Pith. sign in

REVIEW 4 major objections 5 minor 39 references

A2TTA turns topology-induced forecasting errors into an expandable output-calibration problem and separates long-term drift from short-term context during test-time adaptation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 01:13 UTC pith:IJXSAILQ

load-bearing objection A careful empirical study whose headline TTA mechanism is not actually load-bearing; the gains come from a warmed-up calibrator. the 4 major comments →

arxiv 2607.25875 v1 pith:IJXSAILQ submitted 2026-07-28 cs.LG cs.AI

A2TTA: Anchored-and-Agile Test-Time Adaptation for Evolving Traffic Sensor Networks

Du Yin , Xiachong Lin , Yue Tan , Jinliang Deng , Estrid He , Hao Xue , Flora D. Salim This is my paper
classification cs.LG cs.AI
keywords Traffic ForecastingEvolving Sensor NetworksTest-Time AdaptationContinual LearningGraph Neural NetworksDelayed FeedbackFiLM CalibrationSensor Networks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper is trying to establish that test-time adaptation on traffic sensor networks can stay cheap and reliable even as the network itself grows. It proposes freezing the pretrained forecaster and attaching a small, expandable calibrator that rescales and shifts each node's predictions using recent observations, temporal statistics, and a learned per-node embedding. Because the shared calibrator parameters do not depend on the number of sensors, adding a new sensor only adds an embedding row while preserving previous calibration knowledge. The paper then splits adaptation into a persistent global state for long-term drift and a disposable local clone for current context, with labels used only after a full forecast horizon has elapsed. Across ten real-world traffic networks, this reduces mean absolute error by about 9–29 percent relative to the frozen backbone and is best among non-foundation methods in 119 of 120 per-horizon cells.

Core claim

The central claim is that topology expansion and multi-scale temporal drift can be absorbed without retraining the forecaster: freeze the backbone, attach an expandable node-conditioned FiLM calibrator, and maintain two complementary calibrator states. The anchored global state is updated causally from a bounded pool of released labels and regularized toward its warm-up state, capturing long-lived drift; the agile local clone is specialized using temporal-phase, pattern-similarity, and recency weighting, used for a single prediction, then discarded so transient contexts do not contaminate the global state. On ten networks, this reduces Avg-MAE over the frozen STAEFormer backbone by 9.4% to 2

What carries the argument

The engine is an expandable node-conditioned FiLM (Feature-wise Linear Modulation) calibrator. A shared two-layer MLP takes the base forecast, input history, four temporal summaries, and a learnable node embedding, and outputs per-horizon scale and shift vectors; zero initialization makes the calibrator the identity mapping at start. Because the shared MLP is independent of sensor count, adding a node only appends an embedding row, so the module carries calibration knowledge across yearly graph snapshots. Around it, a delayed-feedback pool gates labels until a horizon elapses, an anchored global state absorbs persistent drift, and a disposable local clone applies context-weighted specializat

Load-bearing premise

The load-bearing premise is that every new sensor already has some labeled traffic data from the year it appears, so its embedding row can be warmed up before deployment; if a new sensor must be predicted with no labels at all, the paper gives no evidence the expandable-embedding mechanism can initialize it.

