Pith. sign in

REVIEW 6 major objections 5 minor 35 references

Forecasting Multivariate Urban Data via Decomposition and Spatio-Temporal Graph Analysis

T0 review · 6 major / 5 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 This paper claims that decomposing urban time series into trend, seasonal, and residual parts, each with its own learned graph, improves long-term multivariate forecasts by roughly 3 to 9 percent.

desk verdict The per-component graph learning idea is real, but the headline 2.89–9.10% improvement isn't reproducible from the paper's own table, and the evaluation lacks error bars and a key graph baseline. read the letter →

arxiv 2505.22474 v2 pith:OG22MKQY submitted 2025-05-28 cs.LG

Amirhossein Sohrabbeig , Omid Ardakanian , Petr Musilek This is my paper
classification cs.LG
keywords multivariatetime-seriesforecastingspatio-temporalgraphneuralnetworksattentiontemporalconvolutionalnetworkdecompositiondynamictimewarpinglong-termurbandata
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

This paper argues that long-term multivariate urban forecasting becomes more accurate when the input is first decomposed into trend, seasonal, and residual components and each component is given its own graph structure. The proposed model, DST, learns one directed graph per component from historical data via dynamic time warping and nearest neighbors, then runs graph attention over that graph and temporal convolutions along time. On four real-world datasets covering electricity, weather, carbon intensity, and air pollution, the authors report average error reductions of 2.89% to 9.10% against strong baselines across forecast horizons from 96 to 720 steps. The practical payoff is that sensor dependencies differ across temporal regimes, so modeling them separately should improve long-range forecasts used in urban planning and energy decisions.

What carries the argument

The central mechanism is per-component graph structure learning. A time series is decomposed in an MSTL-inspired way: a weighted moving average with tricube weights extracts the trend, the detrended series is averaged over sliding windows aligned with the seasonality period to extract the seasonal component, and the residual is what remains. Each component is then downsampled, pair-wise dynamic time warping distances are computed between all variates, and each node is connected to its K nearest neighbors to form a directed weighted graph. These static graphs are learned offline and then used as the fixed adjacency for a single GATv2 layer, which lets each sensor assign different attention to its neighbors, while a temporal convolutional network provides causal dilated convolutions over time. The decomposition is what makes the graphs meaningful: trend co-movements and seasonal co-movements can have different structures, so giving each component its own graph is the key idea that carries the forecast improvement.

What would settle it

Take the Electricity dataset, learn the per-component graphs from the first twelve months, then recompute them from the most recent twelve months, and compare DST's MSE at horizon 720 under the frozen early graphs versus the freshly estimated graphs. If the frozen-graph version does not lose, the static-graph assumption is not limiting; if the fresh-graph version improves, the paper's reported accuracy is bounded by how often the graphs must be re-estimated.

Watch

Extended reading notes

Core claim

DST is a multivariate forecasting model that decomposes each input window into trend, seasonal, and residual components, infers a directed graph per component by computing dynamic time warping distances among downsampled variates and keeping the K nearest neighbors, and then forecasts each component with a shared architecture of GATv2 for spatial message passing, a temporal convolutional network for temporal features, a date-time embedding, and a linear projection that produces the full forecast horizon at once. The component forecasts are summed to give the final prediction. The central empirical claim is that this decomposition-with-per-component-graph design beats strong baselines, including s-Mamba, DLinear, Autoformer, PatchTST, zero-shot TimesFM, and Repeat Last, in almost all dataset and horizon settings, with the largest margins at longer horizons and a claimed average improvement of 2.89% to 9.10%. Ablation results show that removing decomposition or temporal feature extraction raises MSE, so both components are load-bearing for the reported accuracy.

Load-bearing premise

The graph of relationships among sensors is fixed for the whole training and forecasting period, so if the true dependencies drift over seasons or across the long forecast horizon, the static per-component graphs become stale and the claimed accuracy gains shrink.

