acceptodds
Under review as a conference paper at ICLR 2027

EvoST: Learning Evolving Graph Structures and Temporal Patterns for Continual Spatio-Temporal Traffic Forecasting

Abstract

Continual spatio-temporal traffic forecasting is ought to accommodate expanding traffic sensor networks and adapt to spatial and temporal distribution shifts across successive periods. Existing methods address network evolution while underexploit node-specific spatial correlations and explicit temporal patterns. In this paper, we propose EvoST, a spatio-temporal learning paradigm that jointly learns evolving graph structures and temporal patterns for continual traffic forecasting. For spatial modeling, EvoST maintains two learnable node embeddings to parameterize adaptive graph adjacencies. At each period transition, embeddings of existing nodes are inherited and augmented with randomly initialized embeddings of newly deployed traffic sensors. The optimization in the current period retains and refines previously learned relations and discovers new dependencies involving newly added nodes. For temporal modeling, a calendar embedding dictionary independent of the sensor count is inherited and updated across periods, transferring learned periodic regularities while adapting to changing temporal patterns. We instantiate this paradigm with linearized kernel graph convolution for efficient spatial aggregation and lightweight weekly time embeddings that combine discrete low-rank representations with continuous phase features. These components enable scalable modeling of spatial correlations and temporal regularities as the traffic network expands. Comprehensive experiments on three real-world continuous traffic flow datasets demonstrate the effectiveness, versatility, efficiency, and interpretability of our method. The source codes of EvoST are available at https://anonymous.4open.science/r/EvoST for review.

Primary area:
learning on time series and dynamical systems
Keywords:
spatio-temporal traffic forecasting, continual 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_34qYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EvoST for traffic forecasting when sensor networks expand across successive periods. It inherits embeddings for existing nodes, initializes embeddings for new nodes, and updates a shared weekly time representation. A linearized kernel graph convolution avoids explicitly constructing a dense adjacency matrix. On three PEMS-based datasets, EvoST reports the best average MAE, RMSE, and MAPE among the listed methods.

Critical strengths
  1. The problem is practically meaningful. When new sensors are added, a forecasting model has to deal with both a changing parameter shape and a changing dependency structure. The paper focuses on this issue directly instead of treating network expansion as a minor implementation detail.
  2. The proposed update rule is simple enough to be useful. At a period transition, old rows of the node embeddings are copied, new rows are appended, and the time dictionary can be reused because it does not depend on the number of sensors. This makes the method easier to deploy than approaches that require a separate replay or pattern-retrieval pipeline.
  3. The empirical results are encouraging. EvoST has the lowest average errors in Table 1 on all three datasets, and the EAC+ experiment is useful because it shows that the evolving graph module is not tied exclusively to the EvoST backbone. The memory measurements on PEMS12-Stream also support the motivation for avoiding an explicit dense adjacency.
Critical weaknesses
  1. The continual-learning protocol is not sufficiently clear. The paper does not specify how much data from existing sensors is available when a new period begins. It also reports aggregate current-period errors without separate results for existing and newly added sensors or a forgetting metric.
  2. The ablations do not fully separate the contributions of the different transfer mechanisms. The retraining variant evaluates the whole inheritance strategy, while removing the graph convolution does not directly compare the proposed linearized kernel with an exact or standard adaptive graph.
  3. The comparison with STBP needs clarification. STBP is reported with an unfrozen backbone on two datasets because the original version produced large errors. The authors should explain this modification and provide results under a consistent comparison setting.
  4. The efficiency analysis does not report the random-feature dimension used in the kernel approximation. A controlled study of approximation accuracy and memory usage would make the scalability claim easier to verify.
  5. The figure descriptions are incomplete. Some figures do not provide sufficiently clear captions. The authors should add complete captions to improve readability.
Questions for authors
  1. What data from existing sensors are available during training at each period transition? Please report performance separately for old and new sensors and, if possible, include a forgetting measure.
  2. Can the authors provide ablations that separately test inherited node embeddings, inherited time embeddings, and inherited shared parameters?
  3. What random-feature dimension is used, and how does changing this dimension affect accuracy and memory?
  4. Please clarify the STBP implementation on each dataset and report results under comparable training settings.
9 comments
AuthorsOpenReview ↗

Response to reviewer 34qY

Dear reviewer 34qY,

We sincerely thank for your insightful and valuable comments from which we learn a lot to improve the manuscript. We will provide point-by-point responses as following.

We also sincerely apologize for our late reply. We are hopeful our responses will address your concerns. If you have further questions, we are still available and willing to make timely and detailed replies.

Best regards,

All authors


Response to W1:

We sincerely thank you for pointing out this concern. We provide separate results for existing and newly added sensors in Response of Q1. When a new period begins, the model is trained on data of all sensors, including all existing sensors and new sensors. Thus, our setting is different from TrafficStream and STKEC while in accordance with EAC and STBP. The main challenge we want to address is how to adapt to the evolving distributions in expanding networks, including both the spatial and temporal aspects. The most important contribution of our work is proposing the expansible adaptive adjacency learning method, which deals with a long-standing critical limitation in continual learning scenario caused by the changing sensor counts across periods. Thus, our method enables adaptive adjacency learning deployable in continual scenarios, aligning the model architecture design in classic traffic forecasting and continual forecasting scenarios.

Except for EvoST, we also provide Ada-ST and EAC+ in the initial manuscript and GWNet+ in this discussion phase. These models both equip with the expansible adaptive adjacency learning and can be directly applied in continual learning scenario. Specifically, Ada-ST shows that the classic surrogate STGNN used in TrafficStream, STKEC, and naive Retrain-ST, Pretrain-ST, Online-ST-AN, Online-ST-NN can be boosted through adaptive adjacency learning. EAC+ shows that the expansible adaptive adjacency learning plays function in harmony with the learnable node prompts proposed in EAC. The two learnable techniques are not conflicting and can both improve the performance of continual learning models. GWNet+ shows that modern STGNNs which has fixed adaptive adjacency learning can be adapted to continual learning scenario, thus aligning the advanced model designs in stationary traffic forecasting and continual scenarios. The last but not least, EvoST further provides a scalable solution of expansible adaptive adjacency learning through linearized kernel graph convolution, which is more efficient when dealing with large-scale node counts or longer period span.