What would settle it

Run the same evaluation on a holdout network where new sensors receive no labels in their first year: initialize their embedding rows with zeros or with the nearest existing sensor's embedding, update only from causally released labels, and compare new-sensor MAE against the frozen backbone. If the gap to frozen disappears, the warm-start label access is load-bearing.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, already-deployed fixed-graph forecasters can be upgraded in deployment by attaching the calibrator, without retraining or replacing the backbone.
  • New sensors that appear in a year's labeled training partition can be incorporated by adding embedding rows; on the new-sensor cohort, MAE drops by 8.8% on PEMS07 and 33.7% on TFNSW relative to frozen STAEFormer.
  • The global/local split prevents transient contexts from contaminating the persistent state: removing the local clone raises Avg-MAE by 1.2–1.6%, and replacing FiLM with a static affine transform raises it by 3.7–4.7%.
  • The method is deployment-cheap: it updates about 33.2K calibrator parameters versus 2.06M backbone parameters for STAEFormer, adding roughly 1.12 ms per test window.
  • Long-term use is stable across 21–25 yearly snapshots in the paper's evaluation: gains over the frozen backbone are positive in every year tested, with the largest benefits in high sensor-churn quartiles.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The setting is warm-started: new-sensor embeddings are learned on that sensor's labeled data from its first year. A fair next test would initialize new embeddings from neighbor averages or from the first few causally released labels and measure whether the gains survive; the paper's own limitation section says it should not be read as zero-shot node commissioning.
  • The same anchored/disposable split could transfer to other evolving multivariate time-series problems where nodes arrive over time, such as power meters or environmental sensors, because the calibrator consumes only the base forecast rather than architecture-specific internals.
  • The mechanism visualization suggests the local clone is a small refinement (about 20% of the global correction magnitude); a cheaper deployment could skip the clone on low-drift periods and activate it only when pool effective-sample-size or drift indicators signal a regime change.
  • Because the method is backbone-agnostic, a natural extension is to attach it to very large pretrained forecasters in settings where labels are delayed by days rather than by steps, so the persistence of the global state can be stress-tested under longer release latencies.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes A2TTA, a post-hoc calibration wrapper for frozen traffic forecasters on evolving sensor graphs. It attaches an expandable, node-conditioned FiLM module to a frozen backbone, warm-starts the calibrator on each year's labeled training partition, and then maintains a persistent global calibrator plus a disposable local clone that are updated only from causally released labels (Algorithm 1). Experiments on the EvoXXLTraffic districts and TFNSW report large Avg-MAE reductions over the frozen backbones and strong performance relative to a broad set of baselines and foundation models.

Significance. If the test-time component were load-bearing, this would be a valuable contribution: the causal delayed-feedback protocol is carefully enforced, the calibrated parameter count is small, ablation controls are present, and the evaluation breadth (ten networks, five seeds, matched checkpoints) is commendable. However, the paper's own ablations show that the online adaptation steps contribute only a small fraction of the reported gains; most of the improvement comes from the per-year supervised warm-up on the same year's training partition. The method can still be useful as a supervised calibration wrapper, and the benchmark/protocol may be a useful resource, but the 'test-time adaptation' and 'anchored-and-agile' framing overstates what the experiments establish.

major comments (4)
  1. [§6.4, Fig. 5 and §A.8, Fig. 6] The central skeptical concern lands. Fig. 5 shows that removing online labels entirely (warm-up only) increases Avg-MAE by only 0.44% (OL-AN) and 0.80% (STAE) at the median. The component ablation in §A.8 shows freezing FiLM after warm-up costs only 0.03–0.11 MAE for OL-AN and 0.09–0.13 for STAE on PEMS03–05, and is actually better on PEMS06/STAE. These effects are on the order of 0.5–1.2% relative, whereas the headline reductions over the frozen backbone are 9.4–29.4%. The experiments therefore do not support the claim that the online 'anchored global state + disposable local clone' mechanism is the source of the measured improvement. Please either provide a setting in which causal online updates matter (e.g., no-warm-up deployment, longer streams, or a change-point stress test) or explicitly reframe the contribution as supervised calibration with optional incremental updates and temper
  2. [§4.1, §7] The topology-expansion claim is only tested for warm-started sensors. §4.1 defines new sensors in ΔV_y as sensors appearing in year y's training, validation, and test partitions, and §7 explicitly states that the setting 'should not be interpreted as zero-shot node commissioning.' Yet the abstract and introduction present topology expansion as a deployment-time obstacle that A2TTA solves through expandable node embeddings. Because each new sensor's embedding row is initialized from labeled data in the same year, the paper provides no evidence for cold-start nodes whose labels are unavailable before the first forecast. The warm-up-free control reported in §A.6 should be applied to the new-sensor cohorts, or the topology-expansion claims should be qualified accordingly.
  3. [§6.8, Fig. 10 and §5.5] The evidence for the anchored-versus-agile separation is weak. Fig. 10 shows that gains over global-only TTA are much smaller and less uniform than gains over the frozen backbone, with some years negative. Since the local clone is the main algorithmic novelty over a simple global calibrator, the paper should report a systematic comparison against global-only TTA across all ten networks and all seeds, and quantify how often the local clone improves on the global state. Without that, the 'persistent global correction and agile context-specific specialization' design is not empirically established.
  4. [§A.7] The hyperparameter sensitivity aggregate in Fig. 17 is hard to interpret: each dataset-backbone pair is normalized by 'its own paper configuration,' so a point on the learning-rate axis may be a default for only some pairs. This can make the aggregate curve misleading. Please report per-dataset/per-backbone results or a clearly defined common normalization.
minor comments (5)
  1. [Tables 2, 5, 6] Table 2 is captioned 'part 1 of 3' but the corresponding parts 2 and 3 are not labeled in the appendix; similar issues affect Tables 5 and 6. Please align captions with the actual appendix structure.
  2. [Fig. 4 and Fig. 12] The main per-horizon curves do not show error bars or confidence bands despite five-seed runs. Adding bands would make the magnitude of the improvement and its variability clearer.
  3. [§6.4] The 'No labels' result is currently one point in a sensitivity figure; given its importance for interpreting the headline claim, consider reporting the warm-up-only ablation in the main text rather than only in the appendix.
  4. [§5.5, Eq. (11)] The notation 'S_day' appears after it is used in the definition of ω_tod; please define S_day before Equation (11). Also, τ and ε are set but not swept; a brief sentence on why would be helpful.
  5. [§A.6] The warm-up-free control is run on only PEMS05, PEMS08, and TFNSW, and the TFNSW/OL-AN variant shows an 11.4% degradation. This suggests strong backbone/dataset interactions; a broader warm-up-dependence table would be more informative.

