Pith. sign in

REVIEW 4 major objections 9 minor 1 cited by

STAMImputer: Spatio-Temporal Attention MoE for Traffic Data Imputation

T0 review · 4 major / 9 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read STAMImputer proposes that routing spatio-temporal attention through a mixture of experts, with low-rank guided sampling graph attention building dynamic graphs, yields lower imputation error than existing state-of-the-art methods on four…

desk verdict A plausible new architecture for traffic imputation that overclaims its 'significant' gains and needs error bars and a sampling ablation, but deserves peer review. read the letter →

arxiv 2506.08054 v2 pith:QCM7WXKH submitted 2025-06-09 cs.LG cs.AI

Yiming Wang , Hao Peng , Senzhang Wang , Haohua Du , Chunyang Liu , Jia Wu , Guanlin Wu This is my paper · ORCID
classification cs.LGcs.AI
keywords trafficdataimputationmixtureofexpertsspatio-temporalattentionlow-rankgraphdynamicstructurelearningmissingintelligenttransportationsystems
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that traffic data imputation improves when the model dynamically balances temporal and spatial evidence instead of processing time before space in a fixed order. STAMImputer is a mixture-of-experts network in which an observation expert scores how much to trust the temporal and spatial attention experts based on the current missing pattern and sparsity. The spatial expert, LrSGAT, scores sensors by local graph attention, samples the top $S=\lceil\log N\rceil$ nodes plus an equal number of random nodes, and uses the sampled attention both to drive a low-rank re-attention step and to build a semi-adaptive dynamic graph of real-time road-network correlations. On four benchmarks, under point missing (25% and 60%) and block missing (failure probabilities 0.2% and 1%, roughly 10% and 30% total), the paper reports the lowest MAE in nearly every configuration, with a few second-place results, and shows the learned dynamic graph improves downstream Graph-Wavenet prediction. If true, the significance is that block missing and nonstationary traffic can be handled by letting the data decide which spatial and temporal signals to trust.

What carries the argument

The load-bearing mechanism is LrSGAT (Low-rank guided Sampling Graph ATtention), a spatial attention block that scores each sensor's local influence through graph attention over a static topology, then keeps the top $\lceil\log N\rceil$ sensors plus an equal number of randomly sampled sensors. The sampled attention vectors act as a compact projector: a re-attention layer uses them to compress the global spatial matrix into a low-rank form and then reconstruct node features, filtering redundant global relationships while propagating hub information. The same sampled attention feeds a DGSL (Dynamic Graph Structure Learning) function that produces a semi-adaptive adjacency matrix at each time step. Around this sits the MoE framework, whose observation expert consumes raw values and computed sparsity features and outputs softmax confidence scores that weight the temporal and spatial expert outputs before the readout MLP.

What would settle it

Replace the significance-scored half of the sampled nodes with random nodes of the same size on DiDi-SZ under block missing; if MAE does not rise substantially, the importance-scoring and hub-propagation mechanism is not what drives the reported gains.

Watch

Extended reading notes

Core claim

The central claim is that a mixture-of-experts arrangement of attention experts generalizes better than fixed time-to-space imputation pipelines. The model pairs a temporal expert (multi-head self-attention) and a spatial expert (LrSGAT) under the control of a feed-forward observation expert that softmax-weights each expert's output per node and time step. LrSGAT is the load-bearing piece: it computes local attention over a static K-nearest-neighbour graph, scores each sensor, samples the top $\lceil\log N\rceil$ sensors plus $\lceil\log N\rceil$ random ones, and projects the attention through that sample to re-attend under a low-rank factorization, which the paper interprets as compressing and reconstructing global spatial information; the same sampled attention generates a semi-adaptive dynamic graph. The paper reports that STAMImputer generally achieves the best MAE across PemsD8, SZ-Taxi, DiDi-SZ, and NYC-Taxi for both point and block missing, and that adding its dynamic graph to Graph-Wavenet improves downstream traffic prediction, particularly under high sparsity or block missing.

