Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Dynamic Modes as Time Representation for Spatiotemporal Forecasting

T0 review · 5 major / 6 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 The paper claims that time embeddings built from the dominant oscillation modes of the observed data—cosine and sine components extracted by Dynamic Mode Decomposition—improve long-horizon spatiotemporal forecasts and reduce residual…

desk verdict DMD-derived sinusoidal time embeddings are a plausible plug-in for spatiotemporal forecasting, but the paper must clarify whether DMD is fit on the training split and reconcile Eq. 13 with Algorithm 1 before the reported gains can be trusted. read the letter →

arxiv 2506.01212 v2 pith:FJF56D3C submitted 2025-06-01 cs.LG stat.ML

Menglin Kong , Vincent Zhihao Zheng , Xudong Wang , Lijun Sun This is my paper
classification cs.LGstat.ML
keywords dynamicmodedecompositionspatiotemporalforecastingtimeembeddingKoopmanoperatortheoryseasonalityspectralanalysislong-horizonresidualautocorrelation
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 the periodic structure a forecasting model needs can be extracted from the data itself, rather than supplied as timestamps or hand-crafted calendar features. It claims that Dynamic Mode Decomposition (DMD) yields a compact set of oscillation frequencies, and that feeding their cosine and sine components to any spatiotemporal forecasting model as extra input channels consistently improves long-horizon accuracy on urban metro ridership, highway traffic, and daily climate data. The reported gains are largest at the longest forecast horizon (12 steps) on the most strongly periodic dataset, and the residual autocorrelation at daily and weekly lags shrinks, meaning the model has internalized the seasonality instead of leaving it in the error term. A sympathetic reader would care because this is a lightweight, model-agnostic replacement for time features that any architecture accepting covariates could adopt.

What carries the argument

The load-bearing object is the DMD time covariate, a $2r$-dimensional vector of the cosine and sine components of the $r$ dominant oscillation frequencies of the observed signal. It is produced by a four-stage chain: a circulant Hankel embedding of the raw signal to lift the spatial dimension from $N$ to $N\tau$; Total DMD (TDMD), a noise-aware de-biased spectral decomposition; Sparsity-Promoting DMD (SPDMD), which prunes to the dominant modes; and finally the Vandermonde temporal dynamics matrix $\bar{C}$, whose eigenvalue arguments $\omega_k = \arg(\lambda_k)$ are read off as frequencies. The construction is what carries the argument: the frequencies are the only information passed downstream, the amplitude terms $e^{\mu_k t}$ are dropped for numerical stability and interpretability, and the resulting covariates concatenate into any model input of shape $(N, T, m)$, widening it to $(N, T, m + 2r)$.

What would settle it

Recompute the DMD modes using only data strictly before each test block, on a dataset whose dominant seasonality shifts over time (for example metro ridership spanning a holiday or schedule change), and measure the same 3-, 6-, and 12-step MAE/RMSE deltas; if the train-only embedding shows no gain or a smaller gain than reported, the temporal-generalization claim fails. A cheaper check is to inspect the residual autocorrelation at lags 72 and 504 after refitting the modes per training fold: if the periodic ACF peaks return, the embedding has not internalized the seasonality it is credited with.

Watch

Extended reading notes

Core claim

The paper's central claim is that a time embedding built from the dominant DMD modes captures multi-scale periodicity that hand-crafted features (time of day, day of week) and learnable embeddings (Time2Vec) miss. Concretely, the embedding is $c_t^{(\mathrm{DMD})} = [\cos(\omega_1 t), \ldots, \cos(\omega_r t), \sin(\omega_1 t), \ldots, \sin(\omega_r t)] \in \mathbb{R}^{2r}$, where each $\omega_k = \arg(\lambda_k)$ is the frequency of a dominant eigenvalue of a reduced Koopman operator fitted to a Hankel-lifted version of the observations; the growth and decay amplitudes are deliberately discarded so the covariate is stationary. Appending these $2r$ channels to FC-LSTM, DCRNN, AGCRN, and Graph WaveNet lowers MAE and RMSE on GZ-METRO, PEMS04, and Daymet across 3-, 6-, and 12-step horizons, with the clearest wins in the 12-step metro case (Graph WaveNet RMSE 93.62 to 87.79; DCRNN 105.48 to 91.98). The residual analysis backs the mechanism: autocorrelation peaks at lags 72 and 144 (daily and bi-daily) nearly disappear, and the learned Time2Vec frequencies are found concentrated at 5–15 steps while DMD resolves daily and weekly modes.

