Pith. sign in

REVIEW 3 major objections 43 references

A bike-demand model uses a few agent tokens so global station and time interactions cost only linear work, and it beats strong baselines on NYC and Chicago systems.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 01:40 UTC pith:BSQPLWRZ

load-bearing objection Solid applied packaging of agent attention for bike demand; gains look real under the authors’ protocol, but single-run tables and no agent-count sensitivity leave the “linear yet fully global” claim only provisionally supported. the 3 major comments →

arxiv 2607.06614 v1 pith:BSQPLWRZ submitted 2026-07-07 cs.LG cs.AI

STAGformer: A Spatio-temporal Agent Graph Transformer for Micro Mobility Demand Forecasting

Ye Zihao This is my paper
classification cs.LG cs.AI
keywords bike-sharingdemand forecastinggraph neural networkagent attentionspatio-temporal transformerlinear complexitymicro-mobility
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

City bike-sharing demand is hard to forecast because stations influence one another across long distances and time, while standard attention grows too expensive as the network grows. This paper claims that a small set of learnable spatial and temporal “agent” tokens can first gather information from every station and time step, then broadcast a global summary back, so long-range interactions are captured at linear cost O(NT) instead of quadratic. The resulting STAGformer stacks that agent attention with a feature encoder that also uses weather, time and points of interest, a graph-propagation step for local spatial neighbors, and a temporal convolution for short patterns. On two large real systems—NYC Citi-Bike and Chicago Divvy-Bike—the full model reports lower RMSE and MAE than a range of recent baselines, and ablations show the biggest drop when the agent attention is removed. If the claim holds, operators can keep global modeling power while still running forecasts on city-scale networks.

Core claim

STAGformer shows that two-step Softmax agent attention, mediated by a small fixed number of spatial and temporal agent tokens, can replace full self-attention for station-level bike demand forecasting, cutting complexity from O((NT)2) to O(NT) while still producing lower RMSE and MAE than strong baselines on NYC and Chicago data.

What carries the argument

Spatio-temporal agent attention: learnable spatial agents (ns ≪ N) and temporal agents (nt ≪ T) first aggregate global keys/values, then broadcast the summary back to every node and time step; a depthwise-convolution residual keeps feature diversity.

Load-bearing premise

That a small fixed number of agent tokens plus a simple residual convolution is enough to keep the full global modeling power of Softmax attention for real bike-demand patterns.

What would settle it

Re-run the same NYC and Chicago experiments with the agent-attention module replaced by full separated Softmax attention (or with agent counts driven to N and T) and check whether RMSE/MAE improve enough to erase the claimed accuracy advantage of the linear version.

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

If this is right

  • City-scale bike rebalancing systems can run multi-step station forecasts that still see long-range station interactions without quadratic attention cost.
  • The same linear agent pattern can be stacked with ordinary graph propagation and temporal convolution, so local and global cues remain complementary rather than mutually exclusive.
  • External context (weather, time, POIs) can be fused once at the encoder and still benefit from the global agent stage.
  • Ablation evidence implies that removing the agent stage hurts accuracy more than removing local graph or convolution modules, so global modeling is not optional for these systems.

Where Pith is reading between the lines

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

  • If agent counts stay small while city networks keep growing, the same linear recipe may transfer to e-scooter or dockless fleets that share the same spatio-temporal structure.
  • Sensitivity of accuracy to the exact agent count (ns, nt) is left open; an adaptive or learned agent budget would be a natural next test.
  • The same two-step aggregate-then-broadcast idea could be tried on other large spatio-temporal graphs (traffic sensors, EV chargers) where full attention is already known to be too costly.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. STAGformer is a spatio-temporal forecasting architecture for station-level bike-sharing demand that fuses a multi-source encoder (node features + weather/time/POIs), graph propagation (fixed or adaptive adjacency with K-step fusion), temporal 1D convolution, and a two-step agent attention module. Separate learnable spatial and temporal agent tokens (ns, nt ≪ N, T) first aggregate then broadcast Softmax attention, reducing joint self-attention cost from O((NT)²) to O(NT) while a depthwise-convolution residual is claimed to preserve feature diversity. The model is evaluated on NYC Citi-Bike and Chicago Divvy-Bike (Sept/Oct 2025 windows) against nine baselines; Table 2 reports lower RMSE/MAE for STAGformer on all four city-month splits, and ablations attribute the largest drop to removal of agent attention.