In addition to the spatial modeling, the temporal component in EvoST is also an useful solution to learn temporal patterns and temporal distribution shifts. It is simple yet effective without requiring pattern bank and complex pipeline.

We have added contents to articulate the continual learning protocol, data access, and the performance results of old and new old in section A.4.1 in the revised manuscript. We also added contents to stress our contributions to the continual learning scenario in section 1, 2, 4, 5.2, 6 in the revised manuscript.

Performance comparison of GWNet and GWNet+ on PEMS-Stream:

MethodMetric3612Avg.
GWNetMAE11.10 ± 0.0311.76 ± 0.0312.83 ± 0.0311.77 ± 0.03
GWNetRMSE18.47 ± 0.0419.76 ± 0.0521.56 ± 0.0619.73 ± 0.05
GWNetMAPE(%)14.95 ± 0.2915.69 ± 0.4317.28 ± 0.2415.83 ± 0.35
GWNet+MAE10.83 ± 0.0211.35 ± 0.0112.18 ± 0.0411.34 ± 0.02
GWNet+RMSE18.12 ± 0.0319.23 ± 0.0420.73 ± 0.0819.18 ± 0.05
GWNet+MAPE(%)14.34 ± 0.2914.91 ± 0.2016.17 ± 0.2215.03 ± 0.25

Response to W2:

We agree with this concern and provide further comparison results in Response of Q2 and Q3. In Response of Q2, we conduct ablations on inherited node embeddings, temporal components, and shared parameters, separately. In Response of Q3, we compare our approximation kernel with exact kernel, and also the approximation kernel itself under different number of random features, according to the prediction accuracy, training epoch time and GPU memory usage.

We have added ablation variants in section 5.3 and comparisons to exact kernel in section 5.4 in the revised manuscript.

AuthorsOpenReview ↗

Response to reviewer 34qY

Response to W3:

We agree with this concern and understand the importance of fairness in baseline comparison. In Response of Q4, we clarify our settings of STBP and report the experimental results of both frozen and unfrozen versions of STBP in our initial manuscript. Besides, we tune the hyperparameter of STBP and report the results of both frozen and unfrozen versions. Finally, we explain our motivation of modifying the original version of STBP and diagnose the reason of abnormal accuracy metrics. These analyses guarantee the fairness in baseline comparison and performance advantage of our method.

We also want to advertise our source codes, as provided in the anonymous repository. Current public codes of continual traffic forecasting methods mostly follow the pipeline of TrafficStream. It is valuable and important, but is kinda old and not easy to read. Our framework is based on the pipeline of LargeST and has many useful modifications by ourselves to guarantee the efficiency and readability, and is easy to use. We carefully sort source codes of all baselines into our framework in a unified way. We are hopeful that our codes will benefit further researches in the continual forecasting filed.

We have added clarifications and diagnoses of STBP results in section 5.2, A.4.3 and revised the results of STBP on PEMS12-Stream dataset in Table 1 and Table 10 in the revised manuscript.


Response to W4:

We sincerely thank you for pointing out this question and make clarification in Response of Q3. The random feature dimension used in the kernel approximation is 64. In Response of Q3, we compare the approximation kernel with exact kernel. We also compare approximation kernel itself under different random feature dimensions in aspects of prediction error, training epoch time, and GPU memory usage.

We have added the used random feature dimension in section 5.1 and the controlled study of our approximation kernel in section 5.4 in the revised manuscript.


Response to W5:

We sincerely thank you for pointing out this concern. In the revised manuscript, we have added complete captions to improve readability, including Figure 5, 6, 7, 8, 9, 10, 11, 12.

AuthorsOpenReview ↗

Response to reviewer 34qY

Response to Q1:

We sincerely thank you for pointing out this question and make clarification as following.

During training at each period transition, the current period data of existing sensors are available. Thus, our model is trained on data of all nodes in each period, including existing sensors and newly added sensors. This setting is different from TrafficStream and STKEC which are trained only on newly added sensors, while in accordance with EAC and STBP which are trained on all sensors.

We report performance separately for old and new sensors in the following table. It can be seen that our method has satisfied results on both old, new, and all sensors. Considering the application of forecasting in real-world continual scenario, we think the model tends to predict on current period. It does not function on historical periods. In addition, the protocols are also not compatible between current and historical periods since the sensor counts are different. Thus, we are not able provide forgetting measure.

We have added clarification of data access and provide separate results of old and new nodes in section A.4.1 in the revised manuscript.

Prediction performance of EvoST on old, new, and all nodes:

PEMS-Stream

MethodMetric3612Avg.
EvoST-Old NodesMAE11.06 ± 0.0311.72 ± 0.0512.75 ± 0.0811.71 ± 0.05
RMSE18.69 ± 0.0519.97 ± 0.0721.68 ± 0.1019.90 ± 0.07
MAPE(%)14.22 ± 0.1415.07 ± 0.1716.69 ± 0.2215.15 ± 0.14
EvoST-New NodesMAE8.73 ± 0.029.18 ± 0.029.94 ± 0.049.18 ± 0.03
RMSE14.54 ± 0.0315.28 ± 0.0416.42 ± 0.0815.28 ± 0.04
MAPE(%)29.11 ± 0.4431.75 ± 0.9336.89 ± 1.2732.26 ± 0.86
EvoST-All NodesMAE11.12 ± 0.0311.83 ± 0.0512.95 ± 0.0711.83 ± 0.05
RMSE18.60 ± 0.0619.96 ± 0.0921.78 ± 0.1019.89 ± 0.08
MAPE(%)14.67 ± 0.1315.60 ± 0.1717.39 ± 0.2115.69 ± 0.14