Load-bearing premise

The load-bearing premise is that the oscillation frequencies estimated from the historical record are the right frequencies for the test and future periods, and the paper does not demonstrate this under shifting seasonality; it also never states explicitly that the modes were computed from the training split alone, so part of the reported gain could come from the embedding having seen test-period data.

Editorial extensions

If this is right

  • Any forecasting architecture that accepts time covariates can adopt the method by concatenating $2r$ extra channels, so the reported accuracy gains transfer without architectural change.
  • The gains concentrate at the longest horizon (12 steps) on strongly periodic data, making the method most valuable precisely where hand-crafted features fail.
  • Residual autocorrelation at daily and weekly lags drops, so models trained with the embedding produce errors closer to temporally independent noise, which should improve the trustworthiness of interval estimates.
  • Because no timestamps or calendar semantics are needed, the embedding applies to systems whose periods are not calendar-aligned or whose metadata is missing.
  • On noisy, weakly periodic data the benefit is explicitly smaller, so the method is a tool for periodic systems rather than a general-purpose fix.

Reading between the lines

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

  • The paper does not state whether the DMD modes are estimated on the training split only; if the full series is used, part of the improvement could be leakage of the test period's periodic structure into the covariates, and a train-only re-estimation is the check that would settle it.
  • Since the amplitude terms $e^{\mu_k t}$ are discarded, the embedding presumes stationary periodicity; re-estimating modes on a sliding window and tracking how $\omega_k$ drifts is the natural extension toward the non-stationary settings the paper names as future work.
  • The spectral covariates could double as a diagnostic: comparing DMD-extracted periods with the residual ACF of any forecaster would reveal whether that model is leaving seasonal structure on the table.
  • The method's value should scale with forecast horizon and with the variance share of the dominant modes (the paper reports over 90% of variance in a few singular components); a testable prediction is that gains grow with $Q/P$ and with that low-rank share.
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

5 major / 6 minor

Summary. The paper proposes a data-driven time embedding for spatiotemporal forecasting based on Dynamic Mode Decomposition (DMD). The pipeline builds a circular Hankel matrix from the observed multivariate signal, applies DMD-style spectral analysis, and uses the arguments of the dominant eigenvalues to define sinusoidal time covariates. These covariates are concatenated with model inputs and evaluated on GZ-METRO, PEMS04, and Daymet with FC-LSTM, DCRNN, AGCRN, and Graph WaveNet as backbones. The paper reports improved MAE/RMSE at 3-, 6-, and 12-step horizons relative to no time embedding and to hand-crafted and Time2Vec embeddings, together with reduced residual autocorrelation.

Significance. If the reported gains are valid, the method is a useful, lightweight, model-agnostic contribution: it replaces hand-crafted calendar features with a compact spectral representation learned from data, and it is compatible with any architecture that accepts time covariates. The paper evaluates across three datasets, four backbone models, and multiple horizons, and it includes a qualitative residual-ACF analysis that directly targets the claimed mechanism. These strengths are meaningful. However, the central claim currently rests on an incompletely specified and potentially non-causal estimation procedure, so the quantitative results do not yet establish the method's temporal-generalization advantage.

