE ↓\downarrowSeen MAE / RMSE ↓\downarrowOverall MAE / RMSE ↓\downarrow
ESDMoGERNN54.61 / 80.2527.02 / 41.8030.75 / 48.78
UniST53.50 / 76.7124.44 / 39.9628.36 / 46.63
SemProtoST39.37 / 58.4617.00 / 27.6120.02 / 33.47
EGBAMoGERNN43.89 / 59.8125.07 / 37.5925.91 / 38.85
UniST44.06 / 62.2322.22 / 35.6123.20 / 37.20
SemProtoST37.49 / 53.9318.07 / 29.3418.93 / 30.86
EGLAMoGERNN35.24 / 52.0527.88 / 40.8028.63 / 42.09
UniST42.83 / 63.4127.53 / 44.0629.10 / 46.42
SemProtoST28.71 / 45.9918.76 / 30.2719.78 / 32.24
ECAMoGERNNOOMOOMOOM
UniST39.81 / 60.9126.11 / 43.2927.50 / 45.39
SemProtoST33.96 / 52.7920.93 / 36.5522.25 / 38.52

OOM denotes that the models incur out-of-memory issue even when we set batch size to 1.

Compared with these two methods, SemProtoST offers the following advantages:

  1. Higher accuracy on both unseen and seen sensors. SemProtoST achieves the lowest MAE and RMSE for Unseen, Seen, and Overall in every available comparison. Against the better-performing available baseline among these two methods for each dataset and metric, unseen node MAE decreases by 14.58%–26.41% and RMSE by 9.83%–23.79%. The corresponding unweighted mean reductions across the four datasets are 18.55% and 14.65%, respectively.

  2. Metadata-conditioned transfer tailored to new sensors. MoGERNN [1] transfers information through graph aggregation, while UniST [2] learns transferable knowledge through spatio-temporal pretraining and knowledge-guided prompting. SemProtoST explicitly uses sensor metadata to query shared spatial prototypes and temporal representations to query a separate temporal prototype bank. This design connects sensor characteristics with reusable traffic patterns without requiring observations from new sensors for model training or adaptation.

  3. Deployment without an explicit sensor graph. SemProtoST constructs node representations from metadata and composes shared prototypes without constructing or updating a sensor graph, simplifying the integration of newly deployed sensors when graph information is unavailable. On ECA, with 8,600 sensors, SemProtoST completes the experiment, whereas MoGERNN encounters OOM under the tested configuration. This demonstrates feasibility at this network size.

References

[1] Zhou, Qishen, et al. "MoGERNN: An inductive traffic predictor for unobserved locations." Transportation Research Part C: Emerging Technologies 174 (2025): 105080.

[2] Yuan, Yuan, et al. "Unist: A prompt-empowered universal model for urban spatio-temporal prediction." SIGKDD, 2024.

AuthorsOpenReview ↗

Response to Reviewer G4t6

Response to W2/Q2:

We added six replacement variants to directly evaluate the spatial and temporal prototype branches on all four datasets in the few-shot setting, keeping the remaining prediction architecture and the branch input/output tensor dimensions unchanged:

  • Zero-memory-S / T / ST: replace the spatial, temporal, or both prototype banks with zero vectors.
  • MLP-S / T / ST: replace prototype-based composition in the spatial, temporal, or both branches with direct MLP mappings, whose outputs enter the same downstream concatenation and prediction modules.

ESD (Few-shot)

VariantSpatial moduleTemporal moduleUnseen MAE / RMSE ↓\downarrowSeen MAE / RMSE ↓\downarrowOverall MAE / RMSE ↓\downarrow
SemProtoSTPrototypePrototype35.88 / 54.5217.34 / 27.9519.89 / 32.81
Zero-memory-SZeroPrototype38.07 / 57.5117.34 / 28.0420.14 / 33.51
Zero-memory-TPrototypeZero37.58 / 57.2017.38 / 28.0520.11 / 33.49
Zero-memory-STZeroZero39.76 / 59.5817.34 / 27.9720.36 / 33.99
MLP-SMLPPrototype36.42 / 55.3417.35 / 27.9619.92 / 33.00
MLP-TPrototypeMLP36.56 / 54.7017.39 / 28.0619.91 / 33.01
MLP-STMLPMLP37.17 / 55.7317.35 / 28.0220.02 / 33.14

EGBA (Few-shot)

