Pith. sign in

REVIEW 3 major objections 5 minor 48 references

A lightweight plug-in can make inductive spatio-temporal kriging work under incomplete, block-missing sensor data by regulating unreliable inputs and calibrating value-dependent residual bias.

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

2026-07-15 14:39 UTC pith:AOSJN75G

load-bearing objection Solid plug-in engineering for inductive kriging under block missingness; useful multi-backbone gains, not a foundational result—the inductive transfer of residual prototypes is the real soft spot. the 3 major comments →

arxiv 2603.05310 v3 pith:AOSJN75G submitted 2026-03-05 cs.SD cs.AI

Latent-Mark: An Audio Watermark Robust to Neural Codec Compression

Yen-Shan Chen , Shih-Yu Lai , Ying-Jung Tsou , Yi-Cheng Lin , Bing-Yu Chen , Yun-Nung Chen , Hung-yi Lee , Shang-Tse Chen This is my paper
classification cs.SD cs.AI
keywords inductive spatio-temporal krigingincomplete observationsblock missingnessreliability-guided regulationresidual bias calibrationgraph neural networksplug-and-play frameworktraffic sensing
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.

Real sensor networks rarely give complete readings, and block-wise gaps from failures or maintenance hurt inductive kriging: inferring signals at locations that never had sensors. A common fix—impute first, then krige—creates an objective mismatch: better reconstruction on observed sensors does not reliably improve predictions at unobserved ones, and imputation can inject value-dependent bias that spatial models then spread. This paper argues that missingness and spatial inference should be handled together. It introduces UNISTOK, a plug-and-play wrapper around existing kriging backbones. Reliability-guided Signal Regulation scores each observation from temporal continuity and spatial support and damps weak or long-gap entries before they propagate. Residual Bias Calibration then learns, after the main model converges, how strongly to correct systematic over- or under-estimation in different value ranges using residual prototypes and context-aware amplitudes. Across traffic and solar datasets and many backbones, the wrapper consistently lowers error with almost no extra parameters, including under real block missingness.

Core claim

UNISTOK consistently improves multiple inductive spatio-temporal kriging backbones under incomplete observations—including block-wise missingness—by combining reliability-guided input regulation with post-hoc residual bias calibration, without changing the backbone architecture and with negligible added trainable parameters.

What carries the argument

Reliability-guided Signal Regulation (RSR) plus Residual Bias Calibration (RBC): RSR builds an entry-wise reliability field from temporal continuity and spatial support to reweight inputs before spatial propagation; RBC freezes the main predictor, estimates value-binned residual prototypes from training residuals, and learns context-conditioned correction amplitudes so systematic over-/under-estimation is reduced at unobserved targets.

Load-bearing premise

The reliability scores and value-conditioned residual patterns learned from incomplete observed sensors still apply at locations that were never used as supervised targets during training.

What would settle it

Train UNISTOK under the paper’s inductive split, then measure whether residual prototypes and reliability cues estimated only on training nodes still reduce MAE/RMSE on held-out unobserved nodes when block-missing patterns and value ranges at those nodes differ sharply from the training sensors; if gains vanish or reverse under that mismatch, the transfer assumption fails.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Existing forecasting and kriging GNNs can be upgraded for incomplete sensors without redesigning their internal layers.
  • Imputation accuracy on observed nodes is not a safe proxy objective for downstream inductive kriging under block missingness.
  • Value-dependent residual tables plus soft retrieval can serve as a general post-hoc calibrator for graph-based spatial inference.
  • Real deployments with maintenance-driven block gaps (as on the NZ-Highway data) can adopt the same plug-in without requiring complete historical coverage.
  • Parameter budgets stay nearly flat, so accuracy gains need not come from larger backbones.

Where Pith is reading between the lines

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

  • The same reliability-plus-bias-calibration pattern may transfer to other inductive graph tasks where inputs are partially observed and errors are value-skewed (e.g., air quality or energy demand at virtual nodes).
  • If residual prototypes must be re-estimated when the backbone or missingness regime shifts, online or continual prototype updates would be a natural next experiment.
  • Jointly learning the reliability field with the backbone end-to-end, instead of as a wrapper, could reveal whether the plug-in separation leaves performance on the table.
  • Stronger tests would stress non-stationary bias (regime changes, sensor drift) rather than only static block masks.

