Pith. sign in

REVIEW 4 major objections 6 minor 32 references

ADFormer: Aggregation Differential Transformer for Passenger Demand Forecasting

T0 review · 4 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 This paper claims ADFormer, a transformer combining differential attention with spatial and temporal aggregation, outperforms seven baselines in passenger-demand forecasting on NYC and Xi'an datasets while running faster than the…

desk verdict A plausible incremental architecture for demand forecasting that needs two fixes before its headline claim can be trusted: specify the DTW split and report uncertainty. read the letter →

arxiv 2506.02576 v3 pith:YRLWBKDK submitted 2025-06-03 cs.AI

Haichen Wang , Liu Yang , Xinyuan Zhang , Haomin Yu , Ming Li , Jilin Hu This is my paper
classification cs.AI
keywords passengerdemandforecastingspatio-temporaldifferentialattentionspatialclusteringtemporalaggregationdynamictimewarpingurbanmobility
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 passenger-demand forecasting improves when a model separates genuine spatial correlations from attention noise and then adds coarse-grained correlations between groups of regions. To do this it introduces ADFormer, a transformer that uses differential attention, the difference of two softmax score maps, for region-level spatial relations, groups regions with similar historical demand patterns through dynamic time warping, and combines these cluster-level relations with temporal self-attention and learned temporal aggregation. On the NYC taxi, NYC bike, and Xi'an taxi datasets, the reported results show lower MAE, RMSE, and MAPE than seven baselines at 30-minute, 90-minute, and 3-hour horizons, and the runtime table reports faster training and inference than PDFormer on the two NYC datasets. If the results hold, attention-based forecasters should model both fine-grained and coarse-grained spatio-temporal structure instead of relying on fixed masks.

What carries the argument

The load-bearing mechanism is differential spatial attention combined with a DTW-derived cluster map. Differential attention splits each query and key into two halves, computes two softmax attention matrices, and subtracts the second from the first, scaled by a learnable lambda, so spurious region-region correlations cancel out and the surviving map is denoised. In parallel, dynamic time warping on historical demand produces a distance matrix, regions are merged hierarchically into balanced clusters, and a cluster map M_cls aggregates region features into cluster features; cluster attention then projects the high-level correlations back to regions through a learnable separation matrix constrained by M_cls. The temporal branch applies ordinary self-attention across time steps and a separate aggregation attention with learned queries, with a restoration matrix built from time-of-day and day-of-week features.

What would settle it

Inspect the released implementation to determine whether the DTW distance matrix and cluster maps are computed before the 7:1:2 train/validation/test split, then rerun the experiments with clusters built exclusively from training data; if the cluster structure changes or the performance gap narrows substantially under that rerun, the claimed gains are at least partly an artifact of test-period leakage.

Watch

Extended reading notes

Core claim

The central claim is that high-level, non-pairwise dependencies, such as demand co-variation among regions that share a functional role like two distant commercial centers, carry forecasting information that ordinary region-level attention misses, and that this information can be extracted by aggregation. ADFormer builds a cluster map from historical demand similarity, runs attention at both the region and cluster levels, and fuses the two. Differential attention, computed as the difference of two softmax attention maps with a learnable scalar, removes part of the attention noise that the authors argue hurts naive attention and fixed-mask methods. The temporal side uses self-attention plus a learnable temporal aggregation query, so time steps are related both directly and through a hierarchical temporal matrix. The authors report consistent gains over seven baselines and interpret the ablation results as showing that spatial aggregation matters more for taxi demand while temporal aggregation matters more for bike demand.

Load-bearing premise

The method assumes the region groupings computed from historical demand by dynamic time warping are derived from training data only; if the full dataset including the test period determines those groupings, test information leaks into the model's spatial structure and the reported forecast comparison is invalid.