VariantSpatial moduleTemporal moduleUnseen MAE / RMSE ↓\downarrowSeen MAE / RMSE ↓\downarrowOverall MAE / RMSE ↓\downarrow
SemProtoSTPrototypePrototype31.97 / 47.4518.23 / 29.4118.84 / 30.45
Zero-memory-SZeroPrototype33.55 / 48.4918.27 / 29.4818.96 / 30.59
Zero-memory-TPrototypeZero33.23 / 48.9318.27 / 29.5118.92 / 30.63
Zero-memory-STZeroZero34.30 / 50.1018.31 / 29.4919.02 / 30.71
MLP-SMLPPrototype32.74 / 49.1318.47 / 29.8019.11 / 30.92
MLP-TPrototypeMLP34.53 / 50.3618.27 / 29.4718.89 / 30.71
MLP-STMLPMLP34.54 / 51.2118.59 / 29.9419.30 / 31.20

EGLA (Few-shot)

VariantSpatial moduleTemporal moduleUnseen MAE / RMSE ↓\downarrowSeen MAE / RMSE ↓\downarrowOverall MAE / RMSE ↓\downarrow
SemProtoSTPrototypePrototype27.57 / 44.4419.01 / 30.5219.89 / 32.23
Zero-memory-SZeroPrototype28.82 / 46.8519.05 / 30.5620.04 / 32.54
Zero-memory-TPrototypeZero28.69 / 46.6219.03 / 30.5120.02 / 32.53
Zero-memory-STZeroZero29.00 / 47.1019.05 / 30.5520.05 / 32.54
MLP-SMLPPrototype28.88 / 46.7519.06 / 30.6120.04 / 32.58
MLP-TPrototypeMLP28.58 / 46.3919.07 / 30.5520.02 / 32.47
MLP-STMLPMLP29.14 / 47.0719.10 / 30.6820.10 / 32.71

The remaining results are provided in Part 2/2 below.

AuthorsOpenReview ↗

Response to Reviewer G4t6

Response to W2/Q2 (Part 2/2)

This comment continues our response to W2/Q2 and reports the results on ECA.

ECA (Few-shot)

VariantSpatial moduleTemporal moduleUnseen MAE / RMSE ↓\downarrowSeen MAE / RMSE ↓\downarrowOverall MAE / RMSE ↓\downarrow
SemProtoSTPrototypePrototype31.49 / 49.7320.94 / 35.8622.01 / 37.51
Zero-memory-SZeroPrototype33.15 / 51.9420.94 / 35.9422.15 / 37.90
Zero-memory-TPrototypeZero32.82 / 51.3420.94 / 35.8722.14 / 37.79
Zero-memory-STZeroZero34.14 / 53.7020.94 / 37.1422.31 / 39.28
MLP-SMLPPrototype32.65 / 50.9521.07 / 36.1222.15 / 37.83
MLP-TPrototypeMLP32.06 / 50.6520.94 / 36.0022.09 / 37.68
MLP-STMLPMLP33.66 / 52.4121.17 / 36.2522.24 / 38.23

The main observations are

  • Both prototype branches contribute to unseen node forecasting. Zero-memory-S increases unseen node MAE by 4.53–6.10%, and Zero-memory-T increases it by 3.94–4.74% across the four datasets. Unseen node RMSE also worsens in every case.
  • Direct MLP mappings do not recover the performance of prototype-based composition at the same feature dimensions. Replacing both branches with MLPs increases unseen node MAE/RMSE by 6.05%/5.36% on average. Replacing both banks with zero vectors produces average increases of 7.93%/7.21%.

Meanwhile, we added an ablation experiment with orthogonal loss.

Ablation Study of Orthogonal Loss (Few-shot)

DatasetVariantsUnseen MAE / RMSE ↓\downarrowSeen MAE / RMSE ↓\downarrowOverall MAE / RMSE ↓\downarrow
ESDw/o Ortho37.97 / 56.3717.39 / 28.0420.16 / 33.30
SemProtoST35.88 / 54.5217.34 / 27.9519.89 / 32.81
EGBAw/o Ortho32.80 / 48.2918.28 / 29.5318.93 / 30.61
SemProtoST31.97 / 47.4518.23 / 29.4118.84 / 30.45
EGLAw/o Ortho27.63 / 44.4919.05 / 30.5919.93 / 32.26
SemProtoST27.57 / 44.4419.01 / 30.5219.89 / 32.23
ECAw/o Ortho31.72 / 49.9620.97 / 35.9322.06 / 37.62
SemProtoST31.49 / 49.7320.94 / 35.8622.01 / 37.51
Reviewer_yY1EOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focus on predicting traffic flow on sensors with no or less data available. The authors leverage the metadata of a sensor e.g. location, road type, nearby places to describe each sensor instead of providing fixed id which allow new sensor to be represented easily. A set of common traffic patterns is learned and sensors are mapped to each of those learned traffic pattern and sensors with different traffic pattern are mapped in different pattern. The orthogonality loss and semnatic constraint loss are designed to ensure this objective which I feel should be described more detail. The knowledge distillation loss is added to adapt to new sensors while avoiding the catastropic forgetting.