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

3 major / 5 minor

Summary. The manuscript proposes UNISTOK, a plug-and-play framework for inductive spatio-temporal kriging when observed sensors have incomplete (especially block-wise) measurements. It argues that the common impute-then-krige pipeline suffers from objective mismatch and propagates value-dependent imputation bias. UNISTOK wraps existing ISK backbones with two modules: Reliability-guided Signal Regulation (RSR), which builds an entry-wise reliability field from temporal continuity and spatial support and regulates inputs before spatial propagation, and Residual Bias Calibration (RBC), which freezes the main predictor, estimates value-binned residual prototypes from training residuals, and learns context-dependent correction amplitudes. Experiments on METR-LA, PEMS-BAY, NREL-AL, and the real block-missing NZ-Highway dataset report consistent gains across nine backbones (Table 1; Tables 7–9), with ablations, hyperparameter checks, efficiency analysis, and asymptotic analysis of soft residual retrieval (Appendix C).

Significance. If the reported gains hold under the inductive split, the work is a useful systems contribution: it targets a practically important failure mode (block missingness on observed sensors), documents objective mismatch between Stage-1 imputation rank and Stage-2 kriging rank (Appendix D.2), and delivers a non-intrusive plug-in with negligible parameter overhead (Figure 5; Appendix B.5). Strengths include multi-backbone evaluation, random/block/no-missing and real missingness settings, module ablations (D.5), and a formal treatment of reliability boundedness and soft-retrieval consistency (Theorems 4.1–4.2 / Appendix C). The result is primarily empirical engineering rather than a new kriging theory, but the problem framing and plug-in design are of clear interest to the spatio-temporal learning community.

major comments (3)
  1. [Appendix B.6; §B.4; RBC / Theorem analysis in C.2] Load-bearing inductive-transfer assumption for RBC (Appendix B.6; protocol B.4). Residual prototypes and related statistics are collected from training residuals on supervised (observed-node) entries after the main predictor converges, then frozen and applied at inference to unobserved target nodes that never receive residual supervision. The multi-backbone gains (Tables 1, 7–9) and soft-retrieval asymptotics (C.2) support the method if that value-conditioned bias surface transfers, but the manuscript does not directly compare residual-vs-prediction curves (or bin-wise conditional bias) on held-out unobserved nodes against the training prototype table. Please add a diagnostic (e.g., scatter or bin-wise GME/CB on validation/test targets vs. training prototypes) and discuss failure cases when the bias surface differs across the inductive split.
  2. [§1; RSR / Theorem 4.1; D.1; Table 1] RSR reliability is defined from temporal continuity and spatial support on the observed graph (motivation in §1; RSR description and Theorem 4.1). Under simultaneous block missingness on neighboring observed sensors, both cues can degrade together, so the reliability field may systematically over- or under-suppress anchors. Table 1 and D.1 show gains under block missingness, but there is no controlled stress test that varies co-missing of spatial neighbors or long-gap length while holding overall missing rate fixed. A short experiment or analysis quantifying RSR behavior when spatial support is itself incomplete would strengthen the claim that regulation, not just capacity, drives the improvement.
  3. [§5 RQ2; Appendix D.2; Table 4 / Fig. 8] End-to-end vs. two-stage claim (RQ2). Appendix D.2 convincingly shows weak/negative rank correlation between Stage-1 imputation MAE and Stage-2 kriging MAE, which motivates joint modeling. The main tables, however, primarily report +Ours on backbones rather than a head-to-head of UNISTOK against the best impute-then-krige pipeline under matched compute and the same backbone. Please add (or elevate) a compact comparison table: best Stage-1 imputer + backbone vs. UNISTOK-wrapped backbone under identical block-missing protocols, so the end-to-end advantage is quantified rather than only the objective-mismatch diagnosis.
minor comments (5)
  1. [Title / Abstract vs. full text] The submission package abstract/title refer to an unrelated audio-watermarking paper (Latent-Mark), while the body, tables, and arXiv line are UNISTOK (arXiv:2603.05301). Ensure the camera-ready abstract, title, primary category, and body are consistent before any further review cycle.
  2. [B.1; Tables 1, 7–9; Figures 5–8] Several figures and tables in the extracted text show placeholder or garbled glyphs (e.g., dataset sizes as boxes, missing rates, and some metric cells). Replace with clean numeric values and readable axis labels in the PDF.
  3. [Appendix A; B.4–B.6] Notation for reliability field, residual bins K, soft temperature τ, and peak-weighting factor is introduced across §4 and Appendix B; a single notation table (Appendix A is listed but thin in the extract) would help readers implement the plug-in.
  4. [§5.1; Tables 1, 7–8] MAPE is omitted on some datasets/settings without a one-line justification (e.g., near-zero targets on NREL). State when MAPE is undefined or unstable.
  5. [§5.2; §5.6; header] Minor prose issues: spacing in “UNISTOKconsistently”, “+Ourspoints”, and repeated “Preprint.” headers. Standard copy-edit pass recommended.

