pith. sign in

arxiv: 2603.13609 · v2 · pith:IK4BIF3Gnew · submitted 2026-03-13 · 💻 cs.CV

A Grid-Based Framework for E-Scooter Demand Representation and Temporal Input Design for Deep Learning: Evidence from Austin, Texas

Mohammad Sahnoon , Merkebe Getachew Demissie , Roberto Souza This is my paper

Pith reviewed 2026-05-21 11:12 UTC · model grok-4.3

classification 💻 cs.CV
keywords e-scooter demand predictionspatiotemporal modelingtemporal input designgrid-based representationdeep learningUNETmicromobilityAustin Texas
0
0 comments X

The pith

A grid-based pipeline with statistically validated temporal inputs reduces deep learning error for e-scooter demand prediction by up to 37 percent.

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

The paper shows how to turn raw e-scooter trip records into a consistent set of hourly grid images that represent pickup and dropoff demand across Austin. It then uses a correlation-plus-error procedure plus an ablation study with statistical tests to choose which past hours to feed into the model instead of relying on simple adjacent-hour or fixed-period choices. This matters because better demand forecasts let operators move vehicles to where they will be needed, cutting empty travel and improving availability. The authors demonstrate that the selected temporal structures capture short-term persistence as well as daily and weekly cycles, yielding clear gains over baselines on both next-hour and next-24-hour tasks.

Core claim

Using large-scale e-scooter data from Austin, Texas, the authors convert trip records into hourly pickup and dropoff demand images on a grid and apply a combined correlation- and error-based procedure to identify informative historical inputs. An ablation study on a baseline UNET model with paired non-parametric tests and Holm correction determines the optimal temporal depth. The resulting structures reduce mean squared error by up to 37 percent for next-hour prediction and 35 percent for next-24-hour prediction compared with adjacent-hour and fixed-period baselines.

What carries the argument

The grid-based spatiotemporal demand images created from trip records, together with the correlation-and-error procedure and ablation study that selects optimal historical time steps for the UNET model.

If this is right

  • The selected temporal structures capture short-term persistence together with daily and weekly cycles.
  • The global activity mask limits evaluation to historically active areas while preserving demand patterns.
  • The reproducible pipeline supports consistent spatial learning across different prediction horizons.
  • Principled dataset construction improves both next-hour and next-24-hour forecast accuracy over heuristic baselines.

Where Pith is reading between the lines

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

  • Repeating the selection process for each new city or model family may be necessary rather than transferring the Austin-derived inputs directly.
  • The same grid-plus-statistical-input approach could be tested on bike-share or car-share data to check whether similar cycle patterns appear.
  • Extending the framework to include weather or event data as additional channels in the grid images might further reduce error.

Load-bearing premise

The temporal structures identified by the ablation study on a single UNET model in Austin will remain optimal when the model architecture, city, or data period changes.

What would settle it

Re-running the full ablation study and statistical tests on a different model architecture or on e-scooter data from another city and finding that a different set of historical depths performs best would show the selected inputs are not generally optimal.

Figures

Figures reproduced from arXiv: 2603.13609 by Merkebe Getachew Demissie, Mohammad Sahnoon, Roberto Souza.

Figure 1
Figure 1. Figure 1: Trip counts per vehicle type in Austin’s raw micromobility dataset [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Spatial coverage of the final 2019 e-scooter dataset. The figure [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of core trip characteristics for the processed 2019 e-scooter dataset: (a) trip duration in minutes, (b) trip distance in kilometers, and (c) [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Average daily e-scooter demand in 2019, obtained by averaging daily [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Average hourly e-scooter demand in 2019, calculated by averaging [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Spatial distribution of (a) trip origin and (b) trip destination intensities for Austin’s 2019 e-scooter dataset at the Census Tract level, highlighting [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Generation of the 240 × 220 square meters rectangular grid over the [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Sample pick-up demand image generated for 4:00 p.m. on March 17, 2019. Subplot (a) uses pixel-based indexing of the raster grid, subplot (b) [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Global binary activity mask used to define the operational footprint of the historically active e-scooter demand locations to guide model training and [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Channel-based input representation for spatiotemporal demand prediction. The figure illustrates the pick-up demand case, in which the model input [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
read the original abstract