Critical strengths
  1. This paper targets the problem of predicting traffic flow at sensors with little or no previous data

  2. This paper aims to describe each sensor by its features (location, road type, nearby places) instead of a fixed ID, so new sensors work automatically.

  3. This paper learns a shared set of common traffic patterns and match each sensor to the right ones.

  4. The paper aims to ensure sensors with similar features use similar patterns, so the model handles new sensors better.

  5. Knowledge distillation is proposed to adapt to new sensors without forgetting what was learned for the old ones.

Critical weaknesses
  1. The description of figures should be more detailed that can make the main contribution of the paper easily understandable to the reader.

  2. The motivation of adding LorthoL_{ortho} to ensure prototype diversity should be discussed in more detail.

  3. The semantic consistency constraint part and the prototype diversity part seems to add two different property in the representation. This part needs to be discussed clearly

  4. Write the conclusion part of the paper which cover the main points of the paper.

  5. The authors can target larger dataset with 10000 sensors to understand how their setup scale on larger setups.

Questions for authors

Among orthogonality, semantic consistency, knowledge distillation which part do you think impacts for the performance of your model for better traffic flow prediction?

5 comments
AuthorsOpenReview ↗

Response to Reviewer yY1E

Dear Reviewer yY1E,

We sincerely thank the reviewer for the insightful comments and constructive suggestions. Below, we provide a point-by-point response to each concern.


Response to W1:

We have expanded the caption and accompanying description of Fig. 3 to explain the purpose of each component:

"The framework of SemProtoST. Metadata and Temporal embeddings query shared spatial and temporal pattern prototypes for forecasting. Orthogonality promotes prototype diversity, semantic consistency regularizes prototype usage, and knowledge distillation preserves previously learned knowledge while adapting to unseen nodes."

Response to W2:

The motivation for Lortho\mathcal{L}_{\mathrm{ortho}} is to reduce redundancy within the shared prototype banks. Without a diversity constraint, multiple prototypes may learn highly correlated representations, reducing the effective capacity of the pattern dictionary. Soft orthogonality regularization discourages correlations between normalized prototype vectors, encouraging diverse latent components that can be composed through cross-attention to represent different traffic patterns.

Our sensitivity experiments support sufficiently weighting this regularizer.

ESD (Few-shot)

Settingλortho\lambda_{\mathrm{ortho}}Unseen MAE / RMSE ↓\downarrowSeen MAE / RMSE ↓\downarrowOverall MAE / RMSE ↓\downarrow
Default100035.88 / 54.5217.34 / 27.9519.89 / 32.81
Lower ortho. weight10036.51 / 55.0317.46 / 28.1120.03 / 33.04
Higher ortho. weight1000035.99 / 54.6117.46 / 28.1319.96 / 32.96

EGBA (Few-shot)

Settingλortho\lambda_{\mathrm{ortho}}Unseen MAE / RMSE ↓\downarrowSeen MAE / RMSE ↓\downarrowOverall MAE / RMSE ↓\downarrow
Default100031.97 / 47.4518.23 / 29.4118.84 / 30.45
Lower ortho. weight10035.26 / 51.6918.29 / 29.5719.05 / 30.89
Higher ortho. weight1000032.96 / 48.1618.28 / 29.5218.94 / 30.60

EGLA (Few-shot)

Settingλortho\lambda_{\mathrm{ortho}}Unseen MAE / RMSE ↓\downarrowSeen MAE / RMSE ↓\downarrowOverall MAE / RMSE ↓\downarrow
Default100027.57 / 44.4419.01 / 30.5219.89 / 32.23
Lower ortho. weight10028.35 / 45.8919.01 / 30.5219.97 / 32.43
Higher ortho. weight1000027.09 / 44.0619.03 / 30.5419.86 / 32.19