Significance. If the linear agent construction truly retains the long-range modeling power of full Softmax attention, the work supplies a practical, scalable architecture for city-scale micro-mobility forecasting that can be deployed without the quadratic memory wall of standard Transformers. The hybrid design (local graph + local conv + global agents + external fusion) is a coherent engineering contribution, and the two-city empirical protocol with module ablations is stronger than many pure architectural papers. The paper does not ship code, multi-seed statistics, or agent-count sensitivity, so the significance remains provisional on those missing checks.

major comments (3)
  1. Table 2 and §5.4 claim that STAGformer “consistently outperforms” nine baselines with “significant improvements” in RMSE/MAE. All numbers appear to be single-run point estimates; no multi-seed means, standard deviations, or statistical tests are reported. Given the free parameters listed in §5.2 (learning rate, batch size, epochs, early stopping) and the small absolute margins (e.g., NYC Sept RMSE 0.4874 vs STAEformer 0.5106), the superiority claim is not yet secured against run-to-run variance.
  2. §4.3.2–4.3.5 and the “w/o Attn” ablation assert that a fixed small agent set (example ns = nt = 32) plus DWC residual fully preserves the global Softmax receptive field that the paper itself treats as critical. No sensitivity sweep over ns/nt, no comparison against full Softmax or separated attention under matched compute, and no analysis of information loss appear. Without that evidence the central efficiency-plus-expressiveness claim remains an untested modeling assumption rather than a demonstrated result.
  3. §5.1–5.2 describe data preparation (hourly alignment, 150 m POI radius, external adjacency) but omit the concrete train/validation/test temporal split, the exact values of T and H used for the multi-horizon claim, and whether the same hyper-parameters were used for every baseline. Reproducibility of Table 2 therefore cannot be verified from the manuscript alone.

Circularity Check

0 steps flagged

No circularity: standard supervised spatio-temporal forecasting model with free parameters trained and evaluated on held-out data; agent attention imported from external citation.

full rationale

STAGformer is an empirical deep-learning architecture paper. Historical node features X, global factors G and adjacency A_ext are mapped by a trainable encoder + graph propagation + temporal convolution + two-step agent attention (learnable agent tokens P_s, P_t of fixed small size n_s, n_t together with free projection matrices and DWC residual) to multi-horizon demand forecasts Ŷ. All parameters are optimized by supervised loss on training windows and evaluated on held-out months (Table 2, ablations §5.5). No equation equates a claimed prediction to a fitted constant by construction; complexity reduction O(NT) follows directly from the two Softmax steps with constant agents (Eqs. 12–15) and is independent of the numerical accuracy claims. The sole external reference for the agent mechanism is Han et al. [31] (ECCV 2024), whose authors do not overlap with the present single-author paper; no uniqueness theorem, ansatz or load-bearing result is imported from self-citation. Consequently the derivation chain contains no self-definitional, fitted-as-prediction, or self-citation circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The central empirical claim rests on standard deep-learning training assumptions plus several architectural free choices (agent counts, graph order, fusion of external features) that are not derived from first principles. No new physical entities are postulated; the “agents” are learnable tokens. Domain assumptions about station graphs and external covariates are conventional for the traffic-forecasting literature.