Editorial extensions

If this is right

  • DST reduces mean squared error and mean absolute error at horizons of 96 to 720 steps across electricity, weather, carbon intensity, and air pollution data, so long-range urban forecasts can be made more accurate with this architecture.
  • The gap between DST and the second-best baseline generally widens as the horizon lengthens, indicating the decomposition-plus-graph design helps most where error accumulation is worst.
  • A single GATv2 layer consistently beats deeper stacks in the experiments, so the gain comes from the graph structure and decomposition rather than from a deeper attention network.
  • Ablations show that temporal feature extraction and decomposition are the most influential modules, meaning both are necessary to realize the reported accuracy gains.
  • The linear direct multi-step forecasting head produces the entire horizon in one pass, making DST computationally lightweight compared with transformer-based and foundation-model baselines.

Reading between the lines

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

  • If per-component graphs genuinely encode different dependency structures, the same decomposition trick could be dropped into other graph-based forecasters by replacing a single learned graph with separate trend, seasonal, and residual graphs; this transfer is not tested in the paper.
  • A natural stress test the paper does not run is to recompute the DTW-derived graphs at multiple points over the three-year datasets and measure how much the edge sets drift, which would bound the forecast horizon over which the static-graph assumption is safe.
  • The reported 2.89% to 9.10% improvement is measured on standardized urban benchmarks with frozen graphs; in deployment, the gains will depend on how stationary the sensor dependencies are over time.
  • The date-time embedding contributes most on strongly periodic datasets, so on non-cyclic urban signals the net benefit of DST relative to a plain temporal-convolution-plus-graph model could be smaller.
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

6 major / 5 minor

Summary. The paper proposes DST, a multivariate time-series forecasting model that first decomposes each series into trend, seasonal, and residual components, learns a separate graph for each component via DTW and K-nearest-neighbors from training data, and then applies GATv2 for spatial feature extraction and a TCN for temporal feature extraction, followed by a linear predictor. The model is evaluated on four urban datasets (electricity, weather, carbon intensity, air pollution) across four long-term horizons (96, 192, 336, 720). The paper claims an average improvement of 2.89% to 9.10% over state-of-the-art baselines, with DST achieving the best MSE/MAE in almost all settings. An ablation study examines the contributions of decomposition, date-time embedding, temporal feature extraction, and spatial feature extraction.

Significance. If the quantitative claims are corrected, the paper makes a useful engineering contribution: the idea of learning distinct graph structures for different time-series components is a sensible inductive bias, and the model is relatively simple compared with recent transformer- and state-space-based baselines. The evaluation spans several diverse urban datasets and horizons, which is a strength. However, the central numeric claim is not reproducible from the reported tables under a clear protocol, the results are reported without variance estimates for the five runs, and there is at least one internally inconsistent table entry. These issues currently undermine the paper's headline and need to be fixed before the contribution can be fully assessed.