ECA (Few-shot)

Settingλortho\lambda_{\mathrm{ortho}}Unseen MAE / RMSE ↓\downarrowSeen MAE / RMSE ↓\downarrowOverall MAE / RMSE ↓\downarrow
Default100031.49 / 49.7320.94 / 35.8622.01 / 37.51
Lower ortho. weight10031.99 / 50.3220.94 / 36.2822.06 / 37.94
Higher ortho. weight1000031.63 / 50.0920.94 / 35.8622.02 / 37.56

Reducing λortho\lambda_{\mathrm{ortho}} from 1000 to 100 increases unseen-node MAE by 1.76%, 10.29%, 2.83%, and 1.59% on ESD, EGBA, EGLA, and ECA, respectively. Increasing it to 10000 provides no consistent improvement. We will expand this motivation and include the supporting results in the revised manuscript.

Response to W3:

The two parts operate on different objects and serve complementary purposes:

  • semantic consistency constraint regularizes how nodes use the prototypes.** It encourages sensors with similar metadata to use consistent prototype combinations.
  • prototype diversity regularizes the prototype vectors, encouraging diversity within the shared dictionary.

These properties are compatible: two semantically similar sensors can assign similar weights to a shared bank of distinct prototypes. Prototype diversity provides expressive components, and semantic consistency guides their use across sensors to support inductive generalization.

AuthorsOpenReview ↗

Response to Reviewer yY1E

Response to W4:

We have revised the Conclusion to summarize the problem, key contributions, and main findings:

"We study few-shot and zero-shot Traffic Expansion Flow Forecasting, where newly deployed sensors have limited or no observations available for training or adaptation. Existing approaches often rely on fixed node identities, struggle to transfer traffic patterns to unseen sensors, and suffer from forgetting during few-shot adaptation. To address these limitations, we propose SemProtoST, which replaces fixed node ID embeddings with metadata-driven representations and transfers patterns learned from seen sensors through shared spatial and temporal prototype banks. Orthogonality regularization encourages prototype diversity, while semantic consistency enhances generalization to unseen nodes. Furthermore, knowledge-distillation-based adaptation preserves forecasting capability on seen nodes while adapting to unseen nodes. Experiments on four large-scale traffic datasets demonstrate its effectiveness and generalization capability."

Response to W5:

The ECA subset of the LargeST benchmark contains 8,600 sensors. To evaluate scalability beyond 10,000 sensors, we construct ECA-Expand, a real-sensor extension of ECA. Specifically, we retain all original ECA sensors and add 2,239 actual Caltrans PeMS sensors whose IDs are absent from ECA. For these additional sensors, we collect traffic records and metadata from the official PeMS platform, align the records with the original ECA timestamps, and aggregate them to the same 15-minute resolution. ECA-Expand therefore contains 10,839 real sensors from October to December 2019; it does not use duplicated or synthetic sensor data. We use the same temporal split, preprocessing, forecasting horizon, and seen/unseen-node protocol as ECA. We use the same model configuration without dataset-specific architectural modifications.

ECA_Expand (10839 Sensors)

DatasetModelSettingUnseen MAE / RMSE ↓\downarrowSeen MAE / RMSE ↓\downarrowOverall MAE / RMSE ↓\downarrow
ECA_ExpandSemProtoST (Our)Zero-shot32.84 / 51.9518.54 / 33.3919.91 / 35.59
STIDZero-shot55.30 / 85.3419.17 / 34.9622.63 / 42.46
SemProtoST (Our)Few-shot28.38 / 45.8218.82 / 32.9219.73 / 34.36
STIDFew-shot31.01 / 49.2019.25 / 35.6520.38 / 37.16

SemProtoST remains more accurate than STID in both settings, particularly on unseen sensors. This confirms that the method remains effective when the network expands beyond 10,000 sensors.

Furthermore, we have added a computational scalability study covering workloads from 716 to 10,839 sensors, with a batch size of 16.

Efficiency Study (Batch Size = 16)

DatasetSensorsTrain/BatchTrain/EpochMem Usage
ESD71628 ms9 s768 MiB
EGBA2,35283 ms27 s2,282 MiB
EGLA3,834137 ms45 s3,424 MiB
ECA8,600336 ms1 min 50 s9,310 MiB
ECA_Expand10,839440 ms2 min 24 s13,430 MiB

Increasing the sensors count from 8,600 to 10,839 (1.25 ××) increases per-batch training time by 1.30 ×× and GPU memory usage by 1.44 ××. Together with the measurements at smaller scales, these results indicate approximately linear growth in computational cost over the evaluated range. This trend is consistent with our architecture: nodes query shared, fixed-size prototype banks without requiring an explicit graph. Moreover, sampled SCC compares each anchor against at most 256 candidate nodes per iteration, avoiding exhaustive all-pairs comparisons. These experiments demonstrate the computational feasibility of SemProtoST beyond 10,000 nodes.