Circularity Check

0 steps flagged

No significant circularity: the calibrator is evaluated causally and the warm-up dependence is disclosed, so no prediction reduces to a fitted input by construction.

full rationale

The paper's headline result is an empirical comparison against frozen backbones under a causal delayed-feedback protocol. The calibrator is warm-started on each year's chronological training partition, but the test-time stream is processed one window at a time and targets are released only after the full horizon has elapsed. Algorithm 1 and Section 5.3 make this explicit: 'records satisfying i+H <= tau_b are released into a bounded first-in, first-out (FIFO) pool' and 'Both adaptation branches read only P_b, so pending targets cannot leak into adaptation.' Predictions are emitted before the corresponding labels are available, so the reported MAE reductions are not forced by construction: the adaptation loss is computed on previously released labels, not on the current target. The warm-up dependence is also disclosed and measured rather than hidden. Section 7 states that 'A2TTA assumes that labels eventually arrive and uses a labeled training partition to warm up the calibrator for each yearly snapshot' and that the new-sensor cohort 'should not be interpreted as zero-shot node commissioning.' Appendix A.6 reports warm-up-free increases of 1.0%, 1.1%, and 2.4% for A2TTA(STAE) and 1.7%, 1.4%, and 11.4% for A2TTA(OLAN), explicitly measuring the contribution of warm-up. The self-citation of EvoXXLTraffic [33] is a benchmark/data resource constructed from Caltrans PeMS and is accompanied by the external TFNSW network; it is not a uniqueness theorem, an ansatz, or a theoretical premise that the paper's derivation depends on. The ablation showing that online adaptation contributes a small fraction of the total gain (w/o online TTA median increases of 0.44% and 0.80%; freezing FiLM after warm-up costs 0.03-0.11 and 0.09-0.13 MAE on PEMS03-05) weakens the attribution of the gain to the test-time adaptation mechanism, but this is a credit-assignment/correctness concern, not a circularity: the evaluation still measures genuinely held-out future windows. No equation or fitted parameter is renamed as a prediction, and no load-bearing argument reduces to a self-citation. Therefore the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

A2TTA introduces no physical entities. The load-bearing choices are evaluation-scope assumptions: yearly snapshots, warm-started new sensors, and the sufficiency of output calibration. Several optimization constants are undisclosed or tuned, which a reader must accept on faith.