Load-bearing premise

The model's edge depends on the assumption that a small sample of sensors—the most locally influential plus a few random ones—carries enough of the city-wide traffic pattern to fill in missing blocks accurately.

Editorial extensions

If this is right

  • In the reported settings (point missing at 25% and 60%; block missing at failure probabilities 0.2% and 1%), STAMImputer achieves the lowest MAE on nearly every dataset-configuration cell and is second in the remaining cells.
  • The semi-adaptive dynamic graph produced by DGSL transfers to downstream prediction: adding the pretrained imputer plus dynamic graph to Graph-Wavenet lowers MAE on NYC-Taxi compared with using the static topology alone, with the largest gains under sparse or block-missing inputs.
  • The total complexity is $O(T^2D+N\log N D+NTD)$, so the hybrid sampling keeps the spatial cost near-linear in the number of sensors despite full attention on the temporal dimension.
  • Ablations attribute the largest loss to removing LrSGAT, especially under block missing; removing MoE, the sampling projector, or the wavelet embedding each also raises MAE, supporting the claim that every component contributes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to make the sample size $S$ or the top/random split depend on the observed sparsity or on the geometry of missing blocks; the paper fixes $\lceil\log N\rceil$ and reports sensitivity to hyperparameters but does not ablate the sampling rule itself.
  • The observation expert's per-node, per-time confidence scores could be exported as a reliability map, which might support sensor maintenance or adaptive data collection decisions.
  • If the dynamic graph is capturing real-time bottlenecks, its edge weights should shift toward a disrupted area after an incident; that is a concrete downstream prediction that could be checked with labelled incident data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 9 minor

Summary. The manuscript proposes STAMImputer, a spatio-temporal traffic data imputation model built on a mixture-of-experts (MoE) framework. The temporal expert uses multi-head self-attention; the spatial expert (LrSGAT) extracts local graph attention via a static topology, samples a small set of significant and random nodes (S = ceil(log N)), and uses the sampled attention vectors both for low-rank guided re-attention and for constructing a semi-adaptive dynamic graph. An observation expert weights the attention experts based on raw input and sparsity features, and wavelet decomposition is used to add frequency features. The paper reports MAE comparisons against nine baselines on PemsD8, SZ-Taxi, DiDi-SZ, and NYC-Taxi under point and block missing patterns, plus a downstream Graph-Wavenet prediction experiment and ablation/robustness studies on NYC-Taxi.

Significance. If the reported gains are reliable, STAMImputer would be a meaningful new architectural contribution to traffic imputation, since the MoE gating of spatio-temporal experts and the sampled low-rank attention are not combined in prior work. The paper has the right shape for such a claim: it evaluates on four public benchmarks against published baselines, includes ablation and downstream-task evidence, and provides a code link. However, the current evidence is single-run MAE with no variance or significance testing, and the paper's own Table 1 shows the main baseline ahead in three of four PemsD8 configurations; the significance claim therefore does not yet hold.