PEMS06-Stream

MethodMetric3612Avg.
EvoST-Old NodesMAE11.10 ± 0.0311.90 ± 0.0413.13 ± 0.0911.90 ± 0.05
RMSE18.13 ± 0.0319.60 ± 0.0621.64 ± 0.1219.59 ± 0.07
MAPE(%)16.33 ± 0.1217.41 ± 0.0919.32 ± 0.1917.48 ± 0.11
EvoST-New NodesMAE9.60 ± 0.2010.46 ± 0.3111.65 ± 0.5110.44 ± 0.32
RMSE15.17 ± 0.2916.68 ± 0.4618.47 ± 0.7216.63 ± 0.48
MAPE(%)15.76 ± 1.7817.52 ± 2.5520.67 ± 4.0517.67 ± 2.62
EvoST-All NodesMAE11.02 ± 0.0311.81 ± 0.0513.01 ± 0.1011.81 ± 0.06
RMSE17.99 ± 0.0319.46 ± 0.0721.47 ± 0.1419.44 ± 0.08
MAPE(%)16.49 ± 0.3117.67 ± 0.3619.73 ± 0.5417.74 ± 0.37

PEMS12-Stream

MethodMetric3612Avg.
EvoST-Old NodesMAE11.85 ± 0.0412.94 ± 0.0714.63 ± 0.1212.94 ± 0.07
RMSE21.18 ± 0.4223.72 ± 1.2027.61 ± 2.7723.92 ± 1.51
MAPE(%)22.15 ± 0.2624.02 ± 0.1927.16 ± 0.3124.13 ± 0.20
EvoST-New NodesMAE10.87 ± 0.0411.98 ± 0.0413.91 ± 0.0512.03 ± 0.04
RMSE20.74 ± 0.0523.23 ± 0.1127.16 ± 0.2023.42 ± 0.11
MAPE(%)17.31 ± 0.3219.00 ± 0.2321.83 ± 0.2919.09 ± 0.26
EvoST-All NodesMAE11.83 ± 0.0412.89 ± 0.0814.52 ± 0.1512.89 ± 0.09
RMSE21.44 ± 0.3823.89 ± 1.1527.70 ± 2.7924.08 ± 1.48
MAPE(%)21.79 ± 0.2323.59 ± 0.1326.63 ± 0.4623.69 ± 0.22
Reviewer_xMDrOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the continuous expansion of traffic sensor networks and the accompanying spatiotemporal distribution changes by proposing a continual traffic forecasting framework, EvoST. Spatially, it inherits existing node embeddings and appends randomly initialized embeddings for newly added nodes to learn continuously evolving node dependencies; temporally, it inherits and updates across periods a temporal embedding composed of a low-rank weekly calendar dictionary and continuous phase features to adapt to changes in periodic patterns. Meanwhile, it employs a linearized kernel graph convolution to improve the computational efficiency of spatial modeling. On three real-world traffic datasets, EvoST reduces the average MAE by 3.82%, 6.27%, and 6.39%, respectively, compared with the best existing baseline in the table.

Critical strengths
  1. The paper focuses on the practically significant problem of sensor network expansion and cross-period distribution changes, and provides a concise framework with reuse potential. Node embedding expansion and temporal representation inheritance enable the model to adapt to an evolving network, and the EAC+ experiments further support the portability of the evolutionary graph learning module. This gives its contribution application value beyond a single prediction model and provides the research community with a useful idea for adapting existing models.

  2. The paper achieves the lowest average MAE on three real traffic datasets with different network scales and time spans, demonstrating competitive predictive performance. The year-by-year results, ablation experiments, and efficiency analysis provide relatively rich empirical evidence for evaluating the method and are the main evidence supporting its effectiveness and application potential; the anonymous code entry point also provides a basis for independent verification and subsequent reuse by the community.

Critical weaknesses
  1. The paper has not yet fully established the independent contribution of the continual learning mechanism. The main technical components already have clear prior research, and the existing overall training-from-scratch ablation does not separately identify the roles of node embedding, temporal embedding, and shared parameter inheritance. In Table 4, EvoST already has a clear advantage in the first year before cross-period transfer has occurred, so the performance improvement is still difficult to attribute to the cross-period mechanism proposed in this paper. At the same time, the prediction results of the current period are not sufficient to directly verify the retention of historical knowledge. The above issues leave the paper's core innovation and its necessity lacking sufficient evidence, and this is the main reason affecting the recommendation.

  2. Key experimental comparisons and validation are still insufficient, limiting confidence in the claims of performance advantage and efficiency. The authors modified the backbone freezing strategy of STBP on PEMS06 and PEMS12, but did not provide a complete comparison between the original and modified versions or diagnose the anomaly, making it difficult to judge the fairness of an important baseline comparison. In addition, key implementation settings such as the number of random features are not fully reported, and the actual impact of the finite feature approximation lacks verification. The efficiency analysis is mainly limited to per-epoch time and GPU memory. These shortcomings affect independent reproduction of the results and reliable judgment of the trade-off between accuracy and computational cost.

Questions for authors
  1. During training in each period, can the model access the data of all old nodes in the current period? Does "retaining historical knowledge" in the paper specifically mean retaining parameter initialization, inter-node relationships, or the predictive capability of historical periods? Considering that newly added nodes change the denominator of adjacency normalization, how do the authors define and verify the claimed knowledge retention?

  2. Table 4 shows that EvoST already has a clear advantage in the first year before cross-period transfer has occurred. In addition to the existing control of overall training from scratch, can the authors provide results in which node embeddings, temporal embeddings, and shared parameters are reset separately, with a consistent training budget? These controls would help determine whether the performance improvement mainly comes from the base architecture or from the cross-period inheritance mechanism proposed in this paper, and would clarify the independent contribution of the latter.