free parameters (7)
  • Node embedding dimension = not reported
    Each sensor gets a learnable embedding row (Eq. 5), but the dimension is not stated.
  • Pool capacity M = 512
    Used in Eq. (8) and Figure 3; stated in §5.2 caption.
  • Global update interval Δ = 64 windows
    Algorithm 1 and §6.1 state global calibration updates every 64 windows.
  • Global/local adaptation steps K_g, K_l = 3 (default per sensitivity sweep)
    Appendix A.7 says the default is three adaptation steps; no main-text value given.
  • Proximal and consistency weights λ_c, λ_p = not reported
    Eq. (9) defines the global objective with these weights, but no values appear in the main text or appendix.
  • Learning rate per backbone/dataset = 1e-3 or 3e-3 depending on dataset/backbone
    Appendix A.7 lists per-dataset learning rates; these are tuned during development.
  • Context weighting constants τ, c, ε, ESS threshold = τ=1, c=5, ε=1e-6, ESS<0.2|P| triggers uniform weights
    Fixed in §5.5; sensitivity to these constants is not reported.
axioms (5)
  • domain assumption Traffic sensor graph evolves only at yearly snapshot boundaries; within-year topology is fixed.
    Problem formulation in Eq. (1) models each district-year as one snapshot; evaluation uses yearly training/deployment cycles.
  • ad hoc to paper New sensors in year y have training/validation/test labels in that same year, so their embeddings can be warm-started.
    §3.2: "Sensors in ΔV_y appear in year y's training, validation, and test partitions but in no earlier snapshot." §7 repeats that this is not zero-shot node commissioning.
  • domain assumption A frozen backbone plus a small output FiLM calibrator can absorb topology-induced and temporal distribution errors.
    Section 5.2 assumes output calibration is sufficient; the backbone is never updated and no internal layers are adapted.
  • domain assumption Labels eventually arrive after a fixed horizon H; without labels the method cannot operate in its main form.
    Stated in §7: "A2TTA assumes that labels eventually arrive."
  • ad hoc to paper The global-consistency and proximal regularizers are benign and can be set without harming performance.
    Eq. (9) introduces λ_c and λ_p; values are not reported, so the exact regularized objective is not reproducible from the text.

pith-pipeline@v1.3.0-alltime-deepseek · 62458 in / 13401 out tokens · 120565 ms · 2026-08-01T01:13:49.183710+00:00 · methodology

0 comments
read the original abstract

Traffic forecasting is important for efficient traffic management and route planning in smart cities. Existing traffic forecasting studies typically assume fixed sensor graphs, overlooking the continuous evolution of real-world traffic networks, e.g., ongoing road network construction and evolving human mobility patterns. These dynamic changes can substantially degrade conventional forecasting models, motivating test-time adaptation (TTA) to efficiently adapt pretrained models during deployment. However, applying TTA to evolving traffic sensor networks remains challenging in two aspects. First, topology expansion introduces new sensors and connections, continuously reshaping the sensor graph. Second, tem- poral shifts vary in time scale and stability, requiring differentiated adaptation to long-term and short-term shifts. In this study, we address these challenges by proposing A2TTA, an Anchored-and-Agile Test-Time Adaptation framework for evolving traffic sensor networks, which transforms topology-induced forecasting errors into an expandable output calibration problem and separates tem- poral adaptation into persistent global correction and agile context-specific specialization. By jointly addressing topology evolution and multi-scale temporal shifts, A2TTA enables efficient and robust adaptation to continuously evolving traffic environments. Extensive experiments on ten real-world traffic networks demonstrate that A2TTA consistently improves forecasting performance across different backbones, datasets, and prediction horizons. Our code is available in https://github.com/lixus7/A2TTA.

Figures

Figures reproduced from arXiv: 2607.25875 by Du Yin, Estrid He, Flora D. Salim, Hao Xue, Jinliang Deng, Xiachong Lin, Yue Tan.