Editorial extensions

If this is right

  • Reported forecast errors (MAE, RMSE, MAPE) drop on NYC taxi, NYC bike, and Xi'an taxi at 30-minute, 90-minute, and 3-hour horizons when region-level and cluster-level spatial attention are combined.
  • On the two NYC datasets, training and inference runtimes per epoch are roughly 30 to 45 percent lower than PDFormer's, so the extra aggregation levels do not necessarily cost speed.
  • Ablation results imply the two correlation types are complementary: spatial aggregation contributes more for taxi demand and temporal aggregation more for bike demand, and dropping either attention branch hurts both datasets.
  • Parameter-study results imply that deeper spatial aggregation helps most at longer horizons, so three-hour forecasts improve more from three-level aggregation than single-step forecasts do.

Reading between the lines

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

  • A direct check of the released code is whether the DTW distance matrix and cluster maps are built from the training split only; the paper does not state this, and using the full period would leak test-period demand patterns into the model's fixed spatial structure.
  • Because the aggregation recipe is not tied to taxi or bike data, the same differential-attention-plus-clustering design could transfer to other spatio-temporal forecasting tasks with functional zones, such as traffic flow, energy demand, or station-level ridership.
  • The paper's attention-noise premise could be probed directly by measuring attention entropy or the concentration of scores before and after the differential operation, a diagnostic the paper does not report.
  • Since deeper hierarchies helped mainly at longer horizons, an extension would be a model that grows the number of clusters with the forecast horizon rather than using one fixed set.
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 / 6 minor

Summary. The paper proposes ADFormer, a transformer architecture for passenger demand forecasting that combines differential spatial attention with spatial cluster attention built from DTW-based region clustering, and couples these with temporal self-attention and a learnable temporal-aggregation attention. The model is evaluated on NYC-Taxi, NYC-Bike, and Xi'an-Taxi at 30-minute, 90-minute, and 3-hour horizons against seven baselines, with additional ablation, parameter, runtime, and case studies. The central claim is that ADFormer consistently outperforms state-of-the-art baselines in forecasting accuracy while maintaining computational efficiency.

Significance. If the empirical claims hold, ADFormer would be a useful contribution to passenger demand forecasting. The architectural idea of combining differential attention with hierarchical spatial and temporal aggregation is reasonable and reasonably novel, the released code is a practical strength, and the three-dataset, seven-baseline, multi-horizon evaluation is broad. The ablation study is informative about the relative contribution of spatial versus temporal aggregation, and the runtime comparison with PDFormer suggests practical efficiency. However, the current empirical support for the headline claim is weakened by a potential leakage risk in the DTW clustering step and by the absence of any uncertainty quantification.