  3. What is the specific reason for the abnormal errors of the original STBP version on PEMS06 and PEMS12? Did the authors verify the implementation and optimization settings, and can they report the results of both the frozen and unfrozen versions as well as their tuning budgets? This information will directly affect judgments about the fairness of the baseline comparison and the performance advantage.

  4. What is the number of random features used in the experiments? Is the random projection kept fixed across periods, and what stabilization measures are used in the exponential computation? For this configuration, are there comparison results with exact kernel computation or with other numbers of features to illustrate the actual impact of the finite random feature approximation on prediction error, training time, and GPU memory?

8 comments
AuthorsOpenReview ↗

Response to reviewer xMDr

Dear reviewer xMDr,

We sincerely thank for your constructive and valuable comments which significantly help us to refine our manuscript. We also sincerely apologize for our late reply.

The following are our point-by-point responses. We are hopeful to address your concerns. If you have further questions, we are available here and willing to make timely replies.

Best regards,

All authors


Response to W1:

We sincerely thank you for pointing out these concerns which help us a lot to further understand our method. We will make clarifications point-by-point as following.


Response to W1-1: The paper has not yet fully established the independent contribution of the continual learning mechanism.

Our model is trained on data of all nodes in each period, including all old and new nodes. This setting is different from TrafficStream and STKEC which are trained only on newly added nodes (and their 2-hop neighbor nodes) while in accordance with EAC and STBP, which freeze the backbone trained on initial period and employ learnable node prompts to adapt to the distributions of successive periods. However, the difference is that we do not freeze the backbone, since the empirical results show that unfrozen backbone tends to achieve better prediction stability and accuracy.

We suppose the independent contribution of the continual learning mechanism of our method exists in two aspects. Firstly, we enable adaptive adjacency learning deployable in continuous scenario where the node network is expanding. The proposed expansible adaptive adjacency learning method, though intuitive and simple, has significant implications in continual learning scenario. It tackles a long-standing critical limitation and aligns the model architecture design in classic traffic forecasting and continual forecasting. To the best of our knowledge, this work is the first to propose this approach and we believe that it will broaden the methodology of continual traffic forecasting and motivate more advances in this field.

Specifically, in addition to EvoST, we also propose two variant models Ada-ST, EAC+ in the initial manuscript, and GWNet+ in this discussion phase which are both equipped with the expansible graph learning module and can be directly applied in continual forecasting scenarios. Ada-ST shows that the classic surrogate STGNN used in TrafficStream, STKEC, and naive Retrain-ST, Pretrain-ST, Online-ST-AN, Online-ST-NN can be boosted through adaptive adjacency learning. EAC+ shows that the expansible graph learning plays function in harmony with the learnable node prompts proposed in EAC. The two learnable techniques are not conflicting and can both improve the performance of continual learning models. GWNet+ shows that modern STGNNs which has fixed adaptive adjacency learning can be adapted to continual learning scenarios, thus aligning the advanced model designs between stationary traffic forecasting and continual forecasting methods. The last but not least, EvoST further provides a scalable solution of expansible graph learning through linearized kernel graph convolution, which is more efficient when dealing with large-scale node counts or longer period span. These instances indicate that our proposed expansible graph learning fills a long-standing gap in continual traffic forecasting methods and has significant implications in graph expansion and continuous scenarios.

Secondly, the proposed weekly time embeddings can be sufficiently served as a continual-pattern learning mechanism. It is also a simple yet effective method to learn current patterns and adapt to the temporal distribution shifts across periods. The rationale is that the unit of one week can well represent the patterns of traffic flow, thus the simple calendar embedding itself can be seen as a pattern bank which learns and stores the week patterns. This approach need not extra pattern bank construction and complex retrieval pipeline. In summary, the expansible adaptive adjacency learning and week time embeddings are designed specifically to the continual scenarios in a simple yet effective way.

We have added contents to articulate the independent contribution of the continual learning mechanism of our method in section 1, 2, 5.2, 6 and the Abstract in the revised manuscript.

Performance comparison of GWNet and GWNet+ on PEMS-Stream.

MethodMetric3612Avg.
GWNetMAE11.10 ± 0.0311.76 ± 0.0312.83 ± 0.0311.77 ± 0.03
GWNetRMSE18.47 ± 0.0419.76 ± 0.0521.56 ± 0.0619.73 ± 0.05
GWNetMAPE(%)14.95 ± 0.2915.69 ± 0.4317.28 ± 0.2415.83 ± 0.35
GWNet+MAE10.83 ± 0.0211.35 ± 0.0112.18 ± 0.0411.34 ± 0.02
GWNet+RMSE18.12 ± 0.0319.23 ± 0.0420.73 ± 0.0819.18 ± 0.05
GWNet+MAPE(%)14.34 ± 0.2914.91 ± 0.2016.17 ± 0.2215.03 ± 0.25
AuthorsOpenReview ↗

Response to reviewer xMDr

Response to W1-2: The main technical components already have clear prior research.

We acknowledge that the linearized kernel graph convolution and calendar embeddings have prior research. However, in addition to their contributions specific to continuous scenarios, the two main technical components also have their own advances.

1.Firstly, the difference of our linearized kernel graph convolution and other kernel graph based methods is the kernel function. They use standard exponential kernel κ(u,v)=exp⁡(u⊤v)\kappa (\mathbf{u},\mathbf{v})=\exp \left( {{\mathbf{u}}^{\top }}\mathbf{v} \right) while ours use exponential kernel with a constant basic item κ(u,v)=α+exp⁡(u⊤v)\kappa (\mathbf{u},\mathbf{v})=\alpha +\exp \left( {{\mathbf{u}}^{\top }}\mathbf{v} \right). The motivation comes from a practical problem that how can we factorize the adaptive adjacency matrix which uses ReLU function, such as the most classic A=SoftMax⁡(ReLU⁡(E1E2⊤))\mathbf{A} = \operatorname{SoftMax}(\operatorname{ReLU}(\mathbf{E}_1\mathbf{E}_2^{\top})).