Circularity Check

0 steps flagged

No significant circularity: UNISTOK is an empirical plug-in (RSR + RBC) evaluated on external kriging benchmarks; residual prototypes are post-hoc training statistics, not self-defining predictions.

full rationale

The paper’s load-bearing claims are (i) that Reliability-guided Signal Regulation plus Residual Bias Calibration improve inductive spatio-temporal kriging under incomplete observations when wrapped around existing backbones, and (ii) that soft residual retrieval is asymptotically consistent for value-conditioned bias. Neither claim reduces by construction to its inputs. RSR reliability is defined from temporal continuity and spatial support and used to reweight inputs; it is not defined as the kriging error it is meant to reduce. RBC residual prototypes are empirical bin-wise averages of training residuals collected after the main predictor converges (Appendix B.6, Eqs. 12–19), then frozen; a separate amplitude network is trained and model selection uses validation MAE. Evaluation is against ground-truth signals on held-out unobserved nodes and against external baselines (Tables 1, 7–9), so reported gains are not forced by the residual fit. Theorems 4.1–4.2 (Appendix C) prove boundedness/monotonicity of the reliability field and consistency of soft retrieval under stated regularity conditions; they characterize the estimators rather than smuggling the target metric into the definition. No uniqueness theorem, load-bearing self-citation chain, or renaming of a known result as a first-principles derivation appears. Mild use of training residual statistics for post-hoc calibration is standard and does not constitute circularity under the stated criteria.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

UNISTOK is an empirical ML method. Load-bearing content is modeling choices and fitted hyperparameters rather than physical postulates. Reliability is defined from temporal continuity and spatial support; residual bias is assumed to be largely value-conditioned and bin-capturable; inductive transfer of those statistics to unobserved nodes is assumed. Free parameters are the usual training and module knobs. Invented entities are methodological constructs (reliability field, residual prototype table), not new physical objects.

free parameters (6)
  • RSR reliability regulation strength (α)
    Controls how strongly reliability reweights inputs; set by default (text cites ~0.5-class setting) and affects suppression of long-gap entries.
  • Number of residual bins K
    Discretizes prediction range for residual prototypes; default around 20; bin width drives approximation error in the theory.
  • Soft-retrieval temperature τ
    Controls softness of bin assignment in RBC; default small positive value; limits concentration of prototype retrieval.
  • Peak-weighting factor for epoch residual aggregation
    Weights epoch-wise residual tables around the validation-best epoch when building final prototypes.
  • Attention residual scale / fusion MLP sizes
    Bounds residual compensation and sets capacity of lightweight affine/MLP plug-in heads; chosen by configuration rather than derived.
  • Training schedule (lr, decay, patience, subgraphs per epoch)
    Standard optimization knobs that influence which backbone checkpoint and residual table are obtained.
axioms (5)
  • domain assumption Entry-wise reliability can be estimated from temporal continuity and spatial neighbor support sufficiently well to improve spatial propagation under block missingness.
    Core premise of RSR in the introduction and method; not proved, motivated by block-gap analysis (Fig. 6, D.1).
  • domain assumption Main-predictor residual bias is largely value-conditioned and approximately Lipschitz in the predicted value, so finite bins yield a useful prototype table.
    Used in RBC design and in the Lipschitz argument for prototype approximation error (theoretical appendix).
  • domain assumption Standard inductive kriging protocol: spatial node split, subgraph sampling, observed nodes as anchors for unobserved targets.
    Inherited from IGNNK-style ISK setup (Appendix B.4); defines the evaluation game.
  • standard math Almost-sure convergence of empirical bin prototypes to conditional bias under i.i.d.-style residual collection as sample size grows.
    Invoked in the consistency argument for soft residual retrieval (Theorem-style analysis in appendix).
  • ad hoc to paper Plug-in modules may freeze the backbone after main training and apply post-hoc calibration without re-deriving backbone optimality.
    Two-stage train-then-calibrate procedure in B.6; practical design choice central to RBC.
invented entities (2)
  • Reliability-guided Signal Regulation (RSR) reliability field no independent evidence
    purpose: Score and reweight each observation before spatial-temporal propagation to suppress long-gap or weakly supported entries.
    Method-specific construct defined by the paper’s temporal/spatial reliability formulas; evaluated only via downstream kriging metrics.
  • Value-binned residual prototype table with context-aware correction amplitudes (RBC) no independent evidence
    purpose: Capture systematic over/under-estimation by predicted-value range and apply selective post-hoc correction.
    New calibration object for this framework; prototypes are empirical buffers from training residuals, not independently measured physical quantities.