major comments (4)
  1. [3.2 (Eqs. 8-12); 4.1] The DTW cluster map M_cls and the separation mask M_S_sep are static model inputs, but the manuscript never states that M_sim is computed from the training portion only. Under the chronological 7:1:2 split described in Section 4.1, if M_sim is computed over the full time series, the test-period labels influence the cluster assignments and the static mask through Eqs. (8)-(12), which would inflate ADFormer's results relative to baselines that do not use such test-derived structure. The authors should state explicitly that M_sim, M_cls, and M_S_sep are derived only from the training split, specify the normalization applied before the DTW computation, and confirm this behavior in the released code.
  2. [4.2, Table 1] The statement in Section 4.2 that 'our model consistently outperforms baselines' is contradicted by the reported numbers. On Xi'an-Taxi 30-min, ADFormer's MAPE is 22.848, which is worse than GWNet (22.631), MTGNN (22.159), and AGCRN (22.353); on NYC-Taxi 90-min, ADFormer's MAPE is 20.348, which is worse than MTGNN (19.758). The abstract and Section 1 also claim that ADFormer 'surpasses state-of-the-art baselines,' which is too strong as stated. The claims should be qualified to the settings where ADFormer is actually best, or the experiments should be extended to support the stronger claim.
  3. [4.2, Table 1 and Table 3] No standard deviations, confidence intervals, or significance tests are reported for any of the accuracy comparisons. Every number in Tables 1 and 3 appears to come from a single run, and several of the reported improvements are small in absolute terms. Without repeated runs and some form of paired test, the reader cannot judge whether the differences are stable. Please report mean and standard deviation over at least three to five seeds for the full comparison, and provide significance tests for the datasets and horizons where ADFormer is claimed to be best.
  4. [3.2, Eq. (9)] Equation (9) is dimensionally inconsistent as written: M_cls X_raw is in R^{T x M x D}, while T_d and T_w remain in R^{T x N x 1} and R^{T x N x 7}, so the concatenation cannot produce a tensor in R^{T x M x D'}. Please specify whether the time features are also aggregated to the cluster level (for example, by applying M_cls to the full X_full or by averaging within each cluster) and correct the notation accordingly. As written, the formula cannot be implemented without additional unstated steps.
minor comments (6)
  1. [4.1, Metrics] The exclusion threshold of 5 for low-demand points should be justified, and the sensitivity of the conclusions to this threshold should be checked. Since the same threshold is applied to all models, this is not a fatal flaw, but the choice deserves a sentence of justification.
  2. [4.2, Table 2] The computational-efficiency claim is based only on a runtime comparison with PDFormer; the other baselines are not included in the runtime measurements. The claim in the abstract and Section 1 that ADFormer maintains computational efficiency should therefore be scoped accordingly, or the runtime comparison should be extended.
  3. [4.3, Figure 4] The ablation results are shown only as inverted and normalized scores, with no numeric values reported. This makes effect sizes difficult to assess; please include the underlying numbers in the text or in a supplementary table.
  4. [3.2, DTW clustering] Please clarify what 'historical demand data' means in the DTW step: how many time steps are used, whether the same chronological window as the training set is used, and how the number of clusters and the cluster-rebalancing threshold are chosen in practice.
  5. [3.3, Eq. (17)] The restoration matrix M_temp_sep is described only briefly; its role and its dimensions (N x T x P) should be explained more carefully, since Eq. (17) projects the time features through W_sep but the text does not make explicit how the result acts as a separation mask.
  6. [4.1, Experimental Settings] Baseline hyperparameters, training epochs, early-stopping criteria, and dataset normalization details are not reported. Adding a reproducibility paragraph with these details would strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ADFormer's components are independently defined and evaluated against external baselines; the DTW-based cluster map is feature construction, not a circular prediction.

full rationale

The paper's central claim is an empirical one: ADFormer outperforms state-of-the-art baselines on public demand-forecasting datasets. The derivation chain consists of architectural choices (differential attention, spatial cluster attention, temporal aggregation attention) whose definitions do not presuppose the reported outcomes. The DTW-based spatial cluster map (Eqs. 8-12) is computed from historical demand and used as a static aggregation input; this is a feature-construction step rather than a fitted parameter renamed as a prediction. The model still learns attention parameters and is evaluated against held-out test data, so no reported metric reduces by construction to its inputs. Self-citations appear in the references but are not load-bearing for the main result; the differential attention mechanism is cited from external work (Ye et al., 2024), and DTW is cited from Berndt and Clifford. The concern that the cluster map may be computed over the full period including the test set is a data-leakage/soundness issue, not circularity, and is not evidenced by the paper's equations as a definitional equivalence. Therefore the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim relies on hand-chosen cluster counts, an unspecified rebalancing threshold, and the domain assumption that demand-similarity clusters are temporally stable. There are no new physical or conceptual entities.

free parameters (5)
  • Spatial cluster counts and hierarchy levels = [64,16] and [96,16,8] tested on Xi'an; no universal setting
    The model's high-level spatial branch exists only because these counts are chosen by hand; performance varies with the choice.
  • Low-demand evaluation threshold = 5 (counts per 30-minute interval)
    Points below this count are excluded from all metrics, following Yao et al.; the value is not justified and can change MAPE rankings.
  • Cluster rebalancing size threshold = unspecified
    Regions are moved between clusters when a cluster exceeds an unspecified size threshold; this affects the cluster map and downstream attention.
  • Architecture hyperparameters = hidden dimension in {32,64,128}, depth in {4,6,8}
    These are tuned per dataset, not derived; final selected values are not reported in the paper.
  • DTW similarity settings = not stated
    The distance matrix M_sim depends on DTW window and distance choices that are not reported.
assumptions (5)
  • domain assumption Regions with similar historical demand will continue to have similar future demand
    The entire spatial aggregation branch assumes that DTW-based functional similarity is stable over time; if it shifts, cluster attention adds noise instead of signal.
  • domain assumption Differential attention reduces harmful attention noise in spatial forecasting
    Borrowed from Ye et al. (2024) for LLMs; the paper does not prove it transfers to urban demand data.
  • domain assumption Six past 30-minute steps contain enough information to forecast up to six future steps
    The problem setup fixes T=6 and H up to 6 without sensitivity analysis on input length.
  • domain assumption Evaluation that excludes low-demand points is a fair basis for comparing models
    The threshold changes the effective test set; all models are evaluated on the same remaining points, but this is a modeling choice that can favor high-volume regions.
  • standard math Softmax attention, dynamic time warping, and agglomerative clustering work as expected
    Standard algorithms are used without re-derivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ADFormer: Aggregation Differential Transformer for Passenger Demand Forecasting." pith.science (2026). https://pith.science/paper/YRLWBKDK