If we use the standard exponential kernel κ(u,v)=exp⁡(u⊤v)\kappa (\mathbf{u},\mathbf{v})=\exp \left( {{\mathbf{u}}^{\top }}\mathbf{v} \right), thus we obtain Aij=exp⁡(Sij)∑k=1Nexp⁡(Sik)A_{ij} = \frac{\exp(S_{ij})}{\sum_{k=1}^{N}\exp(S_{ik})}, where SijS_{ij} denotes the similarity between receiving node ii and sending node jj. The derived adaptive adjacency matrix is A=SoftMax⁡(E1E2⊤)\mathbf{A} = \operatorname{SoftMax}(\mathbf{E}_1\mathbf{E}_2^{\top}). We can see that no ReLU function is expressed.

However, if we use the exponential kernel with a constant item 1, we can obtain Aij=1+exp⁡(Sij)∑k=1N(1+exp⁡(Sik))=exp⁡(Softplus⁡(Sij))∑k=1Nexp⁡(Softplus⁡(Sik))A_{ij} = \frac{1+\exp(S_{ij})}{\sum_{k=1}^{N}(1+\exp(S_{ik}))} = \frac{\exp(\operatorname{Softplus}(S_{ij}))}{\sum_{k=1}^{N}\exp(\operatorname{Softplus}(S_{ik}))}. Thus, the derived adaptive adjacency matrix is A=SoftMax⁡(Softplus⁡(S))\mathbf{A} = \operatorname{SoftMax}(\operatorname{Softplus}(\mathbf{S})).

Since Softplus⁡(s)=ReLU⁡(s)+log⁡(1+exp⁡(−∣s∣))\operatorname{Softplus}(s) = \operatorname{ReLU}(s) + \log(1+\exp(-|s|)), and the second term approaches zero as ∣s∣|s| increases, Softplus function provides a smooth approximation to ReLU. Accordingly, we can approximately obtain A=SoftMax⁡(ReLU⁡(E1E2⊤))\mathbf{A} = \operatorname{SoftMax}(\operatorname{ReLU}(\mathbf{E}_1\mathbf{E}_2^{\top})).

Recently, there is also a concern (MAGE, NeurIPS 2025) that the ReLU function used in adaptive adjacency matrix generation causes edge noise in the graph. However, our spatial operator is not limited to the ReLU-based adjacencies.

Since our kernel function is κ(u,v)=α+exp⁡(u⊤v)\kappa (\mathbf{u},\mathbf{v})=\alpha +\exp \left( {{\mathbf{u}}^{\top }}\mathbf{v} \right), we can derive Aij=α+exp⁡(Sij)∑k=1N(α+exp⁡(Sik))A_{ij} = \frac{\alpha+\exp(S_{ij})}{\sum_{k=1}^{N}(\alpha+\exp(S_{ik}))}. When α=0\alpha=0, A=SoftMax⁡(E1E2⊤)\mathbf{A} = \operatorname{SoftMax}(\mathbf{E}_1\mathbf{E}_2^{\top}). When α=1\alpha=1, we can obtain a smooth surrogate of A=SoftMax⁡(ReLU⁡(E1E2⊤))\mathbf{A} = \operatorname{SoftMax}(\operatorname{ReLU}(\mathbf{E}_1\mathbf{E}_2^{\top})). More generally, for α>0\alpha>0, A=SoftMax⁡(Softplus⁡(S−log⁡α))\mathbf{A} = \operatorname{SoftMax}(\operatorname{Softplus}(\mathbf{S}-\log\alpha)).

Therefore, α=0\alpha=0 recovers the dot-product softmax adjacency, while α=1\alpha=1 gives a Softplus-based approximation to the ReLU-based adaptive adjacency. Varying α\alpha provides additional flexibility in modeling shared basic correlations among node pairs. Thus, the usually used standard exponential kernel is a particular case of our proposed kernel function. To the best of our knowledge, this work is the first to investigate the approximation of ReLU-based adaptive adjacency matrix and propose a general kernel function.

2.Secondly, the week time embedding, though effective, has huge parameter counts. Thus, we propose a low-rank method to substantially reduce the parameters. For instance, consider traffic data sampled every 5 minutes, yielding S=7×288=2016S=7\times 288=2016 weekly time slots. With D=128D=128 and r=8r=8, the parameter counts are Pfull=2016×128=258,048{{P}_{\text{full}}}=2016\times 128=258,048.

While Plow=2016×8+8×128+128=17,280{{P}_{\text{low}}}=2016\times 8+8\times 128+128=17,280. Thus, the resulting parameter reduction is η=1−17,280/258,048≈93.3\eta =1-{17,280}/{258,048}\approx 93.3%.

Besides, we combine the discrete time embeddings and continuous phase features to better represent week patterns. To the best of our knowledge, we are first to propose the low-rank time embeddings and the combination of two kinds of time representations.

We have added contents to stress the uniqueness of our technique components in section 1 and 4.3 in the revised manuscript.

AuthorsOpenReview ↗

Response to reviewer xMDr

Response to W1-3: the existing overall training-from-scratch ablation does not separately identify the roles of node embedding, temporal embedding, and shared parameter inheritance. In Table 4, EvoST already has a clear advantage in the first year before cross-period transfer has occurred, so the performance improvement is still difficult to attribute to the cross-period mechanism proposed in this paper.

We agree with that the existing overall training-from-scratch ablation does not separately test three kinds of parameter inheritance. In Response of Q2, we provide the ablations of node embedding, temporal component, and shared parameter inheritances.

We have added these three ablation variants in section 5.3 in the revised manuscript.


Response to W1-4: the prediction results of the current period are not sufficient to directly verify the retention of historical knowledge.

We agree with that we do not test the trained model in historical periods. In real-world continuous forecasting scenarios, we think the model tends to predict on current period. In addition, the node counts of current and historical periods are also different, which may also make it not compatible between current and historical scenarios.