free parameters (5)
  • number of spatial agent tokens ns = example 32
    Fixed small integer (example 32) that controls capacity and cost of spatial agent attention; chosen by design, not derived.
  • number of temporal agent tokens nt = example 32
    Symmetric free choice for the temporal branch; example value 32.
  • graph propagation steps K and fusion weights αk
    Order of neighbor aggregation and Softmax-normalized fusion weights are architectural/fitted choices (§4.2.1).
  • adaptive graph temperature τ and node embeddings E
    Learnable parameters that define  when adaptive graph learning is used (§4.2.1).
  • hidden dimension d / C, heads H, conv kernel kt, learning rate, batch size, epochs = lr=1e-3, batch=8, 300 epochs
    Standard training and capacity hyperparameters that affect reported RMSE/MAE; only partial values given (lr 1e-3, batch 8, 300 epochs).
axioms (4)
  • domain assumption Bike-sharing demand can be usefully modeled as a graph G=(V,E) with node features and an external or adaptive adjacency matrix.
    Stated in §3.1–3.2 and used throughout the graph propagation module.
  • ad hoc to paper A two-step Softmax agent attention with a small agent set preserves global long-range dependency modeling while reducing complexity to O(NT).
    Imported from [31] and asserted for the spatio-temporal setting in §4.3; the paper’s efficiency and accuracy claims rest on this.
  • domain assumption External covariates (weather, time, POIs within 150 m) are informative for station demand and can be fused by linear projection and concatenation.
    Encoding scheme §4.1 and ablation w/o E.
  • domain assumption Standard supervised sliding-window forecasting with RMSE/MAE on held-out months is a valid measure of operational forecasting quality.
    Problem definition §3.2 and evaluation §5.3.
invented entities (1)
  • Spatial and temporal agent tokens (Ps, Pt) inside STAGformer no independent evidence
    purpose: Mediate two-step aggregation and broadcasting so global attention costs O(NT) instead of quadratic.
    Architectural tokens adapted from Agent Attention [31]; not a new physical object. Independent evidence is only the paper’s own ablations and baselines, not an external falsifiable prediction outside this forecasting task.

pith-pipeline@v1.1.0-grok45 · 19324 in / 3418 out tokens · 43105 ms · 2026-07-11T01:40:32.113223+00:00 · methodology

0 comments
read the original abstract

Accurate station-level demand forecasting is essential for the efficient operation of bike-sharing systems, yet it remains challenging due to complex spatio-temporal dependencies and the large scale of urban networks. This paper presents STAGformer, a Spatio-Temporal Agent Graph Transformer that achieves efficient global modeling with linear computational complexity. The model introduces a two-step agent attention mechanism, where a small set of learnable spatial and temporal agent tokens first aggregate global information and then broadcast it back to individual stations and time steps, effectively capturing long-range interactions while reducing the quadratic cost of standard self-attention to O(NT). STAGformer integrates four core modules: a spatio-temporal encoder that fuses dynamic node features with external contextual factors (weather, time, points of interest), a graph propagation module for spatial neighbor aggregation, a temporal convolution module for local pattern extraction, and the agent attention module for global dependency modeling. Extensive experiments on two real-world datasets -- NYC Citi-Bike and Chicago Divvy-Bike -- demonstrate that STAGformer consistently outperforms state-of-the-art baselines across multiple prediction horizons, achieving significant improvements in both RMSE and MAE. Ablation studies validate the contribution of each component, with the agent attention mechanism proving critical for modeling global spatio-temporal dependencies.

Figures

Figures reproduced from arXiv: 2607.06614 by Ye Zihao.

Figure 1
Figure 1. Figure 1: Spatial imbalance of bike-sharing demand and returns in Chicago on Wednesday, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of model performance and efficiency (NYC, September 2025, MAE; bubble [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overall architecture of Spatio-Temporal Agent Graph Transformer (STAGformer). Yˆ t+1:t+H = f (Xt−T +1:t , Gt−T +1:t , Aext), (1) where: Yˆ t+1:t+H ∈ R B×H×N×D denotes the predicted D-dimensional targets (e.g., rental and return counts) for each station over the future H steps; Xt−T +1:t ∈ R B×T ×N×Fn is the his￾torical node features within the observation window; Gt−T +1:t ∈ R B×T ×Fg is the global feature… view at source ↗
Figure 4
Figure 4. Figure 4: Arrival and departure patterns of bike-sharing stations in New York City. [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation study results on New York and Chicago datasets. Bars represent error metrics [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Forecast morning peak and evening peak departure and arrival heatmaps in New York [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages · 13 internal anchors

  1. [1]

    A comprehensive survey of deep learning for time series forecasting: architectural diversity and open challenges,

    J. Kim, H. Kim, H. Kim, D. Lee, and S. Yoon, “A comprehensive survey of deep learning for time series forecasting: architectural diversity and open challenges,”Artif. Intell. Rev., vol. 58, no. 7, p. 216, 2025

    work page 2025

  2. [2]

    Chatfield and H

    C. Chatfield and H. Xing,The Analysis of Time Series: An Introduction with R, 7th ed. Boca Raton, FL, USA: Chapman and Hall/CRC, 2019

    work page 2019

  3. [3]