@misc{pith2026250602576,
  author       = {Pith},
  title        = {Pith review of: ADFormer: Aggregation Differential Transformer for Passenger Demand Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YRLWBKDK}},
  note         = {Machine review of arXiv:2506.02576}
}
read the original abstract

Passenger demand forecasting helps optimize vehicle scheduling, thereby improving urban efficiency. Recently, attention-based methods have been used to adequately capture the dynamic nature of spatio-temporal data. However, existing methods that rely on heuristic masking strategies cannot fully adapt to the complex spatio-temporal correlations, hindering the model from focusing on the right context. These works also overlook the high-level correlations that exist in the real world. Effectively integrating these high-level correlations with the original correlations is crucial. To fill this gap, we propose the Aggregation Differential Transformer (ADFormer), which offers new insights to demand forecasting promotion. Specifically, we utilize Differential Attention to capture the original spatial correlations and achieve attention denoising. Meanwhile, we design distinct aggregation strategies based on the nature of space and time. Then, the original correlations are unified with the high-level correlations, enabling the model to capture holistic spatio-temporal relations. Experiments conducted on taxi and bike datasets confirm the effectiveness and efficiency of our model, demonstrating its practical value. The code is available at https://github.com/decisionintelligence/ADFormer.

Figures

Figures reproduced from arXiv: 2506.02576 by the authors.

Figure 1
Figure 1. Illustration of spatial attention noise and high-level corre [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of ADFormer. where Xemb ∈ R T ×N×d . Wraw ∈ R D×d , WTd ∈ R 1×d and WTw ∈ R 7×d are learnable parameters. Since the self-attention mechanism does not inherently contain sequential information, we adopt the classical posi￾tional encoding scheme, leveraging the periodicity of sine and cosine functions to generate positional information for the in￾put sequence, namely Xpos ∈ R T ×d . Furthermore, … view at source ↗
Figure 3
Figure 3. Systemic Temporal Attention. where W Tmp Q , WTmp K , WTmp V ∈ R d×d t 1 are learnable map￾ping parameters. Then, we compute the attention scores to obtain the cor￾relations between time steps, thereby deriving the output of temporal self attention: TSA(XT ) = softmax(A Tmp)V Tmp , (14) where ATmp = QTmp(KTmp) T √ d t 1 ∈ R N×T ×T represents the correlations between steps among regions. Temporal Aggregation Attentio… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Ablation study on NYC datasets. (3) The Differential Attention mechanism is particularly effective in spatially dominant scenarios like taxi demand, but less effective when temporal dependencies, such as those in bike demand, play a more significant role. This is mainl…
Figure 5
Figure 5. Figure 5: Spatial aggregation and high-level correlations. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 26 canonical work pages

  1. [1]

    Layer normalization.arXiv preprint arXiv:1607.06450,

    [Ba, 2016] Jimmy Lei Ba. Layer normalization.arXiv preprint arXiv:1607.06450,

    arXiv 2016

  2. [9]

    Stochastic origin- destination matrix forecasting using dual-stage graph con- volutional, recurrent neural networks

    [Huet al., 2020 ] Jilin Hu, Bin Yang, Chenjuan Guo, Chris- tian S Jensen, and Hui Xiong. Stochastic origin- destination matrix forecasting using dual-stage graph con- volutional, recurrent neural networks. In2020 IEEE 36th International conference on data engineering, pages 1417–1428. IEEE,

    work page 2020

  3. [10]