Figure 1
Figure 1. Figure 1: Motivation and design overview of A2TTA. (a) The conventional fixed-graph assumption trains and deploys a forecaster on one unchanged sensor graph. (b) Deployment instead brings topology expansion together with persistent drift and transient deviations, while labels are released only after the forecasting horizon 𝐻. (c) A2TTA keeps the fore￾casting backbone frozen, appends node embeddings to an expandable … view at source ↗
Figure 2
Figure 2. Figure 2: Sensor deployment across nine districts. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of A2TTA. An expandable node-conditioned FiLM calibrator corrects heterogeneous forecast bias as the graph expands. A pending and released feedback pipeline enforces causal label access. A persistent global state and a disposable local clone handle long-term and context-specific temporal shifts, respectively. notation for generality. Each released window 𝑖 is summarized by q𝑖 = [PITH_FULL_IMAGE:f… view at source ↗
Figure 4
Figure 4. Figure 4: All-sensor per-step errors on PEMS03 and TFNSW. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Delayed-label sensitivity; bands show the interquar [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance across drift-severity quartiles. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: examines a high-variance PEMS06-2015 new sensor se￾lected without using either method’s errors. Over seeds 51 and Q1 low Q2 Q3 Q4 high Drift-severity quartile 0 5 10 MAE reduction vs frozen (%) (a) Traffic-distribution drift Q1 low Q2 Q3 Q4 high Drift-severity quartile (b) Sensor churn A²TTA(OL-AN) A²TTA(STAE) [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 7
Figure 7. Figure 7: New-sensor performance. STAEFormer by 8.8% on PEMS07 and 33.7% on TFNSW. Appen￾dix A.3.2 gives all ten datasets. High-drift periods [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 10
Figure 10. Figure 10: Long-term five-seed stability. 6.8 Long-Term Adaptation The legacy-batched five-seed analysis in [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Structural evolution of the sensor graphs. Average [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: All-sensor per-step results on ten networks. [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: New-sensor per-step results on ten networks. [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: New-sensor Avg errors on the remaining eight datasets. [PITH_FULL_IMAGE:figures/full_fig_p017_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Drift-severity diagnostics. A 2TTA(OLAN) A 2TTA(STAE) PEMS03 PEMS04 PEMS05 PEMS06 PEMS07 PEMS08 PEMS10 PEMS11 PEMS12 TFNSW 0.1 0.3 -0.3 0.2 0.4 0.3 -0.0 0.5 0.1 0.2 0.2 0.1 0.1 0.1 0.1 0.2 0.1 0.2 -1.2 -0.5 Mean MAE gain over global-only TTA (%) A 2TTA(OLAN) A 2TTA(STAE) 6.9 6.6 7.7 9.2 5.0 4.6 6.1 5.8 6.7 6.6 5.0 5.2 4.0 3.7 7.3 9.2 5.7 5.8 26.8 19.1 Mean MAE gain over backbone (%) A 2TTA(OLAN) A 2TTA(ST… view at source ↗
Figure 16
Figure 16. Figure 16: Long-term single-seed breadth check. Each dataset-backbone pair is normalized by its own paper config￾uration before averaging, so an aggregate learning-rate point that is a default for only some pairs need not equal zero. Bands show variation across datasets. The learning-rate changes in Avg-MAE for A2TTA(OLAN) and A 2TTA(STAE), respectively, are +0.51% and +0.39% at 10−4 , +0.11% and −0.12% at 3 × 10−4 … view at source ↗
Figure 18
Figure 18. Figure 18: Global and local calibration mechanisms on [PITH_FULL_IMAGE:figures/full_fig_p019_18.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 3 linked inside Pith

  1. [1]

    George EP Box, Gwilym M Jenkins, Gregory C Reinsel, and Greta M Ljung. 2015. Time series analysis: forecasting and control. John Wiley & Sons

    2015

  2. [2]

    Pinlong Cai, Yunpeng Wang, Guangquan Lu, Peng Chen, Chuan Ding, and Jian- ping Sun. 2016. A spatiotemporal correlative k-nearest neighbor model for short-term traffic multistep forecasting.Transportation Research Part C: Emerging Technologies62 (2016), 21–34

    2016

  3. [3]

    Chao Chen, Karl Petty, Alexander Skabardonis, Pravin Varaiya, and Zhanfeng Jia. 2001. Freeway performance measurement system: mining loop detector data. Transportation research record1748 (2001), 96–102

    2001

  4. [4]

    Wei Chen and Yuxuan Liang. 2025. Expand and compress: Exploring tuning prin- ciples for continual spatio-temporal graph forecasting. InInternational Conference on Learning Representations, Vol. 2025. 81631–81656

    2025

  5. [5]

    Wei Chen and Yuxuan Liang. 2025. Learning with calibration: Exploring test- time computing of spatio-temporal forecasting.Advances in Neural Information Processing Systems38 (2025), 155895–155929

    2025

  6. [6]

    Xu Chen, Junshan Wang, and Kunqing Xie. 2021. TrafficStream: A Streaming Traf- fic Flow Forecasting Framework Based on Graph Neural Networks and Continual Learning. InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, Zhi-Hua Zhou (Ed.). International Joint Conferences on Artificial Intelligence Organization,...

    doi:10.24963/ijcai.2021/498 2021

  7. [7]

    Zheng Dong, Renhe Jiang, Haotian Gao, Hangchen Liu, Jinliang Deng, Qingsong Wen, and Xuan Song. 2024. Heterogeneity-informed meta-parameter learning for spatiotemporal time series forecasting. InProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining. 631–641