major comments (4)
  1. [Abstract; Section 5.1, Table 1] The central claim that STAMImputer 'achieves significantly performance improvement compared with existing SOTA approaches' is not supported by the reported evidence. Table 1 lists only one MAE per condition with no standard deviations, number of seeds, or statistical tests, and the paper's own numbers contradict the claim on PemsD8: ImputeFormer is better at 25% point missing (11.01 vs 11.41), at 0.2% block missing (15.07 vs 15.38), and at 1% block missing (17.56 vs 18.06). Since the paper's contribution is empirical superiority, the evaluation must report variance over multiple runs and either perform significance tests or temper the abstract's wording.
  2. [Sections 5.2, 5.3, and Table 3] The robustness, ablation, and downstream claims suffer from the same missing statistical evidence. Figure 2 and Figure 3 show MAE curves and bars without error bars or seed counts, so the purported robustness of STAMImputer and the contribution of each component cannot be separated from run-to-run noise. Table 3 reports single MAE values (e.g., 8.03 vs 8.39 for GWNet+STAMImputer+DGSL vs GWNet+STAMImputer at 25% point missing), making the claimed DGSL improvement unverifiable.
  3. [Section 4.2, Eq. (7)] The hybrid sampling rule underlying LrSGAT is a heuristic: S = ceil(log N), taking the S highest-significance nodes plus S randomly sampled remaining nodes. The paper gives no proof or systematic ablation that this rule preserves the global spatial correlations needed for the low-rank guided re-attention and dynamic graph construction. The authors should either provide a formal characterization of the approximation error relative to full attention or ablate the components of the sampling strategy (top-only, random-only, hybrid, and varying S) with variance reported; the brief mention of hyperparameter sensitivity in Figure 5 is not sufficient.
  4. [Section 4.2, Eq. (14)] The dynamic graph construction depends on the toph(·) function, described only as 'setting terms less than the median to zero.' It is unclear whether the median is taken over the whole matrix or per row, and whether the threshold is a hard mask or a differentiable approximation. This must be specified for reproducibility, and the choice should be justified with an ablation.
minor comments (9)
  1. [Abstract] The phrase 'significantly performance improvement' should be 'significant performance improvement' (and, given the statistical concerns above, the claim should be revised).
  2. [Section 1] The word 'demension' should be 'dimension'.
  3. [Section 5, Experimental Setup] The baseline name is given as 'ImputerFormer' but everywhere else it is 'ImputeFormer'; use consistent spelling.
  4. [Table 3 title] The title contains the typo 'Graph-Wavenent'; it should be 'Graph-Wavenet'.
  5. [Figure 1 caption and Section 4] The in-text references to 'Figure.1', 'Figure 1.1', and 'Figure 1.2' are confusing; use standard subfigure labels.
  6. [Section 4.2 and Eq. (7)] The notation ⟨·∥·⟩ is used for concatenation in Eq. (7) and Eq. (18), but the same angle-bracket notation is also used for tuples; please define this notation once and use it consistently.
  7. [Abstract, code link] The code URL, https://github.com/RingBDStack/STAMImupter, appears to misspell 'Imputer' as 'Imupter'; please verify the link.
  8. [Section 5.1] The description of block missing as 'both with a random missing rate of 5%' is unclear; specify how the 5% random missing rate relates to the failure probabilities 0.2% and 1%.
  9. [Figure 4 caption] The caption says 'Case Study of Efficiency' but the figure shows imputation visualizations, not runtime or efficiency; rename the caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims rest on external benchmark comparisons and independent ablations, not on self-referential definitions or fitted inputs.

full rationale

The paper's central claims are empirical: STAMImputer achieves lower MAE than published baselines on four external benchmarks (Table 1) and that its DGSL module improves downstream Graph-Wavenet prediction (Table 3). These claims are evaluated against external benchmark datasets and prior published methods, not against quantities defined by the model itself. The low-rank guided sampling and dynamic graph construction are architectural proposals tested by ablation and downstream experiments; no prediction is produced by fitting a parameter to the target metric and then renaming it as a result. Self-citations appear only in related work and design motivation (e.g., [Zou et al. 2023] on limitations of adaptive adjacency matrices), and none of them supplies the benchmark results or constitutes the load-bearing evidence for the SOTA claim. There is no equation-level identity between inputs and outputs that would make the reported improvements true by construction. The absence of variance reporting and significance tests, and the inconsistency on PemsD8 where ImputeFormer is better in 3 of 4 settings, are correctness and robustness concerns, not circularity. Hence score 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The ledger is light: the paper relies on standard benchmarks and well-known architectural building blocks. The most debatable assumption is the low-rank sampling heuristic in LrSGAT, which is central to the spatial expert and dynamic graph construction but lacks formal grounding. Hyperparameters such as S, wavelet level, and attention settings are chosen by hand or tuned, and exact values are not reported.