Despite progress in deep learning for shared micromobility demand prediction, the systematic design and statistical validation of temporal input structures remain underexplored. Temporal features are often selected heuristically, even though historical demand strongly affects model performance and generalizability. This paper introduces a reproducible data-processing pipeline and a statistically grounded method for designing temporal input structures for image-to-image demand prediction. Using large-scale e-scooter data from Austin, Texas, we build a grid-based spatiotemporal dataset by converting trip records into hourly pickup and dropoff demand images. The pipeline includes trip filtering, mapping Census Tracts to spatial locations, grid construction, demand aggregation, and creation of a global activity mask that limits evaluation to historically active areas. This representation supports consistent spatial learning while preserving demand patterns. We then introduce a combined correlation- and error-based procedure to identify informative historical inputs. Optimal temporal depth is selected through an ablation study using a baseline UNET model with paired non-parametric tests and Holm correction. The resulting temporal structures capture short-term persistence as well as daily and weekly cycles. Compared with adjacent-hour and fixed-period baselines, the proposed design reduces mean squared error by up to 37 percent for next-hour prediction and 35 percent for next-24-hour prediction. These results highlight the value of principled dataset construction and statistically validated temporal input design for spatiotemporal micromobility demand prediction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces a grid-based pipeline to convert e-scooter trip records from Austin, Texas into hourly spatiotemporal demand images on a Census-tract-derived grid, together with a correlation-plus-error ablation procedure that selects optimal historical temporal depths for a baseline UNet model. The selected inputs are validated with paired non-parametric tests and Holm correction; the resulting structures are reported to reduce MSE by up to 37 % for next-hour and 35 % for next-24-hour prediction relative to adjacent-hour and fixed-period baselines.

Significance. If the temporal-input gains prove robust, the work supplies a reproducible data-construction pipeline and a statistically grounded alternative to heuristic lag selection for spatiotemporal micromobility forecasting. The explicit use of non-parametric tests with multiple-comparison correction is a methodological strength that could be adopted more widely.

major comments (2)
  1. [Ablation study and experimental results] The ablation that identifies the optimal temporal depths is performed exclusively on a single baseline UNet architecture. Because the procedure measures performance improvement on the error surface of that specific model, it is unclear whether the same lag structures would be selected or would produce comparable MSE reductions for architectures with different receptive fields, skip-connection patterns, or loss landscapes. This makes the headline performance claims (37 % and 35 % reductions) potentially architecture-specific rather than a general property of the proposed input design.
  2. [Results and experimental setup] The manuscript provides no tables reporting exact dataset sizes (number of trips, hours, or grid cells), model parameter counts, standard deviations or error bars on the MSE values, or the full set of ablation outcomes. Without these details the quantitative claims cannot be fully assessed and the risk of post-hoc selection effects cannot be ruled out.
minor comments (1)
  1. [Data representation pipeline] Clarify the exact spatial resolution of the grid (cell size in meters) and the construction of the global activity mask in the data-processing pipeline description.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. We address each major point below and outline the revisions we will make to improve transparency and address concerns about generalizability.

read point-by-point responses

  1. Referee: [Ablation study and experimental results] The ablation that identifies the optimal temporal depths is performed exclusively on a single baseline UNet architecture. Because the procedure measures performance improvement on the error surface of that specific model, it is unclear whether the same lag structures would be selected or would produce comparable MSE reductions for architectures with different receptive fields, skip-connection patterns, or loss landscapes. This makes the headline performance claims (37 % and 35 % reductions) potentially architecture-specific rather than a general property of the proposed input design.

    Authors: We selected UNet as a standard and widely adopted baseline for image-to-image spatiotemporal prediction tasks precisely because of its established effectiveness in capturing spatial dependencies via skip connections. The correlation-plus-error ablation procedure itself is model-agnostic: it ranks candidate temporal depths using statistical correlation with the target and empirical error reduction, independent of any particular network's receptive field or loss landscape. The reported MSE reductions therefore demonstrate the value of the selected inputs for this representative architecture and task. We nevertheless acknowledge that explicit verification on additional architectures would further support broader applicability. In the revision we will add a dedicated subsection clarifying the model-agnostic nature of the input-selection method, include a brief discussion of why UNet serves as an appropriate baseline, and report preliminary results on a second architecture (e.g., a ConvLSTM-based encoder-decoder) to illustrate that the same lag structures yield comparable gains. revision: partial

  2. Referee: [Results and experimental setup] The manuscript provides no tables reporting exact dataset sizes (number of trips, hours, or grid cells), model parameter counts, standard deviations or error bars on the MSE values, or the full set of ablation outcomes. Without these details the quantitative claims cannot be fully assessed and the risk of post-hoc selection effects cannot be ruled out.