major comments (5)
  1. [Algorithm 1 / Experiments] Algorithm 1 takes the full raw signal {z_t}_{t=0}^{T-1} and Eq. (8) builds the Hankel matrix from the whole series, while Section 'Experiments' only states that z-score statistics are computed on the training set. It is therefore unclear whether the DMD eigenfrequencies are estimated on the training split only or on the full dataset including validation and test periods. If the latter, the test-time covariates cos(omega_k t), sin(omega_k t) encode information from the test period, so the claimed 'temporal generalization' and the comparisons against Time2Vec and hand-crafted features in Tables 3 and 4 would be inflated by transductive leakage. The authors must specify the DMD fitting split, modify Algorithm 1 and Eq. (8) accordingly, and, if the full series was used, rerun all experiments with training-only frequency estimation.
  2. [Methodology, Eq. (13) vs. Remark/Algorithm 1] The embedding is defined in Eq. (13) as c_t^(DMD) = [Re(C[:,t]); Im(C[:,t])], where C is the Vandermonde matrix of DMD eigenvalues lambda_k, but the Remark and Algorithm 1 instead use c_t = [cos(omega_1 t), ..., sin(omega_r t)] with omega_k = arg(lambda_k), discarding the amplitude/growth factor e^{mu_k t}. These two definitions coincide only if all |lambda_k| = 1; otherwise they produce different covariates. Since the experimental section does not state which definition was evaluated, the reported numbers cannot be unambiguously attributed to the method as described. Please reconcile Eq. (13), the Remark, and Algorithm 1, and report eigenvalue magnitudes or explicitly justify discarding them.
  3. [Eq. (8)] The Hankel matrix in Eq. (8) is circulant: column j is [z_j; z_{j+1}; ...; z_{j+tau-1}] with indices taken modulo T, so the final columns wrap around to the beginning of the series and the snapshot at t = T-1 contains z_0, ..., z_{tau-2}. This imposes a periodic boundary condition z_T = z_0 that is generally false for real spatiotemporal series and can bias the estimated DMD eigenvalues. If the circular construction is intentional, the authors should justify it and explain why it does not distort the extracted frequencies; otherwise the Hankel matrix should be formed from a longer trajectory without wrap-around.
  4. [Methodology / Algorithm 1] The methodology announces a three-stage pipeline of Hankel embedding, Total DMD (TDMD), and Sparsity-Promoting DMD (SPDMD), but Algorithm 1 and Eqs. (9)-(12) implement only standard exact DMD via SVD of H^T H. No equations or algorithmic details are given for the TDMD debiasing step or for how SPDMD selects the r dominant modes, and the choice of r is not discussed beyond aligning embedding dimensions in Table 4. Please provide the full pipeline actually used in the experiments, or state explicitly if TDMD and SPDMD were omitted, and describe how tau and r are selected.
  5. [Experiments / Table 3] All results are reported as averages of three runs with no standard deviations, confidence intervals, or significance tests. In Table 3 several entries show small degradations, e.g., FC-LSTM on Daymet at 3-step (MAE 3.89 vs. 3.90, RMSE 5.11 vs. 5.13) and AGCRN on PEMS04 at 3-step RMSE (29.83 vs. 29.88), which contradicts the abstract's statement that the method 'consistently improves' forecasting accuracy. Please report run-level variation and either provide statistical support for the improvement claim or temper the wording.
minor comments (6)
  1. [Algorithm 1 / Inference] Algorithm 1 returns covariates only for t = 0, ..., T-1; please explain how the fitted frequencies omega_k are used to generate covariates for future time steps t > T-1 at inference time, given that Eq. (1) requires covariates up to t+Q.
  2. [Eq. (1)] Equation (1) includes future covariates C_{t-P+1:t+Q}; please clarify whether the DMD covariates for future timestamps are available at inference time and whether the same global frequencies are used across all timestamps.
  3. [Table 4] In Table 4, the Daymet panel reports only Time2Vec and Ours, whereas the GZ-METRO and PEMS04 panels also include the D and DW baselines; please explain the omission or add the missing comparisons for completeness.
  4. [Related Work / Experiments] The related work discusses Transformers and TCNs, but the experiments cover only LSTM- and GCN-based backbones; please note explicitly that compatibility with Transformer-style architectures is not empirically tested.
  5. [Figure 4] The residual correlation matrices in Figures 4a and 4b lack a colorbar or numeric scale, which makes the reported reductions difficult to interpret; please add a scale or report the exact values in the caption.
  6. [Qualitative Study] The sentence 'We further inspected the learned frequencies of Time2Vec' refers to an analysis that is not shown; please include the corresponding figure or quantitative summary, or remove the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DMD embedding is a feature-engineering step and the central forecasting claim is evaluated on held-out periods; the unresolved train/test split for DMD fitting is a leakage risk, not a circular derivation.