    2024

  8. [8]

    Pengxin Guo, Pengrong Jin, Ziyue Li, Lei Bai, and Yu Zhang. 2024. Online Test- Time Adaptation of Spatial-Temporal Traffic Flow Forecasting.arXiv preprint arXiv:2401.04148(2024)

    arXiv 2024

  9. [9]

    Shengnan Guo, Youfang Lin, Ning Feng, Chao Song, and Huaiyu Wan. 2019. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 33. 922–929

    2019

  10. [10]

    Jiawei Jiang, Chengkai Han, Wayne Xin Zhao, and Jingyuan Wang. 2023. Pdformer: Propagation delay-aware dynamic long-range transformer for traffic flow prediction. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 4365–4373

    2023

  11. [11]

    Renhe Jiang, Du Yin, Zhaonan Wang, Yizhuo Wang, Jiewen Deng, Hangchen Liu, Zekun Cai, Jinliang Deng, Xuan Song, and Ryosuke Shibasaki. 2021. Dl-traff: Survey and benchmark of deep learning models for urban traffic prediction. In Proceedings of the 30th ACM international conference on information & knowledge management. 4515–4525

    2021

  12. [12]

    Selvaraj Vasantha Kumar. 2017. Traffic flow prediction using Kalman filtering technique.Procedia Engineering187 (2017), 582–587

    2017

  13. [13]

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2018. Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting. InInternational Conference on Learning Representations

    2018

  14. [14]

    Aoyu Liu and Yaying Zhang. 2026. A General Spatio-Temporal Backbone with Scalable Contextual Pattern Bank for Urban Continual Forecasting. InThe Four- teenth International Conference on Learning Representations

    2026

  15. [15]

    Hangchen Liu, Zheng Dong, Renhe Jiang, Jiewen Deng, Jinliang Deng, Quan- jun Chen, and Xuan Song. 2023. Spatio-temporal adaptive embedding makes vanilla transformer sota for traffic forecasting. InProceedings of the 32nd ACM international conference on information and knowledge management. 4125–4129

    2023

  16. [16]

    Tengfei Lyu, Weijia Zhang, Jinliang Deng, and Hao Liu. 2025. Autostf: Decoupled neural architecture search for cost-effective automated spatio-temporal forecast- ing. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 985–996

    2025

  17. [17]

    Haoyuan Ma, Mintao Zhou, Xiaodong Ouyang, Du Yin, Renhe Jiang, and Xuan Song. 2022. Forecasting Regional Multimodal Transportation Demand with Graph Neural Networks: An Open Dataset. In2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC). IEEE, 3263–3268. doi:10.1109/ITSC55140. 2022.9922512

    arXiv 2022

  18. [18]

    Minbo Ma, Kai Tang, Huan Li, Fei Teng, Dalin Zhang, and Tianrui Li. 2025. Beyond fixed variables: Expanding-variate time series forecasting via flat scheme and spatio-temporal focal learning. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 2054–2065

    2025

  19. [19]

    Florin Schimbinschi, Luis Moreira-Matias, Vinh Xuan Nguyen, and James Bailey

  20. [20]

    Zezhi Shao, Zhao Zhang, Fei Wang, Wei Wei, and Yongjun Xu. 2022. Spatial- temporal identity: A simple yet effective baseline for multivariate time series forecasting. InProceedings of the 31st ACM international conference on information & knowledge management. 4454–4458

    2022

  21. [21]

    Zezhi Shao, Zhao Zhang, Wei Wei, Fei Wang, Yongjun Xu, Xin Cao, and Chris- tian S Jensen. 2022. Decoupled dynamic spatial-temporal graph neural network for traffic forecasting.Proceedings of the VLDB Endowment15 (2022), 2733–2746. Conference’17, July 2017, Washington, DC, USA Yin et al

    2022

  22. [22]

    Shun-Yao Shih, Fan-Keng Sun, and Hung-yi Lee. 2019. Temporal pattern attention for multivariate time series forecasting.Machine Learning108, 8 (2019), 1421– 1441

    2019

  23. [23]

    Aaron Van Den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, Koray Kavukcuoglu, et al

  24. [24]

    Binwu Wang, Yudong Zhang, Jiahao Shi, Pengkun Wang, Xu Wang, Lei Bai, and Yang Wang. 2023. Knowledge expansion and consolidation for continual traffic prediction with expanding graphs.IEEE Transactions on Intelligent Transportation Systems24, 7 (2023), 7190–7201