free parameters (4)
  • number of sampled nodes S = ceil(log N) per timestep
    Chosen by hand as S=ceil(log N) in Eq. 7; controls the tradeoff between representativeness and complexity.
  • wavelet decomposition level j = not reported
    Hyperparameter selected via sensitivity analysis (Fig. 5); affects low/high frequency decomposition.
  • number of attention heads and layers = not reported
    Hyperparameters likely tuned per dataset; no values given in text.
  • toph threshold in Eq. 14 = median
    Median threshold set to zero; a heuristic choice.
assumptions (4)
  • domain assumption K-nearest neighbors graph G_K is provided and meaningful for spatial correlations
    Used as static topology in Eq. 6; errors in G_K would propagate into attention and imputation.
  • ad hoc to paper Spatial interaction matrix can be approximated by low-rank factorization with sampled nodes as cohesive factors
    Section 4.2 'Low-rank Guided Re-attention' and Appendix A: no proof that sampled attention approximates the low-rank factor; heuristic.
  • domain assumption Missing data generation protocol (point and block missing) mimics real-world sensor failures
    Section 5: following Cini et al. 2021; if real missingness differs, benchmark results may not transfer.
  • domain assumption Wavelet transform with chosen basis provides useful features
    Section 4.3: DWT decomposition; basis and level are not justified in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of STAMImputer: Spatio-Temporal Attention MoE for Traffic Data Imputation." pith.science (2026). https://pith.science/paper/QCM7WXKH

@misc{pith2026250608054,
  author       = {Pith},
  title        = {Pith review of: STAMImputer: Spatio-Temporal Attention MoE for Traffic Data Imputation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QCM7WXKH}},
  note         = {Machine review of arXiv:2506.08054}
}
read the original abstract

Traffic data imputation is fundamentally important to support various applications in intelligent transportation systems such as traffic flow prediction. However, existing time-to-space sequential methods often fail to effectively extract features in block-wise missing data scenarios. Meanwhile, the static graph structure for spatial feature propagation significantly constrains the models flexibility in handling the distribution shift issue for the nonstationary traffic data. To address these issues, this paper proposes a SpatioTemporal Attention Mixture of experts network named STAMImputer for traffic data imputation. Specifically, we introduce a Mixture of Experts (MoE) framework to capture latent spatio-temporal features and their influence weights, effectively imputing block missing. A novel Low-rank guided Sampling Graph ATtention (LrSGAT) mechanism is designed to dynamically balance the local and global correlations across road networks. The sampled attention vectors are utilized to generate dynamic graphs that capture real-time spatial correlations. Extensive experiments are conducted on four traffic datasets for evaluation. The result shows STAMImputer achieves significantly performance improvement compared with existing SOTA approaches. Our codes are available at https://github.com/RingBDStack/STAMImupter.

Figures

Figures reproduced from arXiv: 2506.08054 by the authors.

Figure 1
Figure 1. The overview of STAMImputer and LrSGAT: (1) The framework of MoE contains multi-head self-attention networks as Temporal [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Robustness analysis on NYC-Taxi benchmark. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Ablation studies result on NYC-Taxi benchmark. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Case Study of Hyperparameters on NYC-Taxi benchmark. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 4
Figure 4. Figure 4: Case Study of Efficiency on PEMSD8. Appendix A Discussion of Semi-adaptive Dynamic Graph Structure Learning The semi-adaptive approach reveals how sampled attention approximates the cohesive factor through dynamic commu￾nity representation to maintain spatial relations…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

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

    cs.SD 2026-03 conditional novelty 5.0 of 10

    A reliability-guided regulation plus residual-bias calibration plug-in consistently improves inductive spatio-temporal kriging under incomplete and block-missing sensor observations.