full rationale

The paper's derivation chain is: apply DMD to observed spatiotemporal signals to extract oscillatory frequencies, construct cosine/sine time covariates from those frequencies, concatenate them with existing inputs, and train a forecasting model to predict future observations. The central claim is empirical: the resulting embeddings improve long-horizon accuracy on held-out test periods. This is not a case where the prediction reduces to the fitted input by construction. The DMD step produces auxiliary covariates, not the forecast itself, and the forecasting models (FC-LSTM, DCRNN, AGCRN, Graph WaveNet) must still learn the mapping from covariates to future values. No equation in the paper equates the forecast target with the DMD embedding or with the fitted eigenvalues. The self-citations (Wang and Sun 2022, 2023) concern Hankel embedding and DMD variants used as implementation details; the load-bearing spectral machinery is cited from external work (Schmid 2022; Hemati et al. 2017; Jovanovic et al. 2014). The method is conceptually similar to Fourier features with data-driven frequencies, which is a feature-engineering choice rather than a renaming of a known result. The paper does not explicitly state whether the DMD frequencies are estimated only on the training split; Algorithm 1 takes the full signal and the experimental section only specifies z-score normalization on the training set. That creates a potential transductive leakage concern for the temporal-generalization claim, and it should be addressed, but leakage is a validity/correctness issue rather than circularity: even with full-data frequencies, the reported errors are still produced by models trained on held-out targets, so the claim does not reduce to the fitted input by definition. Overall, the derivation is self-contained with respect to circularity and the score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a small set of standard DMD assumptions and two free hyperparameters (tau and r) that are not reported. No new physical or conceptual entities are introduced. The main unstated assumption is that the spectral structure is stationary across the train/test split.

free parameters (2)
  • Hankel window tau = not reported
    The number of delayed snapshots in the Hankel matrix is a user-chosen hyperparameter that affects the quality of DMD modes. The paper does not report its value for any dataset.
  • Mode count r = not reported
    The number of DMD modes retained after SPDMD determines the embedding dimension. The paper mentions aligning r with the dimension of hand-crafted baselines, but the actual per-dataset values are not given.
assumptions (3)
  • domain assumption Koopman linearization assumption: the nonlinear dynamics can be approximated by a finite-dimensional linear operator on observable space.
    Invoked in Eq. 3-6 of the Methodology. This is a standard approximation for DMD, but it assumes the data is well represented by a low-dimensional linear evolution.
  • domain assumption Stationarity of the periodic structure: the dominant frequencies estimated from training data remain unchanged in the test period.
    The DMD embedding is computed once and then applied to all time steps. If the periodicity shifts over time, the embedding becomes invalid. The paper does not test this.
  • domain assumption Low-rank spatiotemporal structure: the data variance is concentrated in a few singular components.
    The paper justifies DMD by showing a cumulative eigenvalue percentage plot for GZ-METRO, but does not verify this for PEMS04 or Daymet.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamic Modes as Time Representation for Spatiotemporal Forecasting." pith.science (2026). https://pith.science/paper/FJF56D3C

@misc{pith2026250601212,
  author       = {Pith},
  title        = {Pith review of: Dynamic Modes as Time Representation for Spatiotemporal Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FJF56D3C}},
  note         = {Machine review of arXiv:2506.01212}
}
read the original abstract

This paper introduces a data-driven time embedding method for modeling long-range seasonal dependencies in spatiotemporal forecasting tasks. The proposed approach employs Dynamic Mode Decomposition (DMD) to extract temporal modes directly from observed data, eliminating the need for explicit timestamps or hand-crafted time features. These temporal modes serve as time representations that can be seamlessly integrated into deep spatiotemporal forecasting models. Unlike conventional embeddings such as time-of-day indicators or sinusoidal functions, our method captures complex multi-scale periodicity through spectral analysis of spatiotemporal data. Extensive experiments on urban mobility, highway traffic, and climate datasets demonstrate that the DMD-based embedding consistently improves long-horizon forecasting accuracy, reduces residual correlation, and enhances temporal generalization. The method is lightweight, model-agnostic, and compatible with any architecture that incorporates time covariates.