    2023

  25. [25]

    Binwu Wang, Yudong Zhang, Xu Wang, Pengkun Wang, Zhengyang Zhou, Lei Bai, and Yang Wang. 2023. Pattern expansion and consolidation on evolving graphs for continual traffic prediction. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2223–2232

    2023

  26. [26]

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. 2021. Tent: Fully Test-Time Adaptation by Entropy Minimization. In International Conference on Learning Representations

    2021

  27. [27]

    Billy M Williams and Lester A Hoel. 2003. Modeling and forecasting vehicular traffic flow as a seasonal ARIMA process: Theoretical basis and empirical results. Journal of transportation engineering129, 6 (2003), 664–672

    2003

  28. [28]

    Chun-Hsin Wu, Jan-Ming Ho, and Der-Tsai Lee. 2004. Travel-time prediction with support vector regression.IEEE transactions on intelligent transportation systems5, 4 (2004), 276–281

    2004

  29. [29]

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. 2019. Graph wavenet for deep spatial-temporal graph modeling. InProceedings of the 28th International Joint Conference on Artificial Intelligence. 1907–1913

    2019

  30. [30]

    Mingxing Xu, Wenrui Dai, Chunmiao Liu, Xing Gao, Weiyao Lin, Guo-Jun Qi, and Hongkai Xiong. 2020. Spatial-temporal transformer networks for traffic flow forecasting.arXiv preprint arXiv:2001.02908(2020)

    Pith/arXiv arXiv 2020

  31. [31]

    Du Yin, Jinliang Deng, Shuang Ao, Zechen Li, Hao Xue, Arian Prabowo, Renhe Jiang, Xuan Song, and Flora Salim. 2024. Enhancing Spatio-temporal Quan- tile Forecasting with Curriculum Learning: Lessons Learned. InProceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems. 42–53

    2024

  32. [32]

    Du Yin, Renhe Jiang, Jiewen Deng, Yongkang Li, Yi Xie, Zhongyi Wang, Yifan Zhou, Xuan Song, and Jedi S Shang. 2023. MTMGNN: Multi-time multi-graph neural network for metro passenger flow prediction.GeoInformatica27, 1 (2023), 77–105

    2023

  33. [33]

    Du Yin, Hao Xue, Arian Prabowo, Shuang Ao, and Flora Salim. 2026. From XXLTraffic to EvoXXLTraffic: Scaling Traffic Forecasting to Sensor-Evolving Networks.arXiv preprint arXiv:2605.29768(2026)

    Pith/arXiv arXiv 2026

  34. [34]

    Bing Yu, Haoteng Yin, and Zhanxing Zhu. 2018. Spatio-temporal graph convolu- tional networks: a deep learning framework for traffic forecasting. InProceedings of the 27th International Joint Conference on Artificial Intelligence. 3634–3640

    2018

  35. [35]

    Haoyu Zhang, Hao Miao, Xinke Jiang, Yuchen Fang, and Yifan Zhang. 2025. Strap: Spatio-temporal pattern retrieval for out-of-distribution generalization.Advances in Neural Information Processing Systems38 (2025), 118006–118041

    2025

  36. [36]

    Ling Zhao, Yujiao Song, Chao Zhang, Yu Liu, Pu Wang, Tao Lin, Min Deng, and Haifeng Li. 2019. T-GCN: A temporal graph convolutional network for traffic prediction.IEEE transactions on intelligent transportation systems21, 9 (2019), 3848–3858

    2019

  37. [37]

    w/o local clone

    Chuanpan Zheng, Xiaoliang Fan, Cheng Wang, and Jianzhong Qi. 2020. Gman: A graph multi-attention network for traffic prediction. InProceedings of the AAAI conference on artificial intelligence, Vol. 34. 1234–1241. A Appendix A.1 Structural Evolution Details Node counts alone do not characterize this evolution. Figure 11 therefore tracks average node degre...

    2020

  38. [2016]

    Wavenet: A generative model for raw audio.arXiv preprint arXiv:1609.03499 12, 1 (2016)

    Pith/arXiv arXiv 2016

  39. [2017]

    Topology-regularized universal vector autoregression for traffic forecasting in large urban areas.Expert Systems with Applications82 (2017), 301–316

    2017