Reference graph

Works this paper leans on

43 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    Traffic data imputation using deep convolu- tional neural networks.IEEE Access, 8:104740–104752,

    [Benkraoudaet al., 2020 ] Ouafa Benkraouda, Bilal Thon- nam Thodi, Hwasoo Yeo, Monica Menendez, and Saif Ed- din Jabari. Traffic data imputation using deep convolu- tional neural networks.IEEE Access, 8:104740–104752,

    work page 2020

  2. [8]

    Low-rank autoregressive ten- sor completion for spatiotemporal traffic data imputation

    [Chenet al., 2021 ] Xinyu Chen, Mengying Lei, Nicolas Saunier, and Lijun Sun. Low-rank autoregressive ten- sor completion for spatiotemporal traffic data imputation. IEEE Transactions on Intelligent Transportation Systems, 23(8):12301–12310,

    work page 2021

  3. [9]

    [Cichocki and Phan, 2009] Andrzej Cichocki and Anh-Huy Phan. Fast local algorithms for large scale nonnegative ma- trix and tensor factorizations.IEICE transactions on fun- damentals of electronics, communications and computer sciences, 92(3):708–721,

    work page 2009

  4. [11]

    Graph spectral regularized tensor completion for traffic data imputation

    [Denget al., 2021 ] Lei Deng, Xiao-Yang Liu, Haifeng Zheng, Xinxin Feng, and Youjia Chen. Graph spectral regularized tensor completion for traffic data imputation. IEEE Transactions on Intelligent Transportation Systems, 23(8):10996–11010,

    work page 2021

  5. [12]

    Saits: Self-attention-based imputation for time series.Expert Systems with Applications, 219:119619,

    [Duet al., 2023 ] Wenjie Du, David Cˆot´e, and Yan Liu. Saits: Self-attention-based imputation for time series.Expert Systems with Applications, 219:119619,

    work page 2023

  6. [13]

    When spatio-temporal meet wavelets: Disentan- gled traffic forecasting via efficient spectral graph attention networks

    [Fanget al., 2023 ] Yuchen Fang, Yanjun Qin, Haiyong Luo, Fang Zhao, Bingbing Xu, Liang Zeng, and Chenxing Wang. When spatio-temporal meet wavelets: Disentan- gled traffic forecasting via efficient spectral graph attention networks. In2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 517–529. IEEE,

    work page 2023

  7. [14]

    Switch transformers: Scaling to trillion param- eter models with simple and efficient sparsity.Journal of Machine Learning Research, 23(120):1–39,

    [Feduset al., 2022 ] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion param- eter models with simple and efficient sparsity.Journal of Machine Learning Research, 23(120):1–39,

    2022

  8. [15]

    Dy- namic graph convolutional recurrent imputation network for spatiotemporal traffic missing data.Knowledge-Based Systems, 261:110188,

    [Konget al., 2023 ] Xiangjie Kong, Wenfeng Zhou, Guo- jiang Shen, Wenyi Zhang, Nali Liu, and Yao Yang. Dy- namic graph convolutional recurrent imputation network for spatiotemporal traffic missing data.Knowledge-Based Systems, 261:110188,

    work page 2023

Show all 43 references
  1. [16]

    Effi- cient missing data imputing for traffic flow by considering temporal and spatial dependence.Transportation research part C: emerging technologies, 34:108–120,

    [Liet al., 2013 ] Li Li, Yuebiao Li, and Zhiheng Li. Effi- cient missing data imputing for traffic flow by considering temporal and spatial dependence.Transportation research part C: emerging technologies, 34:108–120,

    2013

  2. [18]

    [Lianget al., 2022 ] Yuebing Liang, Zhan Zhao, and Lijun Sun. Memory-augmented dynamic graph convolution net- works for traffic data imputation with diverse missing pat- terns.Transportation Research Part C: Emerging Tech- nologies, 143:103826,

    2022

  3. [19]

    Cross-city few-shot traffic forecasting via traffic pat- tern bank

    [Liuet al., 2023 ] Zhanyu Liu, Guanjie Zheng, and Yanwei Yu. Cross-city few-shot traffic forecasting via traffic pat- tern bank. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 1451–1460,

    2023

  4. [20]

    Cdsa: cross-dimensional self-attention for multivari- ate, geo-tagged time series imputation.arXiv preprint arXiv:1905.09904,