Figures

Figures reproduced from arXiv: 2506.01212 by the authors.

Figure 1
Figure 1. Proposed pipeline. Raw multivariate signals are first converted into a stacked Hankel matrix and then decomposed via [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Koopman operator intuition. (A) A nonlinear spi [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Left: Normalized ridership at all stations of the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: (a)–(b) Residual correlation matrices of GZ-METRO at lags S = 0, 72, 504, without and with our DMD-based embedding. (c)–(d) Autocorrelation functions (ACF) of 12-step prediction residuals from three sensors, showing how our embedding reduces periodic error peaks. spect…

Discussion (0). Sign in 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. Frequency-Constrained Learning for Long-Term Forecasting

    cs.LG 2025-08 reject novelty 3.0 of 10

    Initializing sinusoidal time embeddings with FFT-extracted dominant frequencies and constraining their learning rate yields mixed improvements on traffic forecasting benchmarks, contradicting the paper's claim of cons...

Reference graph

Works this paper leans on

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

  1. [1]

    Data-driven analysis and forecasting of highway traffic dynamics

    Allan M Avila and I Mezi \'c . Data-driven analysis and forecasting of highway traffic dynamics. Nature communications , 11(1):1--16, 2020

    work page 2020

  2. [2]

    Adaptive graph convolutional recurrent network for traffic forecasting

    Lei Bai, Lina Yao, Can Li, Xianzhi Wang, and Can Wang. Adaptive graph convolutional recurrent network for traffic forecasting. Advances in Neural Information Processing Systems , 33:17804--17815, 2020

    work page 2020

  3. [3]

    Pattern recognition and machine learning , volume 4

    Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning , volume 4. Springer, 2006

    2006

  4. [4]

    Time series analysis: forecasting and control

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

    work page 2015

  5. [5]

    Data-driven science and engineering: Machine learning, dynamical systems, and control

    Steven L Brunton and J Nathan Kutz. Data-driven science and engineering: Machine learning, dynamical systems, and control . Cambridge University Press, 2022

    2022

  6. [6]

    Extracting spatial--temporal coherent patterns in large-scale neural recordings using dynamic mode decomposition

    Bingni W Brunton, Lise A Johnson, Jeffrey G Ojemann, and J Nathan Kutz. Extracting spatial--temporal coherent patterns in large-scale neural recordings using dynamic mode decomposition. Journal of neuroscience methods , 258:1--15, 2016

    work page 2016

  7. [7]

    Discovering governing equations from data by sparse identification of nonlinear dynamical systems

    Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the national academy of sciences , 113(15):3932--3937, 2016

    work page 2016

  8. [8]

    Traffic transformer: Capturing the continuity and periodicity of time series for traffic forecasting

    Ling Cai, Krzysztof Janowicz, Gengchen Mai, Bo Yan, and Rui Zhu. Traffic transformer: Capturing the continuity and periodicity of time series for traffic forecasting. Transactions in GIS , 24(3):736--755, 2020

    work page 2020

Show all 46 references
  1. [9]

    Data-driven discovery of coordinates and governing equations

    Kathleen Champion, Bethany Lusch, J Nathan Kutz, and Steven L Brunton. Data-driven discovery of coordinates and governing equations. Proceedings of the National Academy of Sciences , 116(45):22445--22451, 2019

    2019

  2. [10]

    Discovery of nonlinear multiscale systems: Sampling strategies and embeddings

    Kathleen P Champion, Steven L Brunton, and J Nathan Kutz. Discovery of nonlinear multiscale systems: Sampling strategies and embeddings. SIAM Journal on Applied Dynamical Systems , 18(1):312--333, 2019

    2019

  3. [11]