reviewed 2026-07-15 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Latent-Mark: An Audio Watermark Robust to Neural Codec Compression." pith.science (2026). https://pith.science/paper/AOSJN75G

@misc{pith2026260305310,
  author       = {Pith},
  title        = {Pith review of: Latent-Mark: An Audio Watermark Robust to Neural Codec Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AOSJN75G}},
  note         = {Machine review of arXiv:2603.05310}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

While existing audio watermarking techniques have achieved strong robustness against traditional digital signal processing (DSP) attacks, they remain vulnerable to neural compression. This occurs because modern neural audio codecs act as noise filters and discard the imperceptible waveform variations used in prior watermarking methods. To address this limitation, we propose Latent-Mark, the first zero-bit audio watermarking framework designed to survive neural codec compression. Our key insight is that robustness to the encode-decode process requires embedding the watermark within the codec's invariant latent space. We achieve this by optimizing the audio waveform to induce a detectable directional shift in its encoded latent representation, while constraining perturbations to align with the natural audio manifold to ensure imperceptibility. To prevent overfitting to a single codec's quantization rules, we introduce Cross-Codec Optimization, jointly optimizing the waveform across multiple surrogate codecs to target shared latent invariants. Extensive evaluations demonstrate robust zero-shot transferability to unseen neural codecs, achieving competitive resilience against traditional DSP attacks while preserving perceptual imperceptibility. We hope our work will inspire future research into universal watermarking frameworks capable of maintaining integrity across increasingly complex and diverse generative distortions.

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

48 extracted references · 11 linked inside Pith

  1. [1]

    Kriging convolutional networks

    Gabriel Appleby, Linfeng Liu, and Li-Ping Liu. Kriging convolutional networks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 3187–3194, 2020

    2020

  2. [2]

    Linear interpolation revitalized

    Thierry Blu, Philippe Thévenaz, and Michael Unser. Linear interpolation revitalized. IEEE Transactions on Image Processing, 13(5):710–719, 2004

    2004

  3. [3]

    Spectral networks and locally connected networks on graphs

    Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203, 2013

    Pith/arXiv arXiv 2013

  4. [4]

    Tianfeng Chai and Roland R. Draxler. Root mean square error (rmse) or mean absolute error (mae)? arguments against avoiding rmse in the literature. Geoscientific Model Development, 7(3):1247–1250, 2014

    2014

  5. [5]

    Empirical evaluation of gated recurrent neural networks on sequence modeling

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014

    Pith/arXiv arXiv 2014

  6. [7]

    Multivariate time series imputation by graph neural networks

    Andrea Cini, Ivan Marisca, and Cesare Alippi. Multivariate time series imputation by graph neural networks. corr abs/2108.00298 (2021). arXiv preprint arXiv:2108.00298, 2021

    Pith/arXiv arXiv 2021

  7. [8]

    Latent space model for road networks to predict time-varying traffic

    Dingxiong Deng, Cyrus Shahabi, Ugur Demiryurek, Linhong Zhu, Rose Yu, and Yan Liu. Latent space model for road networks to predict time-varying traffic. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1525–1534, 2016

    2016

  8. [9]