major comments (6)
  1. [Abstract, Section 5, Table 2] The claimed 'average improvement of 2.89% to 9.10%' is not reproducible under the natural definition of per-cell relative MSE improvement over the best baseline, averaged over the four horizons. Using Table 2, I obtain approximately Air 2.9%, Carbon 11.4%, Electricity 9.1%, and Weather 7.8%; the Carbon figure falls outside the stated range. The paper never defines the aggregation protocol (per-dataset vs per-horizon, which baseline is the reference, whether MSE or MAE is used). Please state the protocol explicitly and recompute the range, then adjust the abstract and conclusion accordingly.
  2. [Table 2, Weather row, horizon 720, PatchTST] The reported MAE of 0.133 is internally inconsistent with the adjacent MSE of 1.378 and with PatchTST's other Weather MAEs (0.256, 0.459, 0.618). This appears to be a typo, likely 0.733. Because this cell affects whether DST wins that setting and contributes to the headline, the correct value must be provided and the surrounding comparisons re-evaluated.
  3. [Section 5, Table 2] Results are averaged over five runs, but no standard deviations or confidence intervals are reported. Figure 6 shows distributions only for one setting (multivariate output, horizon 96). Several winning margins are small—for example, Air pollution at horizon 96 has DST MSE 0.864 versus DLinear 0.872, a relative difference of about 0.9%—so the claim of being "best in almost all settings" is not statistically supported without variance information. Please report run-level statistics, at least for the marginal cells.
  4. [Section 3.2 and Section 5] The static graph assumption is stated in Section 3.2 but never stress-tested. The graphs are learned offline from the training data via DTW/KNN and then held fixed for forecast horizons up to 720 steps. If inter-variate dependencies drift over that horizon, the message-passing edges become stale. Please add an experiment that re-learns graphs on successive training windows and measures the change in accuracy, or otherwise justify that the sensitivity to this assumption is negligible.
  5. [Section 6, Table 3] The reported MSE values in Table 3 do not match Table 2. For example, DST full model on Air pollution is 0.404 in Table 3 but ranges from 0.864 to 0.945 in Table 2. The ablation apparently uses a different experimental setting (likely multivariate output at horizon 96, as in Figure 6), but the caption and text do not state this. Please specify the exact protocol (output mode, horizon, normalization) used in the ablation so the results are interpretable.
  6. [Section 5, Table 2 caption and Section 3.3.3] The evaluation protocol labeled "multivariate input, univariate output" is ambiguous. The model in Eq. (9) produces forecasts for all D channels, and Section 3.3.3 says each variate is predicted in a univariate fashion. Please clarify whether a separate model is trained per target variate, how the reported MSE/MAE are aggregated across target variates, and how the univariate-output setting applies to the Electricity dataset with 321 variates.
minor comments (5)
  1. [Section 3.1, Eq. (2)] Equation (2) sums B_i over the index j, but the summand should be B_j. Please correct this typographical error.
  2. [Throughout] There are several typos: "degredation" in the Introduction, "methdology" in Section 3, and "GetSeasnonalityPeriods" in Algorithm 1. A careful proofreading pass is recommended.
  3. [Section 4.1, Table 1] Table 1 lists 49 variates for the Carbon Intensity dataset, while the text states data are recorded across 50 U.S. states and four territories. Please reconcile these counts and clarify how the 49 variates are selected.
  4. [Section 4.2] The rationale for excluding MOIRAI (probabilistic forecasting) is reasonable, but TimesFM is evaluated only in zero-shot mode; consider also reporting a fine-tuned version or another foundation model to give a stronger comparison from the foundation-model category.
  5. [Figure 6] The y-axis is exaggerated, as acknowledged in the text. This makes visual differences appear larger than they are. Please use a consistent scale or annotate the actual numerical values so the reader can judge the magnitude of the differences.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: DST's forecasts are evaluated on held-out data against external baselines, and the cited prior work is motivational rather than load-bearing.

full rationale

The derivation chain is self-contained against external benchmarks. Decomposition (Section 3.1) and graph learning (Section 3.2) are preprocessing steps applied to the training series; DTW and K-nearest-neighbor graph construction use only historical inputs and are not functions of the forecast targets. The forecasting model (Section 3.3) is trained with MSE on training samples and evaluated on held-out test samples in Table 2, with results averaged over five runs. No parameter is fitted to the test set and then renamed a prediction; the '2.89% to 9.10%' range is a summary of Table 2 comparisons against external baselines. The self-citations to Sohrabbeig et al. [2023] motivate decomposition ('This decomposition is crucial ... Sohrabbeig et al. [2023]') and justify using a linear forecaster ('This choice is motivated by the demonstrated effectiveness of similar methods in previous studies Sohrabbeig et al. [2023], Zeng et al. [2023]'); these are design-choice citations, not the evidence for DST's claimed accuracy gain, which rests on the held-out comparisons and ablations. The apparent internal inconsistency in Table 2 (Weather/720 PatchTST MAE 0.133 vs neighboring entries) and the under-specified averaging protocol for the headline improvement are correctness/reproducibility concerns, not circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method rests on modeling assumptions about static graphs, shared decomposition structure, and DTW proximity as a dependency proxy. No new physical entities are postulated. The main fitted degree of freedom is graph sparsity K, selected per dataset on validation data.