    Freeway performance measurement system: mining loop detector data

    Chao Chen, Karl Petty, Alexander Skabardonis, Pravin Varaiya, and Zhanfeng Jia. Freeway performance measurement system: mining loop detector data. Transportation Research Record , 1748(1):96--102, 2001

    2001

  4. [12]

    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 , 2014

    2014

  5. [13]

    Time series analysis: with applications in R , volume 2

    Jonathan D Cryer and Kung-Sik Chan. Time series analysis: with applications in R , volume 2. Springer, 2008

    2008

  6. [14]

    Convolutional neural networks on graphs with fast localized spectral filtering

    Micha \"e l Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems , 29, 2016

    2016

  7. [15]

    Long-range transformers for dynamic spatiotemporal forecasting

    Jake Grigsby, Zhe Wang, and Yanjun Qi. Long-range transformers for dynamic spatiotemporal forecasting. arXiv preprint arXiv:2109.12218 , 2021

    2021 arXiv

  8. [16]

    Attention based spatial-temporal graph convolutional networks for traffic flow forecasting

    Shengnan Guo, Youfang Lin, Ning Feng, Chao Song, and Huaiyu Wan. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. In Proceedings of the AAAI conference on artificial intelligence , volume 33, pages 922--929, 2019

    2019

  9. [17]

    De-biasing the dynamic mode decomposition for applied koopman spectral analysis of noisy datasets

    Maziar S Hemati, Clarence W Rowley, Eric A Deem, and Louis N Cattafesta. De-biasing the dynamic mode decomposition for applied koopman spectral analysis of noisy datasets. Theoretical and Computational Fluid Dynamics , 31(4):349--368, 2017

    2017

  10. [18]

    Long Short-Term Memory

    Sepp Hochreiter and Jrgen Schmidhuber. Long Short-Term Memory . Neural Computation , 9(8):1735--1780, 1997

    1997

  11. [19]

    Probabilistic energy forecasting: Global energy forecasting competition 2014 and beyond

    Tao Hong, Pierre Pinson, and Shu Fan. Probabilistic energy forecasting: Global energy forecasting competition 2014 and beyond. International Journal of Forecasting , 32(3):896--913, 2016

    2014

  12. [20]

    Sparsity-promoting dynamic mode decomposition

    Mihailo R Jovanovi \'c , Peter J Schmid, and Joseph W Nichols. Sparsity-promoting dynamic mode decomposition. Physics of Fluids , 26(2):024103, 2014

    2014

  13. [21]

    Time-delay observables for koopman: Theory and applications

    Mason Kamb, Eurika Kaiser, Steven L Brunton, and J Nathan Kutz. Time-delay observables for koopman: Theory and applications. SIAM Journal on Applied Dynamical Systems , 19(2):886--917, 2020

    2020

  14. [22]

    Statistical methods versus neural networks in transportation research: Differences, similarities and some insights

    Matthew G Karlaftis and Eleni I Vlahogianni. Statistical methods versus neural networks in transportation research: Differences, similarities and some insights. Transportation Research Part C: Emerging Technologies , 19(3):387--399, 2011

    2011

  15. [23]

    Time2vec: Learning a vector representation of time

    Seyed Mehran Kazemi, Rishab Goel, Sepehr Eghbali, Janahan Ramanan, Jaspreet Sahota, Sanjay Thakur, Stella Wu, Cathal Smyth, Pascal Poupart, and Marcus Brubaker. Time2vec: Learning a vector representation of time. arXiv preprint arXiv:1907.05321 , 2019

    1907 arXiv

  16. [24]

    Semi-supervised classification with graph convolutional networks

    Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International conference on learning representations , 2016

    2016

  17. [25]

    Deep learning

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature , 521(7553):436--444, 2015

    2015

  18. [26]

    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. In International conference on learning representations , 2018

    2018

  19. [27]

    Learnable fourier features for multi-dimensional spatial positional encoding

    Yang Li, Si Si, Gang Li, Cho-Jui Hsieh, and Samy Bengio. Learnable fourier features for multi-dimensional spatial positional encoding. Advances in Neural Information Processing Systems , 34:15816--15829, 2021