    [Maet al., 2019 ] Jiawei Ma, Zheng Shou, Alireza Zareian, Hassan Mansour, Anthony Vetro, and Shih-Fu Chang. Cdsa: cross-dimensional self-attention for multivari- ate, geo-tagged time series imputation.arXiv preprint arXiv:1905.09904,

    2019 arXiv

  5. [21]

    Learning to reconstruct missing data from spa- tiotemporal graphs with sparse observations.Advances in Neural Information Processing Systems, 35:32069–32082,

    [Mariscaet al., 2022 ] Ivan Marisca, Andrea Cini, and Ce- sare Alippi. Learning to reconstruct missing data from spa- tiotemporal graphs with sparse observations.Advances in Neural Information Processing Systems, 35:32069–32082,

    2022

  6. [22]

    Missing data: A comparison of neural network and expectation maxi- mization techniques.Current Science, pages 1514–1521,

    [Nelwamondoet al., 2007 ] Fulufhelo V Nelwamondo, Shakir Mohamed, and Tshilidzi Marwala. Missing data: A comparison of neural network and expectation maxi- mization techniques.Current Science, pages 1514–1521,

    2007

  7. [27]

    Self-attention graph convolution imputation network for spatio-temporal traffic data.IEEE Transactions on Intelligent Transportation Sys- tems,

    [Weiet al., 2024 ] Xiulan Wei, Yong Zhang, Shaofan Wang, Xia Zhao, Yongli Hu, and Baocai Yin. Self-attention graph convolution imputation network for spatio-temporal traffic data.IEEE Transactions on Intelligent Transportation Sys- tems,

    2024

  8. [28]

    Graph wavenet for deep spatial-temporal graph modeling.arXiv preprint arXiv:1906.00121,

    [Wuet al., 2019 ] Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. Graph wavenet for deep spatial-temporal graph modeling.arXiv preprint arXiv:1906.00121,

    2019 arXiv

  9. [29]

    Traffic speed imputation with spatio-temporal attentions and cycle-perceptual training

    [Xuet al., 2022 ] Qianxiong Xu, Sijie Ruan, Cheng Long, Liang Yu, and Chen Zhang. Traffic speed imputation with spatio-temporal attentions and cycle-perceptual training. InProceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 2280– 2289,

    2022

  10. [30]

    Hrst-lr: a hessian regularization spatio- temporal low rank algorithm for traffic data imputation

    [Xuet al., 2023 ] Xiuqin Xu, Mingwei Lin, Xin Luo, and Zeshui Xu. Hrst-lr: a hessian regularization spatio- temporal low rank algorithm for traffic data imputation. IEEE Transactions on Intelligent Transportation Systems, 24(10):11001–11017,

    2023

  11. [31]

    Hierarchical spatio-temporal graph convo- lutional neural networks for traffic data imputation.Infor- mation Fusion, 106:102292,

    [Xuet al., 2024 ] Dongwei Xu, Hang Peng, Yufu Tang, and Haifeng Guo. Hierarchical spatio-temporal graph convo- lutional neural networks for traffic data imputation.Infor- mation Fusion, 106:102292,

    2024

  12. [32]

    Spatial-temporal traffic data imputation via graph attention convolutional network

    [Yeet al., 2021 ] Yongchao Ye, Shiyao Zhang, and James JQ Yu. Spatial-temporal traffic data imputation via graph attention convolutional network. InInternational Con- ference on Artificial Neural Networks, pages 241–252. Springer,

    2021

  13. [33]

    St-mvl: Filling missing values in geo-sensory time series data

    [Yiet al., 2016 ] Xiuwen Yi, Yu Zheng, Junbo Zhang, and Tianrui Li. St-mvl: Filling missing values in geo-sensory time series data. InProceedings of the 25th international joint conference on artificial intelligence,

    2016

  14. [35]

    Gain: Missing data imputation using gen- erative adversarial nets

    [Yoonet al., 2018 ] Jinsung Yoon, James Jordon, and Mi- haela Schaar. Gain: Missing data imputation using gen- erative adversarial nets. InInternational conference on machine learning, pages 5689–5698. PMLR,