free parameters (3)
  • K (number of nearest neighbors in graph construction) = Grid-searched over 10% to 100% of other nodes; final value not reported
    K determines the graph edges and therefore the spatial information flow. It is tuned on validation data, so the test results are conditional on this choice.
  • Look-back window length l = 336
    Fixed for all experiments and datasets. Accuracy depends on this choice, and it also limits the seasonal components that can be extracted in the sample decomposition.
  • Seasonality period p = Daily (24 steps) for all datasets; weekly cycles not extracted
    The paper states a single daily seasonal component is used because the look-back window is 336 steps, but the detection procedure is not reported. The decomposition and per-component graphs depend entirely on this choice.
assumptions (4)
  • domain assumption The graph structure for each time-series component is static over time.
    Invoked in Section 3 ('These connections are assumed to be static') and Section 3.2 ('graph structure learning is an offline task'). If inter-variate dependencies drift, the fixed graphs become stale.
  • domain assumption The number and type of decomposition components are identical across all variates.
    Section 3.2 states 'the number of these components is assumed to be identical across all variates'. Heterogeneous sensors may have different seasonal structures, which would violate this assumption.
  • domain assumption DTW distance over downsampled component series is a valid proxy for inter-variate dependency for forecasting.
    Section 3.2 constructs graphs from DTW-based K-nearest neighbors. No evidence is given that this proximity measure captures the dependencies that matter for prediction.
  • ad hoc to paper A single daily seasonal component suffices for the 336-step look-back window.
    Section 5: 'we decided to extract only one seasonal component due to the short length of the input'. For horizons up to 720 (multiple weeks), missing weekly cycles could handicap the model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Forecasting Multivariate Urban Data via Decomposition and Spatio-Temporal Graph Analysis." pith.science (2026). https://pith.science/paper/OG22MKQY

@misc{pith2026250522474,
  author       = {Pith},
  title        = {Pith review of: Forecasting Multivariate Urban Data via Decomposition and Spatio-Temporal Graph Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OG22MKQY}},
  note         = {Machine review of arXiv:2505.22474}
}
read the original abstract

Long-term forecasting of multivariate urban data poses a significant challenge due to the complex spatiotemporal dependencies inherent in such datasets. This paper presents DST, a novel multivariate time-series forecasting model that integrates graph attention and temporal convolution within a Graph Neural Network (GNN) to effectively capture spatial and temporal dependencies, respectively. To enhance model performance, we apply a decomposition-based preprocessing step that isolates trend, seasonal, and residual components of the time series, enabling the learning of distinct graph structures for different time-series components. Extensive experiments on real-world urban datasets, including electricity demand, weather metrics, carbon intensity, and air pollution, demonstrate the effectiveness of DST across a range of forecast horizons, from several days to one month. Specifically, our approach achieves an average improvement of 2.89% to 9.10% in long-term forecasting accuracy over state-of-the-art time-series forecasting models.

Figures

Figures reproduced from arXiv: 2505.22474 by the authors.

Figure 1
Figure 1. Extracting seasonal patterns using a sliding window of size [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. DTW distance heatmaps of (a) trend and (b) seasonal components of the air pollution dataset (Section 4.1); [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The overall architecture of proposed forecasting model. After decomposition, each component is handled [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Decomposition of an electricity consumption sample over two weeks. The first subplot displays the original [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: One layer of spatio-temporal feature extraction with time embedding. The input to this module is readings [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: MSE comparison of our proposed model against baselines on different datasets. This is a multivariate [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Comparison between the forecasts of DST, and two best performing baselines: s-Mamba and DLinear, on [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 20 canonical work pages

  1. [1]

    Electricity load forecasting: a systematic review

    Isaac Kofi Nti, Moses Teimeh, Owusu Nyarko-Boateng, and Adebayo Felix Adekoya. Electricity load forecasting: a systematic review. Journal of Electrical Systems and Information Technology, 7: 0 1--19, 2020

    work page 2020

  2. [2]

    Guoyan Huang, Xinyi Li, Bing Zhang, and Jiadong Ren. Pm2. 5 concentration forecasting at surface monitoring sites using gru neural network based on empirical mode decomposition. Science of the Total Environment, 768: 0 144516, 2021

    work page 2021

  3. [3]