    2021

  20. [28]

    N-beats: Neural basis expansion analysis for interpretable time series forecasting

    Boris N Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio. N-beats: Neural basis expansion analysis for interpretable time series forecasting. In International conference on learning representations , 2020

    2020

  21. [29]

    Fc-gaga: Fully connected gated graph architecture for spatio-temporal traffic forecasting

    Boris N Oreshkin, Arezou Amini, Lucy Coyle, and Mark Coates. Fc-gaga: Fully connected gated graph architecture for spatio-temporal traffic forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 35, pages 9233--9241, 2021

    2021

  22. [30]

    Satellite remote sensing for applied ecologists: opportunities and challenges

    Nathalie Pettorelli, William F Laurance, Timothy G O'Brien, Martin Wegmann, Harini Nagendra, and Woody Turner. Satellite remote sensing for applied ecologists: opportunities and challenges. Journal of Applied Ecology , 51(4):839--848, 2014

    2014

  23. [31]

    Deep learning and process understanding for data-driven earth system science

    Markus Reichstein, Gustau Camps-Valls, Bjorn Stevens, Martin Jung, Joachim Denzler, Nuno Carvalhais, and Prabhat. Deep learning and process understanding for data-driven earth system science. Nature , 566(7743):195--204, 2019

    2019

  24. [32]

    Dynamic mode decomposition and its variants

    Peter J Schmid. Dynamic mode decomposition and its variants. Annual Review of Fluid Mechanics , 54:225--254, 2022

    2022

  25. [33]

    Spatial-temporal synchronous graph convolutional networks: A new framework for spatial-temporal network data forecasting

    Chao Song, Youfang Lin, Shengnan Guo, and Huaiyu Wan. Spatial-temporal synchronous graph convolutional networks: A new framework for spatial-temporal network data forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 34, pages 914--921, 2020

    2020

  26. [34]

    Sequence to sequence learning with neural networks

    Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. Advances in neural information processing systems , 27, 2014

    2014

  27. [35]

    Fourier features let networks learn high frequency functions in low dimensional domains

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Proc...

    2020

  28. [36]

    Daymet: Monthly climate summaries on a 1-km grid for north america, version 4 r1

    MM Thornton, R Shrestha, Y Wei, PE Thornton, S Kao, and BE Wilson. Daymet: Monthly climate summaries on a 1-km grid for north america, version 4 r1. ornl daac, oak ridge, tennessee, usa, 2022

    2022

  29. [37]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems , 30, 2017

    2017

  30. [38]

    Short-term traffic forecasting: Where we are and where we’re going

    Eleni I Vlahogianni, Matthew G Karlaftis, and John C Golias. Short-term traffic forecasting: Where we are and where we’re going. Transportation Research Part C: Emerging Technologies , 43:3--19, 2014

    2014

  31. [39]

    Extracting dynamic mobility patterns by hankel dynamic modes decomposition

    Xudong Wang and Lijun Sun. Extracting dynamic mobility patterns by hankel dynamic modes decomposition. In The 11th Triennial Symposium on Transportation Analysis, Mauritius Island , 2022

    2022

  32. [40]

    Anti-circulant dynamic mode decomposition with sparsity-promoting for highway traffic dynamics analysis

    Xudong Wang and Lijun Sun. Anti-circulant dynamic mode decomposition with sparsity-promoting for highway traffic dynamics analysis. Transportation Research Part C: Emerging Technologies , 153:104178, 2023

    2023

  33. [41]

    Graph wavenet for deep spatial-temporal graph modeling

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. Graph wavenet for deep spatial-temporal graph modeling. page 1907–1913, 2019

    1907

  34. [42]

    Multi-scale context aggregation by dilated convolutions

    Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. In International conference on learning representations , 2016

    2016

  35. [43]

    Temporal regularized matrix factorization for high-dimensional time series prediction

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

    2016

  36. [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. In Proceedings of the 27th International Joint Conference on Artificial Intelligence , page 3634–3640, 2018

    2018

  37. [45]

    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

  38. [46]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence '...

Pith tools

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