Response to W2:

We sincerely thank you for pointing out these concerns. We fully agree with the importance of fairness in baseline comparison and independent reproduction of the results. We will make clarifications point-by-point as following.


Response to W2-1: The authors modified the backbone freezing strategy of STBP on PEMS06 and PEMS12, but did not provide a complete comparison between the original and modified versions or diagnose the anomaly, making it difficult to judge the fairness of an important baseline comparison.

We agree with that the initial manuscript lacks sufficient reports and diagnoses. In Response of Q3, we provide detailed analyses. Specifically, we first report the initial manuscript results of both frozen and unfrozen versions of STBP on PEMS06-Stream and PEMS12-Stream datasets, and explain why we choose to modify the original version of STBP. Then, we tune the hyperparameter of STBP and report the results of both frozen and unfrozen versions on PEMS06-Stream and PEMS12-Stream. Finally, we make diagnoses of the encountered abnormal results. We believe that these analyses will guarantee the fairness in baseline comparison and performance advantage of our method.

We have added clarifications of STBP results in section 5.2, A.4.3 in the revised manuscript.


Response to W2-2: Key implementation settings such as the number of random features are not fully reported, and the actual impact of the finite feature approximation lacks verification. The efficiency analysis is mainly limited to per-epoch time and GPU memory.

We agree with that the initial manuscript lacks detailed implementation settings, and the impact of the finite feature approximation. In Response of Q4, we report the number of random features used in the experiments, the setting of random projection and stabilization measures. Then, we provide comparison results of approximation kernel with exact kernel. We also report the actual impact of approximation kernel under different numbers of random features according to prediction error, training epoch time, and GPU memory usage. These analyses will support the effectiveness and scalability of the linearized kernel graph convolution of our method.

We have added the used number of random features in section 5.1 and added experimental results and verifications of approximation kernel in section 5.4 in the revised manuscript.

We also want to advertise our source codes, as provided in the anonymous repository. Current public codes of continual traffic forecasting methods mostly follow the pipeline of TrafficStream. It is valuable and important, but is kinda old and not easy to read. Our framework is based on the pipeline of LargeST and has many useful modifications by ourselves to guarantee the efficiency and readability, and is easy to use. We carefully sort source codes of all baselines into our framework in a unified way. We are hopeful that our codes will benefit further researches in the continual forecasting filed.

Reviewer_p47DOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

EvoST is a coherent and potentially useful approach, but its methodological novelty appears limited relative to the paper’s “new paradigm” framing. Its strongest contribution is making node-specific adaptive graph learning practical for expanding sensor networks through a simple parameter-inheritance protocol. The individual spatial and temporal components are largely established techniques or modest extensions. My main concern is that the paper never really pins down what the hard learning problem is here, beyond the fairly routine task of adapting to a growing set of sensors. The application motivation makes sense, but under the stated protocol, you already have current-period training data, you can inherit the existing parameters, and you can fine-tune the model after assigning embeddings to the new sensors. So it's not clear to me what actually distinguishes this setting from a standard warm start with a bigger parameter table. On top of that, the paper doesn't clearly separate two things: retaining knowledge from earlier periods versus simply continuing to predict the existing nodes using their current-period observations. As a result, the proposed solution feels like it falls straight out of the setup, rather than tackling a real stability–adaptation tension or uncovering a new principle for learning on evolving graphs. To make a stronger contribution, the paper would need to spell out much more sharply what makes this setting nontrivial or give a convincing argument for why simple adaptation is enough in a situation where people previously thought specialized continual-learning methods were necessary.

Critical strengths
  1. . The low-rank calendar representation produces genuine parameter savings, and the associative kernel formulation avoids materializing a dense adjacency matrix. Appendix A.2 makes these choices reasonably transparent.

  2. The Ada-ST and EAC+ variants support the practical value of introducing expandable adaptive graph learning. This is a more compelling contribution than presenting the complete architecture as an indivisible package.

Critical weaknesses
  1. The paper does not establish a sufficiently compelling continual-learning challenge. The application motivation: forecasting on an expanding sensor network is reasonable, but the paper does not clearly explain what makes its formulation substantially different from routine periodic model updating. Under the stated protocol, current-period training data are available for both existing and newly added sensors, and the model can be fine-tuned after expanding its node embeddings. Consequently, a straightforward warm-start-and-expand strategy already addresses much of the stated challenge. Crucially, continuing to predict existing sensors using their current-period observations is not the same as retaining knowledge of previous periods. The paper does not sufficiently articulate the tension between historical knowledge retention and adaptation, or explain why that tension requires a distinct continual-learning approach. This weakens the significance of the contribution: the proposed solution follows naturally from the setup, without identifying a nontrivial obstacle or revealing a compelling new insight.

  2. The methodological contribution is largely an incremental combination of established techniques. The expandable embedding protocol is useful, but the paper does not establish what substantive learning principle emerges beyond this integration. Its broad claim of a “new paradigm” therefore exceeds the demonstrated methodological advance. A simple approach could still warrant acceptance if it resolved a sharply defined difficulty or established an important, unexpected finding; the manuscript does not make a sufficiently strong case for either.

Questions for authors
  1. How does the spatial operator differ technically from BigST and other kernel-based adaptive graph methods?

  2. What makes the temporal component a continual-pattern mechanism rather than a compact calendar embedding that is fine-tuned across periods?

7 comments
AuthorsOpenReview ↗

Response to reviewer p47D

Dear reviewer p47D,

We sincerely thank for your time and efforts, which help us substantially to improve our manuscript. As following, we will provide responses to your comments point-by-point.

We also sincerely apologize for our late reply. We are hopeful our responses can address your concerns. If you have further questions, we are available here to make timely replies.

Best regards,

All authors


Response to Paper Summary and W1:

We sincerely thank for your constructive, detailed, and patient comments which help us to ponder our method and manuscript. We will make clarifications as following.