AuthorsOpenReview ↗

Response to Reviewer yY1E

Response to Question:

Their effects depend on the forecasting objective and training stage. Under the tested loss-weight perturbations, the largest deterioration in unseen node forecasting occurs when reducing the orthogonality weight. The existing ablations show complementary effects: removing SCC degrades unseen node forecasting, while removing KD degrades seen node forecasting during few-shot adaptation, consistent with increased forgetting. KD is used only during adaptation and does not contribute to zero-shot inference. Thus, the sensitivity results highlight the importance of prototype diversity, while the ablations support SCC for inductive generalization and KD for preserving existing forecasting capability. These results do not establish a universal ranking of the three components.

Ablation Study

DatasetVariantsUnseen MAE / RMSE ↓\downarrowSeen MAE / RMSE ↓\downarrowOverall MAE / RMSE ↓\downarrow
ESDw/o Ortho37.97 / 56.3717.39 / 28.0420.16 / 33.30
w/o SCC36.80 / 56.0817.82 / 28.5020.38 / 33.57
w/o KD35.93 / 54.4617.63 / 28.1820.10 / 32.97
EGBAw/o Ortho32.80 / 48.2918.28 / 29.5318.93 / 30.61
w/o SCC32.28 / 47.8018.60 / 29.8619.21 / 30.89
w/o KD31.89 / 47.3819.52 / 31.1620.07 / 32.06
EGLAw/o Ortho27.63 / 44.4919.05 / 30.5919.93 / 32.26
w/o SCC27.76 / 45.1019.15 / 30.6419.93 / 32.25
w/o KD27.40 / 44.2119.52 / 31.1620.07 / 32.06
ECAw/o Ortho31.72 / 49.9620.97 / 35.9322.06 / 37.62
w/o SCC31.90 / 50.5321.10 / 36.0722.19 / 37.97
w/o KD31.55 / 50.1122.21 / 37.9923.16 / 39.39

We thank you again for your valuable feedback, which has helped us improve the clarity and rigor of our work.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Response to Reviewer pWEy

Response to Q4:

The coefficient 1000 acts on an elementwise-averaged orthogonality penalty. Our implementation sums the normalized penalties of the spatial and temporal prototype banks:

\sum_{r\in\\{s,t\\}} \frac{1}{K_r^2} \left\lVert \widehat{\mathbf{P}}_r \widehat{\mathbf{P}}_r^{\top} - \mathbf{I} \right\rVert_F^2$$ where each prototype row is L2-normalized. For each bank, multiplying by 1000 corresponds to an unnormalized Frobenius-penalty coefficient of $1000/K_r^2$: approximately 0.977 for $K_r=32$ and 0.244 for $K_r=64$. Coefficient magnitudes alone therefore do not establish the relative influence of differently scaled losses. **The W1/Q1 sensitivity results support 1000 as a reasonable shared default: reducing it worsens unseen-node performance on all four datasets, while increasing it provides no consistent benefit.** --- We thank you again for your valuable feedback, which has helped us improve the clarity and rigor of our work.
AuthorsOpenReview ↗

Response to Reviewer pWEy

Dear Reviewer,

Thank you again for your valuable feedback. We would greatly appreciate it if you could let us know whether our responses and revisions have adequately addressed your concerns.

Thank you for your time and consideration.

Dear Reviewer pWEy,

Thank you again for your valuable feedback. We conducted sensitivity experiments for the three loss weights and clarified the role of the normalized orthogonality loss. We also clarified that “rapid prediction after deployment” refers to enabling a newly deployed sensor to participate in forecasting with little or no data collection for training or adaptation, rather than claiming an inference-speed advantage. Nevertheless, we added efficiency measurements, including training time, inference latency, throughput, and GPU memory usage. Finally, we clarified the Overall metric aggregation and recalculated the reported average improvements after incorporating the additional baselines.

We would greatly appreciate it if you could let us know whether our responses and revisions have adequately addressed your concerns. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.