    Std-plm: Understanding both spa- tial and temporal properties of spatial-temporal data with plm

    [Huanget al., 2025 ] Yiheng Huang, Xiaowei Mao, Sheng- nan Guo, Yubin Chen, Junfeng Shen, Tiankuo Li, Youfang Lin, and Huaiyu Wan. Std-plm: Understanding both spa- tial and temporal properties of spatial-temporal data with plm. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 11817–11825,

    work page 2025

  4. [12]

    Deep multi-view graph- based network for citywide ride-hailing demand predic- tion.Neurocomputing, 510:79–94,

    [Jinet al., 2022 ] Guangyin Jin, Zhexu Xi, Hengyu Sha, Yanghe Feng, and Jincai Huang. Deep multi-view graph- based network for citywide ride-hailing demand predic- tion.Neurocomputing, 510:79–94,

    work page 2022

  5. [14]

    Foundts: Comprehen- sive and unified benchmarking of foundation models for time series forecasting.arXiv preprint arXiv:2410.11802,

    [Liet al., 2024 ] Zhe Li, Xiangfei Qiu, Peng Chen, Yihang Wang, Hanyin Cheng, Yang Shu, Jilin Hu, Chenjuan Guo, Aoying Zhou, Qingsong Wen, et al. Foundts: Comprehen- sive and unified benchmarking of foundation models for time series forecasting.arXiv preprint arXiv:2410.11802,

    arXiv 2024

  6. [15]

    Asstformer: Adaptive sparse spatial-temporal transformer for effective traffic forecasting

    [Lianget al., 2024 ] Ruifeng Liang, Zhiheng Li, and Kai Zhang. Asstformer: Adaptive sparse spatial-temporal transformer for effective traffic forecasting. In2024 5th International Seminar on Artificial Intelligence, Network- ing and Information Technology, pages 1847–1852,

    work page 2024

  7. [16]

    Spatio-temporal adaptive embedding makes vanilla transformer sota for traffic forecasting

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

    work page 2023

  8. [17]

    Stpsformer: Spatial-temporal probsparse transformer for long-term traffic flow forecasting

    [Luet al., 2024 ] Jun Lu, Dan Wang, and Zhanquan Wang. Stpsformer: Spatial-temporal probsparse transformer for long-term traffic flow forecasting. In2024 International Joint Conference on Neural Networks, pages 1–8,

    work page 2024

Show all 32 references
  1. [18]

    Stg4traffic: A survey and bench- mark of spatial-temporal graph neural networks for traffic prediction.CoRR, abs/2307.00495,

    [Luoet al., 2023 ] Xunlian Luo, Chunjiang Zhu, Detian Zhang, and Qing Li. Stg4traffic: A survey and bench- mark of spatial-temporal graph neural networks for traffic prediction.CoRR, abs/2307.00495,

    2023 arXiv

  2. [19]

    Spatial-temporal dynamic graph attention networks for ride-hailing demand predic- tion.arXiv preprint arXiv:2006.05905,

    [Pianet al., 2020 ] Weiguo Pian, Yingbo Wu, Xiangmou Qu, Junpeng Cai, and Ziyi Kou. Spatial-temporal dynamic graph attention networks for ride-hailing demand predic- tion.arXiv preprint arXiv:2006.05905,

    2020 arXiv

  3. [20]

    Tfb: Towards comprehensive and fair benchmarking of time se- ries forecasting methods.Proceedings of the VLDB En- dowment, 17(9):2363–2377,

    [Qiuet al., 2024 ] Xiangfei Qiu, Jilin Hu, Lekui Zhou, Xingjian Wu, Junyang Du, Buang Zhang, Chenjuan Guo, Aoying Zhou, Christian S Jensen, Zhenli Sheng, et al. Tfb: Towards comprehensive and fair benchmarking of time se- ries forecasting methods.Proceedings of the VLDB En- do...