Current continual learning methods provide constructive solutions for traffic forecasting in continuously changing scenarios. However, there still exist limitations in these methods. Firstly, most current methods overlook the direct modeling of expanding structures of sensor network. These methods are node-agnostic which means that the model is not aware of the node counts. The trained model on one period can be directly used or finetuned in other periods even though the node counts are different. Although this design is specifically suitable to the continuous scenarios, it limits the modeling of spatial correlations among traffic sensors that are essential for accurate predictions. Moreover, this limitation also hinders the application of advanced STGNN on continuous scenarios which equips adaptive adjacency learning technique. Secondly, most current methods are lack of explicit temporal patterns modeling in continuous scenarios. Current pattern-based methods construct complex pattern bank from the training data while underexploit the temporal regularity inherent in data. Finally, most current methods tend to design complex workflows or rely on extra data which limits the generalization and application in real-world scenarios.

In this paper, we propose EvoST, which learns evolving graph structures and temporal patterns with the expansion of traffic sensor network across successive periods. It equips with a novel expansible graph learning module which adapts to the expanding structure of sensor network and the accompanying spatial distribution drift. This process is achieved seamlessly across periods via the generation and optimization of graph adjacencies using expansible node embeddings. As for the temporal patterns, we maintain a weekly time dictionary to store periodic regularities of traffic state data. The rationale inside is that we suppose the unit of one week can well represent the temporal patterns of traffic states. The temporal distribution drift can also be reflected by the change of weekly time patterns. Since the embedding dictionary is not related to the number of traffic sensors, it can be directly reloaded and finetuned across successive periods. This process provides a well initialization containing previous temporal patterns and adaption to current patterns via the optimization process. In concrete implementation, we use linearized kernel graph convolution to approximate the standard adaptive graph convolution, guaranteeing both accuracy and efficiency in spatial modeling. The temporal component consists of a discrete low-rank embedding and a continuous phase representation to model complementary temporal patterns in parameter efficient manner.

In addition to EvoST, we also propose three variant models Ada-ST, EAC+, and GWNet+ in this paper which are both equipped with the expansible graph learning module and can be directly applied in continual forecasting scenarios. Specifically, Ada-ST shows that the classic surrogate STGNN used in TrafficStream, STKEC, and naive Retrain-ST, Pretrain-ST, Online-ST-AN, Online-ST-NN can be boosted through adaptive adjacency learning. EAC+ shows that the expansible graph learning plays function in harmony with the learnable node prompts proposed in EAC. The two learnable techniques are not conflicting and can both improve the performance of continual learning models. GWNet+ shows that modern STGNNs which has fixed adaptive adjacency learning can be adapted to continual learning scenarios, thus aligning the advanced model designs between stationary traffic forecasting and continual forecasting methods. The last but not least, EvoST further provides a scalable solution of expansible graph learning through linearized kernel graph convolution, which is more efficient when dealing with large-scale node counts or longer period span. These instances indicate that our proposed expansible graph learning fills a long-standing gap in continual traffic forecasting methods and has significant implications in graph expansion and continuous scenarios.

We have added these contents into section 1 in the revised manuscript.

AuthorsOpenReview ↗

Response to reviewer p47D

Response to W2:

We sincerely thank you for pointing out these concerns and will make clarifications point-by-point as following.


Response to W2-1: The methodological contribution is largely an incremental combination of established techniques.

We acknowledge that the spatial operator and temporal component have prior research. However, in addition to their contributions specific to continual learning scenarios, as provided in Response of W1 and Response of Q2, the two main technical components also have their own advances. Specifically, the linearized kernel graph convolution equips with a unique kernel function which is different from the usually used kernel. It has several useful characteristics in approximation computation. We provide more details in Response of Q1. The week time embedding, though effective, has huge parameter counts. Thus, we propose a low-rank method to substantially reduce the parameters. Besides, we combine the discrete time embeddings and continuous phase features to better represent week patterns. To the best of our knowledge, this work is first to propose the low-rank time embeddings and the combination of two kinds of time representations.

We have added contents to stress the uniqueness of our technique components in section 1 and 4.3 in the revised manuscript.


Response to W2-2: The expandable embedding protocol is useful, but the paper does not establish what substantive learning principle emerges beyond this integration. Its broad claim of a “new paradigm” therefore exceeds the demonstrated methodological advance. A simple approach could still warrant acceptance if it resolved a sharply defined difficulty or established an important, unexpected finding; the manuscript does not make a sufficiently strong case for either.

We sincerely thank you for pointing out this concern and agree with that the initial manuscript does not stress the implication of our method.

We think the most important contribution of our work is the explicit modeling of structure expansion of sensor network and propose the expansible adaptive graph learning approach. It tackles a long-standing and critical limitation in continual forecasting scenarios and will substantially improve the performance of continual forecasting methods. Specifically, we propose four variant models which are both equipped with the expansible adaptive graph learning and can be directly applied in continual forecasting scenarios. The EvoST, Ada-ST, EAC+ are proposed in our initial manuscript and GWNet+ is proposed in this discussion phase. These instances indicate that our proposed expansible adaptive graph learning has significant implications in graph expansion and continual forecasting scenarios and provide the research community with a substantive and useful technique.

Performance comparison of GWNet and GWNet+ on PEMS-Stream:

MethodMetric3612Avg.
GWNetMAE11.10 ± 0.0311.76 ± 0.0312.83 ± 0.0311.77 ± 0.03
GWNetRMSE18.47 ± 0.0419.76 ± 0.0521.56 ± 0.0619.73 ± 0.05
GWNetMAPE(%)14.95 ± 0.2915.69 ± 0.4317.28 ± 0.2415.83 ± 0.35
GWNet+MAE10.83 ± 0.0211.35 ± 0.0112.18 ± 0.0411.34 ± 0.02
GWNet+RMSE18.12 ± 0.0319.23 ± 0.0420.73 ± 0.0819.18 ± 0.05
GWNet+MAPE(%)14.34 ± 0.2914.91 ± 0.2016.17 ± 0.2215.03 ± 0.25

We have deleted the contents of “paradigm” in the revised manuscript. After rethinking our methodology, we have added contents to articulate the contributions of our proposed method to continual learning scenarios in section 1, 2, 5.2, 6 and the Abstract in the revised manuscript.