    Saits: Self-attention-based imputation for time series

    Wenjie Du, David Côté, and Yan Liu. Saits: Self-attention-based imputation for time series. Expert Systems with Applications, 219:119619, 2023

    2023

  9. [10]

    Tsi-bench: Benchmarking time series imputation

    Wenjie Du, Jun Wang, Linglong Qian, Yiyuan Yang, Zina Ibrahim, Fanxing Liu, Zepu Wang, Haoxin Liu, Zhiyuan Zhao, Yingjie Zhou, et al. Tsi-bench: Benchmarking time series imputation. arXiv preprint arXiv:2406.12747, 2024

    Pith/arXiv arXiv 2024

  10. [11]

    Reduction of partial-volume artifacts with zero-filled interpolation in three-dimensional mr angiography

    Yiping P Du, Dennis L Parker, Wayne L Davis, and Guang Cao. Reduction of partial-volume artifacts with zero-filled interpolation in three-dimensional mr angiography. Journal of Magnetic Resonance Imaging, 4 (5):733–741, 1994

    1994

  11. [12]

    Deep learning- based uncertainty quantification for spatio-temporal environmental remote sensing: A systematic literature review

    Aya Ferchichi, Ahlem Ferchichi, Fatma Hendaoui, Mejda Chihaoui, and Radhia Toujani. Deep learning- based uncertainty quantification for spatio-temporal environmental remote sensing: A systematic literature review. Neurocomputing, page 130242, 2025

    2025

  12. [13]

    Neural message passing for quantum chemistry

    Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International conference on machine learning, pages 1263–1272. Pmlr, 2017

    2017

  13. [14]

    Hyndman and Anne B

    Rob J. Hyndman and Anne B. Koehler. Another look at measures of forecast accuracy. International Journal of Forecasting, 22(4):679–688, 2006

    2006

  14. [15]

    Graph neural network for traffic forecasting: A survey

    Weiwei Jiang and Jiayun Luo. Graph neural network for traffic forecasting: A survey. Expert systems with applications, 207:117921, 2022

    2022

  15. [16]

    Spatio-temporal graph neural networks for predictive learning in urban computing: A survey

    Guangyin Jin, Yuxuan Liang, Yuchen Fang, Zezhi Shao, Jincai Huang, Junbo Zhang, and Yu Zheng. Spatio-temporal graph neural networks for predictive learning in urban computing: A survey. IEEE transactions on knowledge and data engineering, 36(10):5388–5408, 2023

    2023

  16. [17]

    A survey on graph neural networks for time series: Forecasting, classification, imputation, and anomaly detection

    Ming Jin, Huan Yee Koh, Qingsong Wen, Daniele Zambon, Cesare Alippi, Geoffrey I Webb, Irwin King, and Shirui Pan. A survey on graph neural networks for time series: Forecasting, classification, imputation, and anomaly detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  17. [18]

    Semi-supervised classification with graph convolutional networks

    TN Kipf. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

    Pith/arXiv arXiv 2016

  18. [19]

    Dynamic graph convolutional recurrent imputation network for spatiotemporal traffic missing data

    Xiangjie Kong, Wenfeng Zhou, Guojiang Shen, Wenyi Zhang, Nali Liu, and Yao Yang. Dynamic graph convolutional recurrent imputation network for spatiotemporal traffic missing data. Knowledge-Based Systems, 261:110188, 2023

    2023

  19. [20]

    High-resolution multi-source traffic data in new zealand

    Bo Li, Ruotao Yu, Zijun Chen, Yingzhe Ding, Mingxia Yang, Jinghua Li, Jianxiao Wang, and Haiwang Zhong. High-resolution multi-source traffic data in new zealand. Scientific Data, 11(1):1216, 2024. 11

    2024

  20. [21]

    A spatiotemporal approach for traffic data imputation with complicated missing patterns

    Huiping Li, Meng Li, Xi Lin, Fang He, and Yinhai Wang. A spatiotemporal approach for traffic data imputation with complicated missing patterns. Transportation research part C: emerging technologies, 119:102730, 2020

    2020

  21. [22]

    Diffusion convolutional recurrent neural network: Data-driven traffic forecasting

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926, 2017

    Pith/arXiv arXiv 2017

  22. [23]