    Advances in neural information processing systems 22,

    D. D. Lee, P. Pham, Y. Largman, and A. Ng, “Advances in neural information processing systems 22,”Neural Inf. Process. Syst., vol. 1, no. 1, pp. 1–11, 2009

    work page 2009

  4. [4]

    Zhou,Machine Learning

    Z.-H. Zhou,Machine Learning. Springer Nature, 2021

    work page 2021

  5. [5]

    Greedy function approximation: a gradient boosting machine,

    J. H. Friedman, “Greedy function approximation: a gradient boosting machine,”Ann. Statist., pp. 1189–1232, 2001

    work page 2001

  6. [6]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,”Neural Comput., vol. 9, no. 8, pp. 1735–1780, 1997

    work page 1997

  7. [7]

    Learning phrase representations using RNN encoder–decoder for statistical machine translation,

    K. Cho et al., “Learning phrase representations using RNN encoder–decoder for statistical machine translation,” inProc. Conf. Empir. Methods Nat. Lang. Process. (EMNLP), 2014, pp. 1724–1734

    work page 2014

  8. [8]

    Deep spatio-temporal residual networks for citywide crowd flows prediction,

    J. Zhang, Y. Zheng, and D. Qi, “Deep spatio-temporal residual networks for citywide crowd flows prediction,” inProc. AAAI Conf. Artif. Intell., vol. 31, no. 1, 2017. 17 STAGformer: Spatio-temporal Agent Graph Transformer

    work page 2017

  9. [9]

    Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting

    Y. Li, R. Yu, C. Shahabi, and Y. Liu, “Diffusion convolutional recurrent neural network: Data-driven traffic forecasting,”arXiv preprint arXiv:1707.01926, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  10. [10]

    Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting

    B. Yu, H. Yin, and Z. Zhu, “Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting,”arXiv preprint arXiv:1709.04875, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  11. [11]

    Attention based spatial-temporal graph con- volutional networks for traffic flow forecasting,

    S. Guo, Y. Lin, N. Feng, C. Song, and H. Wan, “Attention based spatial-temporal graph con- volutional networks for traffic flow forecasting,” inProc. AAAI Conf. Artif. Intell., vol. 33, no. 1, 2019, pp. 922–929

    work page 2019

  12. [12]

    GMAN: A graph multi-attention network for traffic prediction,

    C. Zheng, X. Fan, C. Wang, and J. Qi, “GMAN: A graph multi-attention network for traffic prediction,” inProc. AAAI Conf. Artif. Intell., vol. 34, no. 1, 2020, pp. 1234–1241

    work page 2020

  13. [13]

    TrafficStream: A Streaming Traffic Flow Forecasting Framework Based on Graph Neural Networks and Continual Learning

    X. Chen, J. Wang, and K. Xie, “TrafficStream: A streaming traffic flow forecast- ing framework based on graph neural networks and continual learning,”arXiv preprint arXiv:2106.06273, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  14. [14]

    Graph WaveNet for Deep Spatial-Temporal Graph Modeling

    Z. Wu, S. Pan, G. Long, J. Jiang, and C. Zhang, “Graph WaveNet for deep spatial-temporal graph modeling,”arXiv preprint arXiv:1906.00121, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  15. [15]

    Adaptive graph convolutional recurrent network for traffic forecasting,

    L. Bai, L. Yao, C. Li, X. Wang, and C. Wang, “Adaptive graph convolutional recurrent network for traffic forecasting,” inAdv. Neural Inf. Process. Syst., vol. 33, 2020, pp. 17804– 17815

    work page 2020

  16. [16]

    Attention is all you need,

    A. Vaswani et al., “Attention is all you need,” inAdv. Neural Inf. Process. Syst., vol. 30, 2017

    work page 2017

  17. [17]

    Spatial-Temporal Transformer Networks for Traffic Flow Forecasting

    M. Xu et al., “Spatial-temporal transformer networks for traffic flow forecasting,”arXiv preprint arXiv:2001.02908, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2001

  18. [18]

    Generating Long Sequences with Sparse Transformers

    R. Child, S. Gray, A. Radford, and I. Sutskever, “Generating long sequences with sparse transformers,”arXiv preprint arXiv:1904.10509, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1904

  19. [19]