    2024

  4. [21]

    A comprehensive sur- vey of deep learning for multivariate time series fore- casting: A channel strategy perspective.arXiv preprint arXiv:2502.10721,

    [Qiuet al., 2025a ] Xiangfei Qiu, Hanyin Cheng, Xingjian Wu, Jilin Hu, and Chenjuan Guo. A comprehensive sur- vey of deep learning for multivariate time series fore- casting: A channel strategy perspective.arXiv preprint arXiv:2502.10721,

  5. [22]

    Spatial-temporal identity: A simple yet effective baseline for multivariate time series forecast- ing

    [Shaoet al., 2022 ] Zezhi Shao, Zhao Zhang, Fei Wang, Wei Wei, and Yongjun Xu. Spatial-temporal identity: A simple yet effective baseline for multivariate time series forecast- ing. InProceedings of the 31st ACM International Con- ference on Information & Knowledge Management,...

    2022

  6. [23]

    Graph wavenet for deep spatial-temporal graph modeling

    [Wuet al., 2019 ] Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. Graph wavenet for deep spatial-temporal graph modeling. InInternational Joint Conference on Artificial Intelligence 2019, pages 1907– 1913,

    2019

  7. [24]

    Con- necting the dots: Multivariate time series forecasting with graph neural networks

    [Wuet al., 2020 ] Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, Xiaojun Chang, and Chengqi Zhang. Con- necting the dots: Multivariate time series forecasting with graph neural networks. InProceedings of the 26th ACM SIGKDD international conference on knowledge discov- ery ...

    2020

  8. [25]

    Catch: Channel-aware multivariate time series anomaly detection via frequency patching

    [Wuet al., 2025 ] Xingjian Wu, Xiangfei Qiu, Zhengyu Li, Yihang Wang, Jilin Hu, Chenjuan Guo, Hui Xiong, and Bin Yang. Catch: Channel-aware multivariate time series anomaly detection via frequency patching. InICLR,

    2025

  9. [26]

    Learning dynamic and hierarchical traffic spatiotemporal features with transformer.IEEE Transactions on Intelli- gent Transportation Systems, 23(11):22386–22399,

    [Yanet al., 2021 ] Haoyang Yan, Xiaolei Ma, and Ziyuan Pu. Learning dynamic and hierarchical traffic spatiotemporal features with transformer.IEEE Transactions on Intelli- gent Transportation Systems, 23(11):22386–22399,

    2021

  10. [28]

    Differential trans- former.arXiv preprint arXiv:2410.05258,

    [Yeet al., 2024 ] Tianzhu Ye, Li Dong, Yuqing Xia, Yutao Sun, Yi Zhu, Gao Huang, and Furu Wei. Differential trans- former.arXiv preprint arXiv:2410.05258,

    2024 arXiv

  11. [29]

    Regu- larized graph structure learning with semantic knowledge for multi-variates time-series forecasting

    [Yuet al., 2022 ] Hongyuan Yu, Ting Li, Weichen Yu, Jian- guo Li, Yan Huang, Liang Wang, and Alex Liu. Regu- larized graph structure learning with semantic knowledge for multi-variates time-series forecasting. InInternational Joint Conference on Artificial Intelligence 2022, p...

    2022

  12. [30]

    Dnn-based prediction model for spatio-temporal data

    [Zhanget al., 2016 ] Junbo Zhang, Yu Zheng, Dekang Qi, Ruiyuan Li, and Xiuwen Yi. Dnn-based prediction model for spatio-temporal data. InProceedings of the 24th ACM SIGSPATIAL international conference on advances in ge- ographic information systems, pages 1–4,

    2016

  13. [31]

    Gman: A graph multi- attention network for traffic prediction

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

    2020

  14. [32]

    Soup: Spatial-temporal demand forecasting and com- petitive supply in transportation.IEEE Transactions on Knowledge and Data Engineering, 35(2):2034–2047, 2021

    [Zhenget al., 2021 ] Bolong Zheng, Qi Hu, Lingfeng Ming, Jilin Hu, Lu Chen, Kai Zheng, and Christian S Jensen. Soup: Spatial-temporal demand forecasting and com- petitive supply in transportation.IEEE Transactions on Knowledge and Data Engineering, 35(2):2034–2047, 2021

    2021

  15. [1994]