    2018

  15. [36]

    Temporal regularized matrix factorization for high-dimensional time series prediction.Advances in neu- ral information processing systems, 29,

    [Yuet al., 2016 ] Hsiang-Fu Yu, Nikhil Rao, and Inderjit S Dhillon. Temporal regularized matrix factorization for high-dimensional time series prediction.Advances in neu- ral information processing systems, 29,

    2016

  16. [37]

    Stgan: Spatio- temporal generative adversarial network for traffic data imputation.IEEE Transactions on Big Data, 9(1):200– 211,

    [Yuanet al., 2022 ] 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

  17. [38]

    Moefication: Transformer feed-forward layers are mixtures of experts

    [Zhanget al., 2021 ] Zhengyan Zhang, Yankai Lin, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. Moefication: Transformer feed-forward layers are mixtures of experts. arXiv preprint arXiv:2110.01786,

    2021 arXiv

  18. [39]

    Self-attention graph convo- lution residual network for traffic data completion.IEEE Transactions on Big Data, 9(2):528–541,

    [Zhanget al., 2022 ] Yong Zhang, Xiulan Wei, Xinyu Zhang, Yongli Hu, and Baocai Yin. Self-attention graph convo- lution residual network for traffic data completion.IEEE Transactions on Big Data, 9(2):528–541,

    2022

  19. [40]

    Score-cdm: Score-weighted convolutional diffu- sion model for multivariate time series imputation.arXiv preprint arXiv:2405.13075,

    [Zhanget al., 2024a ] Shunyang Zhang, Senzhang Wang, Hao Miao, Hao Chen, Changjun Fan, and Jian Zhang. Score-cdm: Score-weighted convolutional diffu- sion model for multivariate time series imputation.arXiv preprint arXiv:2405.13075,

    arXiv

  20. [41]

    Traffic data imputation and prediction: An efficient realization of deep learning.IEEE Access, 8:46713–46722,

    [Zhaoet al., 2020 ] Junhui Zhao, Yiwen Nie, Shanjin Ni, and Xiaoke Sun. Traffic data imputation and prediction: An efficient realization of deep learning.IEEE Access, 8:46713–46722,

    2020

  21. [42]

    Se-gsl: A general and effective graph struc- ture learning framework through structural entropy opti- mization

    [Zouet al., 2023 ] Dongcheng Zou, Hao Peng, Xiang Huang, Renyu Yang, Jianxin Li, Jia Wu, Chunyang Liu, and Philip S Yu. Se-gsl: A general and effective graph struc- ture learning framework through structural entropy opti- mization. InProceedings of the ACM Web Conference 2023,...

    2023

  22. [43]

    Multispans: a multi-range spatial-temporal transformer network for traffic forecast via structural entropy optimization

    [Zouet al., 2024 ] Dongcheng Zou, Senzhang Wang, Xue- feng Li, Hao Peng, Yuandong Wang, Chunyang Liu, Ke- hua Sheng, and Bo Zhang. Multispans: a multi-range spatial-temporal transformer network for traffic forecast via structural entropy optimization. InProceedings of the 17th...

    2024

  23. [2007]

    Imputeformer: Low rankness-induced transformers for generalizable spatiotemporal imputation

    [Nieet al., 2024 ] Tong Nie, Guoyang Qin, Wei Ma, Yuewen Mei, and Jian Sun. Imputeformer: Low rankness-induced transformers for generalizable spatiotemporal imputation. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2260– 2271,

    2024

  24. [2009]

    Filling the g ap s: Multivariate time series imputation by graph neural networks.arXiv preprint arXiv:2108.00298,

    [Ciniet al., 2021 ] Andrea Cini, Ivan Marisca, and Cesare Alippi. Filling the g ap s: Multivariate time series imputation by graph neural networks.arXiv preprint arXiv:2108.00298,

    2021 arXiv

  25. [2011]

    Attention is all you need.Ad- vances in Neural Information Processing Systems,

    [Vaswani, 2017] A Vaswani. Attention is all you need.Ad- vances in Neural Information Processing Systems,