    Sta-gann: A valid and generalizable spatio-temporal kriging approach

    Yujie Li, Shao Zezhi, Chengqing Yu, Tangwen Qian, Zhao Zhang, Yifan Du, Shaoming He, Fei Wang, and Yongjun Xu. Sta-gann: A valid and generalizable spatio-temporal kriging approach. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management, pages 1726–1736, 2025

    2025

  23. [24]

    Memory-augmented dynamic graph convolution networks for traffic data imputation with diverse missing patterns

    Yuebing Liang, Zhan Zhao, and Lijun Sun. Memory-augmented dynamic graph convolution networks for traffic data imputation with diverse missing patterns. Transportation Research Part C: Emerging Technologies, 143:103826, 2022

    2022

  24. [25]

    Darkfarseer: Inductive spatio-temporal kriging via hidden style enhancement and sparsity-noise mitigation

    Zhuoxuan Liang, Wei Li, Dalin Zhang, Yidan Chen, Zhihong Wang, Xiangping Zheng, and Moustafa Youssef. Darkfarseer: Inductive spatio-temporal kriging via hidden style enhancement and sparsity-noise mitigation. arXiv preprint arXiv:2501.02808, 2025

    arXiv 2025

  25. [26]

    Cubic spline interpolation

    Sky McKinley and Megan Levine. Cubic spline interpolation. College of the Redwoods, 45(1):1049–1060, 1998

    1998

  26. [27]

    Imputeformer: Low rankness-induced transformers for generalizable spatiotemporal imputation

    Tong Nie, Guoyang Qin, Wei Ma, Yuewen Mei, and Jian Sun. Imputeformer: Low rankness-induced transformers for generalizable spatiotemporal imputation. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 2260–2271, 2024

    2024

  27. [28]

    Ppca-based missing data imputation for traffic flow volume: A systematical approach

    Li Qu, Li Li, Yi Zhang, and Jianming Hu. Ppca-based missing data imputation for traffic flow volume: A systematical approach. IEEE Transactions on intelligent transportation systems, 10(3):512–522, 2009

    2009

  28. [29]

    Mean circulatory filling pressure: its meaning and measurement

    CARL F Rothe. Mean circulatory filling pressure: its meaning and measurement. Journal of Applied Physiology, 74(2):499–509, 1993

    1993

  29. [30]

    Machine learning based approaches for imputation in time series data and their impact on forecasting

    Muhammad Saad, Mohita Chaudhary, Fakhri Karray, and Vincent Gaudet. Machine learning based approaches for imputation in time series data and their impact on forecasting. In 2020 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 2621–2627. IEEE, 2020

    2020

  30. [31]

    Pre-training enhanced spatial-temporal graph neural network for multivariate time series forecasting

    Zezhi Shao, Zhao Zhang, Fei Wang, and Yongjun Xu. Pre-training enhanced spatial-temporal graph neural network for multivariate time series forecasting. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, pages 1567–1577, 2022

    2022

  31. [32]

    Decoupled dynamic spatial-temporal graph neural network for traffic forecasting

    Zezhi Shao, Zhao Zhang, Wei Wei, Fei Wang, Yongjun Xu, Xin Cao, and Christian S Jensen. Decoupled dynamic spatial-temporal graph neural network for traffic forecasting. arXiv preprint arXiv:2206.09112, 2022

    Pith/arXiv arXiv 2022

  32. [33]

    Bidirectional spatial–temporal traffic data imputation via graph attention recurrent neural network

    Guojiang Shen, Wenfeng Zhou, Wenyi Zhang, Nali Liu, Zhi Liu, and Xiangjie Kong. Bidirectional spatial–temporal traffic data imputation via graph attention recurrent neural network. Neurocomputing, 531:151–162, 2023

    2023

  33. [34]

    Graph attention networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

    Pith/arXiv arXiv 2017

  34. [35]

    Inferring inter- section traffic patterns with sparse video surveillance information: An st-gan method

    Pengkun Wang, Chaochao Zhu, Xu Wang, Zhengyang Zhou, Guang Wang, and Yang Wang. Inferring inter- section traffic patterns with sparse video surveillance information: An st-gan method. IEEE Transactions on Vehicular Technology, 71(9):9840–9852, 2022

    2022

  35. [36]