We also want to advertise our source codes, as provided in the anonymous repository. Current public codes of continual traffic forecasting methods mostly follow the pipeline of TrafficStream. It is valuable and important, but is kinda old and not easy to read. Our framework is based on the pipeline of LargeST and has many useful modifications by ourselves to guarantee the efficiency and readability, and is easy to use. We carefully sort source codes of all baselines into our framework in a unified way. We are hopeful that our codes will benefit further researches in the continual forecasting filed.

AuthorsOpenReview ↗

Response to reviewer p47D

Response to Q1:

We sincerely thank you for pointing out this question, which helps us to further explain the uniqueness of our method.

The difference of our linearized kernel graph convolution and BigST is the kernel function. BigST uses standard exponential kernel κ(u,v)=exp⁡(u⊤v)\kappa (\mathbf{u},\mathbf{v})=\exp \left( {{\mathbf{u}}^{\top }}\mathbf{v} \right) while ours use exponential kernel with a constant basic item κ(u,v)=α+exp⁡(u⊤v)\kappa (\mathbf{u},\mathbf{v})=\alpha +\exp \left( {{\mathbf{u}}^{\top }}\mathbf{v} \right). The motivation comes from a practical problem that how can we factorize the adaptive adjacency matrix which uses ReLU function, such as the most classic A=SoftMax⁡(ReLU⁡(E1E2⊤))\mathbf{A} = \operatorname{SoftMax}(\operatorname{ReLU}(\mathbf{E}_1\mathbf{E}_2^{\top})).

If we use the standard exponential kernel κ(u,v)=exp⁡(u⊤v)\kappa (\mathbf{u},\mathbf{v})=\exp \left( {{\mathbf{u}}^{\top }}\mathbf{v} \right), thus we obtain Aij=exp⁡(Sij)∑k=1Nexp⁡(Sik)A_{ij} = \frac{\exp(S_{ij})}{\sum_{k=1}^{N}\exp(S_{ik})}, where SijS_{ij} denotes the similarity between receiving node ii and sending node jj. The derived adaptive adjacency matrix is A=SoftMax⁡(E1E2⊤)\mathbf{A} = \operatorname{SoftMax}(\mathbf{E}_1\mathbf{E}_2^{\top}). We can see that no ReLU function is expressed.

However, if we use the exponential kernel with a constant item 1, we can obtain Aij=1+exp⁡(Sij)∑k=1N(1+exp⁡(Sik))=exp⁡(Softplus⁡(Sij))∑k=1Nexp⁡(Softplus⁡(Sik))A_{ij} = \frac{1+\exp(S_{ij})}{\sum_{k=1}^{N}(1+\exp(S_{ik}))} = \frac{\exp(\operatorname{Softplus}(S_{ij}))}{\sum_{k=1}^{N}\exp(\operatorname{Softplus}(S_{ik}))}. Thus, the derived adaptive adjacency matrix is A=SoftMax⁡(Softplus⁡(S))\mathbf{A} = \operatorname{SoftMax}(\operatorname{Softplus}(\mathbf{S})).

Since Softplus⁡(s)=ReLU⁡(s)+log⁡(1+exp⁡(−∣s∣))\operatorname{Softplus}(s) = \operatorname{ReLU}(s) + \log(1+\exp(-|s|)), and the second term approaches zero as ∣s∣|s| increases, Softplus function provides a smooth approximation to ReLU. Accordingly, we can approximately obtain A=SoftMax⁡(ReLU⁡(E1E2⊤))\mathbf{A} = \operatorname{SoftMax}(\operatorname{ReLU}(\mathbf{E}_1\mathbf{E}_2^{\top})).

Recently, there is also a concern (MAGE, NeurIPS 2025) that the ReLU function used in adaptive adjacency matrix generation causes edge noise in the graph. However, our spatial operator is not limited to the ReLU-based adjacencies.

Since our kernel function is κ(u,v)=α+exp⁡(u⊤v)\kappa (\mathbf{u},\mathbf{v})=\alpha +\exp \left( {{\mathbf{u}}^{\top }}\mathbf{v} \right), we can derive Aij=α+exp⁡(Sij)∑k=1N(α+exp⁡(Sik))A_{ij} = \frac{\alpha+\exp(S_{ij})}{\sum_{k=1}^{N}(\alpha+\exp(S_{ik}))}. When α=0\alpha=0, A=SoftMax⁡(E1E2⊤)\mathbf{A} = \operatorname{SoftMax}(\mathbf{E}_1\mathbf{E}_2^{\top}). When α=1\alpha=1, we can obtain a smooth surrogate of A=SoftMax⁡(ReLU⁡(E1E2⊤))\mathbf{A} = \operatorname{SoftMax}(\operatorname{ReLU}(\mathbf{E}_1\mathbf{E}_2^{\top})). More generally, for α>0\alpha>0, A=SoftMax⁡(Softplus⁡(S−log⁡α))\mathbf{A} = \operatorname{SoftMax}(\operatorname{Softplus}(\mathbf{S}-\log\alpha)).

Therefore, α=0\alpha=0 recovers the dot-product softmax adjacency, while α=1\alpha=1 gives a Softplus-based approximation to the ReLU-based adaptive adjacency. Varying α\alpha provides additional flexibility in modeling shared basic correlations among node pairs. Thus, the kernel function used in BigST is a particular case of our proposed kernel function. To the best of our knowledge, this work is the first to investigate the approximation of ReLU-based adaptive adjacency matrix and propose a general kernel function.

We have added contents to stress the uniqueness of our kernel graph convolution in section 4.3 in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.