    Transformers are RNNs: Fast au- toregressive transformers with linear attention,

    A. Katharopoulos, A. Vyas, N. Pappas, and F. Fleuret, “Transformers are RNNs: Fast au- toregressive transformers with linear attention,” inProc. Int. Conf. Mach. Learn. (ICML), 2020, pp. 5156–5165

    work page 2020

  20. [20]

    Predicting bike usage for New York City’s bike sharing system,

    D. Singhvi, S. Singhvi, P. I. Frazier, S. G. Henderson, E. O’Mahony, D. B. Shmoys, and D. B. Woodard, “Predicting bike usage for New York City’s bike sharing system,” inProc. AAAI Workshop on Computational Sustainability, Austin, TX, USA, 2015

    work page 2015

  21. [21]

    Rethinking Attention with Performers

    K. Choromanski et al., “Rethinking attention with performers,”arXiv preprint arXiv:2009.14794, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2009

  22. [22]

    Graph Attention Networks

    P.Veličković, G.Cucurull, A.Casanova, A.Romero, P.Liò, andY.Bengio, “Graphattention networks,”arXiv preprint arXiv:1710.10903, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  23. [23]

    How Attentive are Graph Attention Networks?

    S. Brody, U. Alon, and E. Yahav, “How attentive are graph attention networks?”arXiv preprint arXiv:2105.14491, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  24. [24]

    GaAN: Gated Attention Networks for Learning on Large and Spatiotemporal Graphs

    J.Zhang, X.Shi, J.Xie, H.Ma, I.King, andD.-Y.Yeung, “GaAN:Gatedattentionnetworks for learning on large and spatiotemporal graphs,”arXiv preprint arXiv:1803.07294, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  25. [25]

    Connecting the dots: Mul- tivariate time series forecasting with graph neural networks,

    Z. Wu, S. Pan, G. Long, J. Jiang, X. Chang, and C. Zhang, “Connecting the dots: Mul- tivariate time series forecasting with graph neural networks,” inProc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Min., 2020, pp. 753–763. 18 STAGformer: Spatio-temporal Agent Graph Transformer

    work page 2020

  26. [26]

    Informer: Beyond efficient transformer for long sequence time-series fore- casting,

    H. Zhou et al., “Informer: Beyond efficient transformer for long sequence time-series fore- casting,” inProc. AAAI Conf. Artif. Intell., vol. 35, no. 12, 2021, pp. 11106–11115

    work page 2021

  27. [27]

    Set transformer: A frame- work for attention-based permutation-invariant neural networks,

    J. Lee, Y. Lee, J. Kim, A. Kosiorek, S. Choi, and Y. W. Teh, “Set transformer: A frame- work for attention-based permutation-invariant neural networks,” inProc. Int. Conf. Mach. Learn. (ICML), 2019, pp. 3744–3753

    work page 2019

  28. [28]

    Predicting travel demand of a bike sharing system using graph convolutional neural networks,

    A. Behroozi and A. Edrisi, “Predicting travel demand of a bike sharing system using graph convolutional neural networks,”Public Transp., vol. 17, no. 1, pp. 281–317, 2025

    work page 2025

  29. [29]

    Contextual data integration for bike-sharing demand prediction with graph neural networks in degraded weather conditions,

    R. Rochas, A. Furno, and N.-E. El Faouzi, “Contextual data integration for bike-sharing demand prediction with graph neural networks in degraded weather conditions,” inProc. IEEE Int. Conf. Intell. Transp. Syst. (ITSC), 2023, pp. 5436–5441

    work page 2023

  30. [30]

    A demand truncation and migration poisson model for real demand inference in free-floating bike-sharing system,

    J. Wang, T. Miwa, and T. Morikawa, “A demand truncation and migration poisson model for real demand inference in free-floating bike-sharing system,”IEEE Trans. Intell. Transp. Syst., vol. 24, no. 10, pp. 10525–10536, Oct. 2023

    work page 2023

  31. [31]

    Agent attention: On the integration of softmax and linear attention,

    D. Han et al., “Agent attention: On the integration of softmax and linear attention,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2024, pp. 124–140

    work page 2024

  32. [32]

    An adaptive spatial-temporal method capturing for short-term bike- sharing prediction,

    J. Feng and H. Liu, “An adaptive spatial-temporal method capturing for short-term bike- sharing prediction,”IEEE Trans. Intell. Transp. Syst., vol. 25, no. 11, pp. 16761–16774, Nov. 2024

    work page 2024

  33. [33]