    Price- and-time-aware dynamic ridesharing

    [Chenet al., 2018 ] Lu Chen, Qilu Zhong, Xiaokui Xiao, Yunjun Gao, Pengfei Jin, and Christian S Jensen. Price- and-time-aware dynamic ridesharing. In2018 IEEE 34th international conference on data engineering, pages 1061–1072. IEEE,

    2018

  16. [2016]

    Adaptive graph convolutional recurrent network for traffic forecasting.Advances in neural infor- mation processing systems, 33:17804–17815,

    [Baiet al., 2020 ] Lei Bai, Lina Yao, Can Li, Xianzhi Wang, and Can Wang. Adaptive graph convolutional recurrent network for traffic forecasting.Advances in neural infor- mation processing systems, 33:17804–17815,

    2020

  17. [2018]

    Machine learning for public transportation de- mand prediction: A systematic literature review.Engineer- ing Applications of Artificial Intelligence, 137:109166,

    [di Torrepadulaet al., 2024 ] Franca Rocco di Torrepadula, Enea Vincenzo Napolitano, Sergio Di Martino, and Nicola Mazzocca. Machine learning for public transportation de- mand prediction: A systematic literature review.Engineer- ing Applications of Artificial Intelligence, 13...

    2024

  18. [2019]

    Self-supervised spatial-temporal bottleneck attentive network for efficient long-term traffic forecasting

    [Guoet al., 2023 ] Shengnan Guo, Youfang Lin, Letian Gong, Chenyu Wang, Zeyu Zhou, Zekai Shen, Yiheng Huang, and Huaiyu Wan. Self-supervised spatial-temporal bottleneck attentive network for efficient long-term traffic forecasting. In2023 IEEE 39th International Conference on ...

    2023

  19. [2020]

    Using dynamic time warping to find patterns in time series

    [Berndt and Clifford, 1994] Donald J Berndt and James Clif- ford. Using dynamic time warping to find patterns in time series. InProceedings of the 3rd international conference on knowledge discovery and data mining, pages 359–370,

    1994

  20. [2021]

    Deep multi-view spatial-temporal network for taxi demand prediction

    [Yaoet al., 2018 ] Huaxiu Yao, Fei Wu, Jintao Ke, Xianfeng Tang, Yitian Jia, Siyu Lu, Pinghua Gong, Jieping Ye, and Zhenhui Li. Deep multi-view spatial-temporal network for taxi demand prediction. InProceedings of the AAAI con- ference on artificial intelligence, volume 32, pa...

    2018

  21. [2022]

    Hexagon- based convolutional neural network for supply-demand forecasting of ride-sourcing services.IEEE Transactions on Intelligent Transportation Systems, 20(11):4160–4173,

    [Keet al., 2018 ] Jintao Ke, Hai Yang, Hongyu Zheng, Xiqun Chen, Yitian Jia, Pinghua Gong, and Jieping Ye. Hexagon- based convolutional neural network for supply-demand forecasting of ride-sourcing services.IEEE Transactions on Intelligent Transportation Systems, 20(11):4160–4173,

    2018

  22. [2023]

    Deep residual learning for image recog- nition

    [Heet al., 2016 ] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog- nition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778,

    2016

  23. [2024]

    Spatiotemporal multi-graph convolution network for ride- hailing demand forecasting

    [Genget al., 2019 ] Xu Geng, Yaguang Li, Leye Wang, Lingyu Zhang, Qiang Yang, Jieping Ye, and Yan Liu. Spatiotemporal multi-graph convolution network for ride- hailing demand forecasting. InProceedings of the AAAI conference on artificial intelligence, volume 33, pages 3656–3663,

    2019

  24. [2025]

    Pdformer: Propagation delay- aware dynamic long-range transformer for traffic flow pre- diction

    [Jianget al., 2023 ] Jiawei Jiang, Chengkai Han, Wayne Xin Zhao, and Jingyuan Wang. Pdformer: Propagation delay- aware dynamic long-range transformer for traffic flow pre- diction. InProceedings of the AAAI conference on artifi- cial intelligence, volume 37, pages 4365–4373,

    2023

Pith tools

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