    Graphy: Graph-based physics-guided urban air quality modeling for monitoring-constrained regions

    Shangjie Du, ZhiZhang Hu, and Shijia Pan. Graphy: Graph-based physics-guided urban air quality modeling for monitoring-constrained regions. In Proceedings of the 11th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation, pages 33--43, 2024

    work page 2024

  4. [4]

    Subseasonalclimateusa: a dataset for subseasonal forecasting and benchmarking

    Soukayna Mouatadid, Paulo Orenstein, Genevieve Flaspohler, Miruna Oprescu, Judah Cohen, Franklyn Wang, Sean Knight, Maria Geogdzhayeva, Sam Levang, Ernest Fraenkel, et al. Subseasonalclimateusa: a dataset for subseasonal forecasting and benchmarking. Advances in Neural Information Processing Systems, 36, 2024

    work page 2024

  5. [5]

    Carboncast: multi-day forecasting of grid carbon intensity

    Diptyaroop Maji, Prashant Shenoy, and Ramesh K Sitaraman. Carboncast: multi-day forecasting of grid carbon intensity. In Proceedings of the 9th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation, pages 198--207, 2022

    2022

  6. [6]

    Connecting the dots: Multivariate time series forecasting with graph neural networks

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, Xiaojun Chang, and Chengqi Zhang. Connecting the dots: Multivariate time series forecasting with graph neural networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 753--763, 2020

    work page 2020

  7. [7]

    Graph attention networks

    Petar Veli c kovi \'c , Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

    arXiv 2017

  8. [8]

    How attentive are graph attention networks? arXiv preprint arXiv:2105.14491, 2021

    Shaked Brody, Uri Alon, and Eran Yahav. How attentive are graph attention networks? arXiv preprint arXiv:2105.14491, 2021

    arXiv 2021

Show all 35 references
  1. [9]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling

    Shaojie Bai, J Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271, 2018

    2018 arXiv

  2. [10]

    Carbon-aware computing for datacenters

    Ana Radovanović et al. Carbon-aware computing for datacenters. IEEE Transactions on Power Systems, 38 0 (2): 0 1270--1280, 2023

    2023

  3. [11]

    Some recent advances in forecasting and control

    George EP Box and Gwilym M Jenkins. Some recent advances in forecasting and control. Journal of the Royal Statistical Society. Series C (Applied Statistics), 17 0 (2): 0 91--109, 1968

    1968

  4. [12]

    Distribution of residual autocorrelations in autoregressive-integrated moving average time series models

    George EP Box and David A Pierce. Distribution of residual autocorrelations in autoregressive-integrated moving average time series models. Journal of the American statistical Association, 65 0 (332): 0 1509--1526, 1970

    1970

  5. [13]

    Forecasting seasonals and trends by exponentially weighted moving averages

    Charles C Holt. Forecasting seasonals and trends by exponentially weighted moving averages. International journal of forecasting, 20 0 (1): 0 5--10, 2004

    2004

  6. [14]

    Forecasting sales by exponentially weighted moving averages

    Peter R Winters. Forecasting sales by exponentially weighted moving averages. Management science, 6 0 (3): 0 324--342, 1960

    1960

  7. [15]

    A training algorithm for optimal margin classifiers

    Bernhard E Boser, Isabelle M Guyon, and Vladimir N Vapnik. A training algorithm for optimal margin classifiers. In Proceedings of the fifth annual workshop on Computational learning theory, pages 144--152, 1992

    1992

  8. [16]

    Random forests

    Leo Breiman. Random forests. Machine learning, 45: 0 5--32, 2001

    2001

  9. [17]

    Greedy function approximation: a gradient boosting machine

    Jerome H Friedman. Greedy function approximation: a gradient boosting machine. Annals of statistics, pages 1189--1232, 2001

    2001

  10. [18]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785--794, 2016

    2016

  11. [19]