    BGM: Demand prediction for expanding bike- sharing systems with dynamic graph modeling,

    Y. Zhao, H. Wen, X. Zhang, and M. Luo, “BGM: Demand prediction for expanding bike- sharing systems with dynamic graph modeling,” inProc. Int. Joint Conf. Artif. Intell. (IJCAI), 2025, pp. 10008–10016

    work page 2025

  34. [34]

    Bike-sharing demand prediction based on dynamic time warping and spatio-temporal graph attention network,

    Z. Xiang, F. Zeng, L. Liu, J. Wu, S. Mumtaz, and V. C. M. Leung, “Bike-sharing demand prediction based on dynamic time warping and spatio-temporal graph attention network,” IEEE Trans. Intell. Transp. Syst., 2025, to be published

    work page 2025

  35. [35]

    Dual stream meta learning for road surface classification and riding event detection on shared bikes,

    S. Jiang, Z. Strout, B. He, D. Peng, P. B. Shull, and B. P. L. Lo, “Dual stream meta learning for road surface classification and riding event detection on shared bikes,”IEEE Trans. Syst. Man Cybern. Syst., vol. 53, no. 11, pp. 7188–7200, Nov. 2023

    work page 2023

  36. [36]

    GeOT: a spatially explicit framework for evaluating spatio-temporal predictions,

    N. Wiedemann, T. Uscidda, and M. Raubal, “GeOT: a spatially explicit framework for evaluating spatio-temporal predictions,”Int. J. Geogr. Inf. Sci., vol. 39, no. 10, pp. 2236– 2266, 2025

    work page 2025

  37. [37]

    Incorporating the impact of spatio-temporal interactions on bicycle sharing system demand: A case study of New York CitiBike system,

    A. Faghih-Imani and N. Eluru, “Incorporating the impact of spatio-temporal interactions on bicycle sharing system demand: A case study of New York CitiBike system,”J. Transp. Geogr., vol. 54, pp. 218–227, 2016

    work page 2016

  38. [38]

    Micromobility Flow Prediction: A Bike Sharing Station-level Study via Multi-level Spatial-Temporal Attention Neural Network

    X. Yang, J. Wang, S. Han, and S. He, “Micromobility flow prediction: A bike shar- ing station-level study via multi-level spatial-temporal attention neural network,”arXiv preprint arXiv:2507.16020, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    MoDE-Boost: Boosting shared mobility demand with edge-ready prediction models,

    A. Tziorvas, G. S. Theodoropoulos, and Y. Theodoridis, “MoDE-Boost: Boosting shared mobility demand with edge-ready prediction models,”arXiv preprint arXiv:2602.16573, 2026

    work page arXiv 2026

  40. [40]

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

    H. Liu et al., “Spatio-temporal adaptive embedding makes vanilla transformer sota for traffic forecasting,” inProc. ACM Int. Conf. Inf. Knowl. Manag. (CIKM), 2023, pp. 4125–4129. 19 STAGformer: Spatio-temporal Agent Graph Transformer

    work page 2023

  41. [41]

    STGformer: Efficient Spatiotemporal Graph Transformer for Traffic Forecasting

    H. Wang, J. Chen, T. Pan, Z. Dong, L. Zhang, R. Jiang, and X. Song, “STGformer: Efficient spatiotemporal graph transformer for traffic forecasting,”arXiv preprint arXiv:2410.00385, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  42. [42]

    STGFormer: Spatio-temporal GraphFormer for 3D human pose estimation in video,

    Y. Liu and Z. Zhang, “STGFormer: Spatio-temporal GraphFormer for 3D human pose estimation in video,”Pattern Recognit., p. 112239, 2025

    work page 2025

  43. [43]

    T-STAR: A Context-Aware Transformer Framework for Short-Term Probabilistic Demand Forecasting in Dock-Based Shared Micro-Mobility

    J. Cheng, G. H. de Almeida Correia, O. Cats, and S. S. Azadeh, “T-STAR: A context-aware transformerframeworkforshort-termprobabilisticdemandforecastingindock-basedshared micro-mobility,”arXiv preprint arXiv:2602.06866, 2026. 20

    work page internal anchor Pith review Pith/arXiv arXiv 2026