    2017

  26. [2013]

    Fine-grained urban flow inference with incomplete data.IEEE Transac- tions on Knowledge and Data Engineering, 35(6):5851– 5864,

    [Liet al., 2022 ] Jiyue Li, Senzhang Wang, Jiaqiang Zhang, Hao Miao, Junbo Zhang, and S Yu Philip. Fine-grained urban flow inference with incomplete data.IEEE Transac- tions on Knowledge and Data Engineering, 35(6):5851– 5864,

    2022

  27. [2016]

    Deep learning on traffic prediction: Methods, analysis, and future directions

    [Yinet al., 2021 ] Xueyan Yin, Genze Wu, Jinze Wei, Yan- ming Shen, Heng Qi, and Baocai Yin. Deep learning on traffic prediction: Methods, analysis, and future directions. IEEE Transactions on Intelligent Transportation Systems, 23(6):4927–4943,

    2021

  28. [2017]

    Generative- free urban flow imputation

    [Wanget al., 2022 ] Senzhang Wang, Jiyue Li, Hao Miao, Junbo Zhang, Junxing Zhu, and Jianxin Wang. Generative- free urban flow imputation. InProceedings of the 31st ACM International Conference on Information & Knowl- edge Management, pages 2028–2037,

    2022

  29. [2018]

    Missing traffic data imputation for artificial intelligence in intelligent trans- portation systems: review of methods, limitations, and challenges.IEEE Access, 11:34080–34093,

    [Chanet al., 2023 ] Robin Kuok Cheong Chan, Joanne Mun- Yee Lim, and Rajendran Parthiban. Missing traffic data imputation for artificial intelligence in intelligent trans- portation systems: review of methods, limitations, and challenges.IEEE Access, 11:34080–34093,

    2023

  30. [2019]

    A nonconvex low-rank tensor completion model for spatiotemporal traffic data imputation.Transportation Re- search Part C: Emerging Technologies, 117:102673,

    [Chenet al., 2020 ] Xinyu Chen, Jinming Yang, and Lijun Sun. A nonconvex low-rank tensor completion model for spatiotemporal traffic data imputation.Transportation Re- search Part C: Emerging Technologies, 117:102673,

    2020

  31. [2020]

    Brits: Bidirectional recurrent im- putation for time series.Advances in neural information processing systems, 31,

    [Caoet al., 2018 ] Wei Cao, Dong Wang, Jian Li, Hao Zhou, Lei Li, and Yitan Li. Brits: Bidirectional recurrent im- putation for time series.Advances in neural information processing systems, 31,

    2018

  32. [2021]

    Traffic flow imputation using parallel data and generative adversarial networks.IEEE Transactions on In- telligent Transportation Systems, 21(4):1624–1630,

    [Chenet al., 2019 ] Yuanyuan Chen, Yisheng Lv, and Fei- Yue Wang. Traffic flow imputation using parallel data and generative adversarial networks.IEEE Transactions on In- telligent Transportation Systems, 21(4):1624–1630,

    2019

  33. [2022]

    Bayesian temporal factorization for multidimensional time series prediction.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4659–4673,

    [Chen and Sun, 2021] Xinyu Chen and Lijun Sun. Bayesian temporal factorization for multidimensional time series prediction.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4659–4673,

    2021

  34. [2023]

    [Chen and Chen, 2022] Yong Chen and Xiqun Michael Chen. A novel reinforced dynamic graph convolutional network model with data imputation for network-wide traffic flow prediction.Transportation Research Part C: Emerging Technologies, 143:103820,

    2022

  35. [2024]

    mice: Multivariate imputation by chained equations in r.Journal of statistical software, 45:1–67,

    [Van Buuren and Groothuis-Oudshoorn, 2011] Stef Van Bu- uren and Karin Groothuis-Oudshoorn. mice: Multivariate imputation by chained equations in r.Journal of statistical software, 45:1–67,

    2011

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.