    Long short-term memory

    Sepp Hochreiter and J \"u rgen Schmidhuber. Long short-term memory. Neural computation, 9 0 (8): 0 1735--1780, 1997

    1997

  12. [20]

    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. In NIPS 2014 Workshop on Deep Learning, December 2014, 2014

    2014

  13. [21]

    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. arXiv preprint arXiv:1905.10437, 2019

    1905 arXiv

  14. [22]

    Informer: Beyond efficient transformer for long sequence time-series forecasting

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 11106--11115, 2021

    2021

  15. [23]

    Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. Advances in Neural Information Processing Systems, 34: 0 22419--22430, 2021

    2021

  16. [24]

    Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. In International Conference on Machine Learning, pages 27268--27286. PMLR, 2022

    2022

  17. [25]

    Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121--11128, 2023

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121--11128, 2023

    2023

  18. [26]

    Cyclenet: Enhancing time series forecasting through modeling periodic patterns

    Shengsheng Lin, Weiwei Lin, Xinyi HU, Wentai Wu, Ruichao Mo, and Haocheng Zhong. Cyclenet: Enhancing time series forecasting through modeling periodic patterns. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/fo...

    2024

  19. [27]

    Multivariate time series forecasting with dynamic graph neural odes

    Ming Jin, Yu Zheng, Yuan-Fang Li, Siheng Chen, Bin Yang, and Shirui Pan. Multivariate time series forecasting with dynamic graph neural odes. IEEE Transactions on Knowledge and Data Engineering, 35 0 (9): 0 9168--9180, 2022

    2022

  20. [28]

    Adaptive spatio-temporal graph convolutional neural network for remaining useful life estimation

    Yuxuan Zhang, Yuanxiang Li, Xian Wei, and Lei Jia. Adaptive spatio-temporal graph convolutional neural network for remaining useful life estimation. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1--7, 2020. doi:10.1109/IJCNN48605.2020.9206739

    2020

  21. [29]

    Spatial-temporal fusion graph neural networks for traffic flow forecasting, 2021

    Mengzhang Li and Zhanxing Zhu. Spatial-temporal fusion graph neural networks for traffic flow forecasting, 2021. URL https://arxiv.org/abs/2012.09641

    2021 arXiv

  22. [30]

    Spatio-temporal predictive modeling techniques for different domains: a survey

    Rahul Kumar, Manish Bhanu, Jo\ a o Mendes-Moreira, and Joydeep Chandra. Spatio-temporal predictive modeling techniques for different domains: a survey. ACM Comput. Surv., 57 0 (2), October 2024. ISSN 0360-0300. doi:10.1145/3696661. URL https://doi.org/10.1145/3696661

    2024 doi

  23. [31]

    A decoder-only foundation model for time-series forecasting

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting. arXiv preprint arXiv:2310.10688, 2023

    2023 arXiv

  24. [32]

    Unified training of universal time series forecasting transformers

    Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. Unified training of universal time series forecasting transformers. In Proceedings of the 41st International Conference on Machine Learning, ICML'24. JMLR.org, 2024

    2024

  25. [33]

    Decompose and conquer: Time series forecasting with multiseasonal trend decomposition using loess

    Amirhossein Sohrabbeig, Omid Ardakanian, and Petr Musilek. Decompose and conquer: Time series forecasting with multiseasonal trend decomposition using loess. Forecasting, 5 0 (4): 0 684--696, 2023. ISSN 2571-9394. doi:10.3390/forecast5040037. URL https://www.mdpi.com/2571-9394/5/4/37

    2023 doi

  26. [34]

    Dynamic time warping

    Meinard M \"u ller. Dynamic time warping. Information retrieval for music and motion, pages 69--84, 2007

    2007

  27. [35]

    Assessing beijing's pm2.5 pollution: severity, weather impact, apec and winter heating

    Xuan Liang, Tao Zou, Bin Guo, Shuo Li, Haozhe Zhang, Shuyi Zhang, Hui Huang, and Song Xi Chen. Assessing beijing's pm2.5 pollution: severity, weather impact, apec and winter heating. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 471, 2015...

    2015

Pith tools

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