    Stamim- puter: Spatio-temporal attention moe for traffic data imputation

    Yiming Wang, Hao Peng, Senzhang Wang, Haohua Du, Chunyang Liu, Jia Wu, and Guanlin Wu. Stamim- puter: Spatio-temporal attention moe for traffic data imputation. arXiv preprint arXiv:2506.08054, 2025

    Pith/arXiv arXiv 2025

  36. [37]

    Task-oriented time series imputation evaluation via generalized representers

    Zhixian Wang, Linxiao Yang, Liang Sun, Qingsong Wen, and Yi Wang. Task-oriented time series imputation evaluation via generalized representers. Advances in Neural Information Processing Systems, 37:137403–137431, 2024

    2024

  37. [38]

    Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance

    Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. Climate research, 30(1):79–82, 2005

    2005

  38. [39]

    Inductive graph neural networks for spatiotemporal kriging

    Yuankai Wu, Dingyi Zhuang, Aurelie Labbe, and Lijun Sun. Inductive graph neural networks for spatiotemporal kriging. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 4478–4485, 2021. 12

    2021

  39. [40]

    Spatial aggregation and temporal convolution networks for real-time kriging

    Yuankai Wu, Dingyi Zhuang, Mengying Lei, Aurelie Labbe, and Lijun Sun. Spatial aggregation and temporal convolution networks for real-time kriging. arXiv preprint arXiv:2109.12144, 2021

    Pith/arXiv arXiv 2021

  40. [41]

    A comprehen- sive survey on graph neural networks

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and Philip S Yu. A comprehen- sive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 32(1): 4–24, 2020

    2020

  41. [42]

    Kits: Inductive spatio- temporal kriging with increment training strategy

    Qianxiong Xu, Cheng Long, Ziyue Li, Sijie Ruan, Rui Zhao, and Zhishuai Li. Kits: Inductive spatio- temporal kriging with increment training strategy. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 12945–12953, 2025

    2025

  42. [43]

    Spatial-temporal traffic data imputation via graph attention convolutional network

    Yongchao Ye, Shiyao Zhang, and James JQ Yu. Spatial-temporal traffic data imputation via graph attention convolutional network. In International Conference on artificial neural networks, pages 241–252. Springer, 2021

    2021

  43. [44]

    Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting

    Bing Yu, Haoteng Yin, and Zhanxing Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. arXiv preprint arXiv:1709.04875, 2017

    Pith/arXiv arXiv 2017

  44. [45]

    Stgan: Spatio-temporal generative adversarial network for traffic data imputation

    Ye Yuan, Yong Zhang, Boyue Wang, Yuan Peng, Yongli Hu, and Baocai Yin. Stgan: Spatio-temporal generative adversarial network for traffic data imputation. IEEE Transactions on Big Data, 9(1):200–211, 2022

    2022

  45. [46]

    Strap: Spatio- temporal pattern retrieval for out-of-distribution generalization, 2025

    Haoyu Zhang, Wentao Zhang, Hao Miao, Xinke Jiang, Yuchen Fang, and Yifan Zhang. Strap: Spatio- temporal pattern retrieval for out-of-distribution generalization, 2025. URL ���������������������� ����������

    2025

  46. [47]

    Nearest neighbor selection for iteratively knn imputation.Journal of Systems and Software, 85(11):2541–2552, 2012

    Shichao Zhang. Nearest neighbor selection for iteratively knn imputation.Journal of Systems and Software, 85(11):2541–2552, 2012

    2012

  47. [48]

    Gman: A graph multi-attention network for traffic prediction

    Chuanpan Zheng, Xiaoliang Fan, Cheng Wang, and Jianzhong Qi. Gman: A graph multi-attention network for traffic prediction. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 1234–1241, 2020

    2020

  48. [49]

    Increase: Inductive graph representation learning for spatio-temporal kriging

    Chuanpan Zheng, Xiaoliang Fan, Cheng Wang, Jianzhong Qi, Chaochao Chen, and Longbiao Chen. Increase: Inductive graph representation learning for spatio-temporal kriging. In Proceedings of the ACM Web Conference 2023, pages 673–683, 2023. 13 Appendices A Notation 15 B Implementation Details 16 B.1 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . ...

    2023

This paper was first reviewed by grok-4.5 on July 15, 2026.