    Authors: We agree that these details are essential for reproducibility and for allowing readers to evaluate the strength of the quantitative claims. In the revised manuscript we will add a new table (or expanded section) that reports: (i) exact dataset statistics including total trips, number of hourly time steps, grid dimensions, and the size of the active-area mask; (ii) model parameter counts for the UNet; (iii) mean MSE values accompanied by standard deviations across the test folds or repeated runs; and (iv) the complete ablation table showing MSE for every candidate temporal depth together with the corresponding p-values before and after Holm correction. These additions will be placed in the main text or as a supplementary table to eliminate any ambiguity regarding post-hoc selection. revision: yes

Circularity Check

1 steps flagged

Ablation-based selection of temporal inputs on the same UNET model used for evaluation introduces moderate circularity in MSE reduction claims

specific steps
  1. fitted input called prediction [Abstract]
    "Optimal temporal depth is selected through an ablation study using a baseline UNET model with paired non-parametric tests and Holm correction. The resulting temporal structures capture short-term persistence as well as daily and weekly cycles. Compared with adjacent-hour and fixed-period baselines, the proposed design reduces mean squared error by up to 37 percent for next-hour prediction and 35 percent for next-24-hour prediction."

    The temporal input configuration is identified by measuring and minimizing error on the identical UNET architecture and dataset later used to compute the reported MSE reductions. Because the 'proposed design' is the winner of that ablation, the claimed percentage improvements are partly guaranteed by the selection process itself rather than arising from an independent test of a pre-specified input structure.

full rationale

The paper's headline result (up to 37% and 35% MSE reduction) is obtained after first running an ablation study on a baseline UNET to select the optimal temporal depth and structures. This selection directly optimizes the inputs against the error metric later used to quantify improvement versus fixed baselines. While the procedure includes statistical tests and the baselines are distinct heuristics, the central performance numbers are measured on the configuration that was chosen to minimize error on that exact model and data, creating a moderate fitted-input issue. No self-citations, definitional loops, or imported uniqueness theorems appear in the provided text; the derivation remains largely empirical and self-contained once the selection step is acknowledged.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that historical demand patterns are stable enough to be captured by a fixed set of past hours and on the modeling choice that a standard UNET suffices to demonstrate the value of the input design.

free parameters (1)
  • selected temporal depth
    Number of historical hours retained after the ablation study; chosen to minimize error on the target prediction task.
axioms (1)
  • domain assumption Historical demand strongly affects model performance and generalizability.
    Invoked in the opening paragraph as the motivation for systematic temporal input design.

pith-pipeline@v0.9.0 · 5787 in / 1460 out tokens · 80844 ms · 2026-05-21T11:12:20.607963+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages

  1. [1]

    The benefits of carpooling,

    S. Shaheen, A. Cohen, and A. Bayen, “The benefits of carpooling,” UC Berkeley: Transportation Sustainability Research Center, 10 2024. [Online]. Available: https://escholarship.org/uc/item/7jx6z631

    work page 2024

  2. [2]

    Utilization rate of the fleet: a novel performance metric for a novel shared mobility,

    A. B. R. Gonz ´alez, M. R. Wilby, J. J. V . D ´ıaz, R. F. Pozo, and C. S. ´Avila, “Utilization rate of the fleet: a novel performance metric for a novel shared mobility,”Transportation, vol. 50, pp. 285–301, 2023. [Online]. Available: https://doi.org/10.1007/s11116-021-10244-x

    work page doi:10.1007/s11116-021-10244-x 2023

  3. [3]

    A systematic literature review on machine learning in shared mobility,

    J. Teusch, J. N. Gremmel, C. Koetsier, F. T. Johora, M. Sester, D. M. Woisetschl¨ager, and J. P. M ¨uller, “A systematic literature review on machine learning in shared mobility,”IEEE Open Journal of Intelligent Transportation Systems, vol. 4, pp. 870–899, 2023

    work page 2023

  4. [4]

    Equity implications of emerging mobility services and public transit coopetition: A review,

    A. D. Beza, M. G. Demissie, and L. Kattan, “Equity implications of emerging mobility services and public transit coopetition: A review,” Transportation Research Part D: Transport and Environment, vol. 144, p. 104751, 2025. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S1361920925001610

    work page 2025

  5. [5]

    On the inefficiency of ride-sourcing services towards urban congestion,

    C. V . Beojone and N. Geroliminis, “On the inefficiency of ride-sourcing services towards urban congestion,”Transportation Research Part C: Emerging Technologies, vol. 124, p. 102890, 2021. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0968090X20307907

    work page 2021

  6. [6]

    Are e-scooters polluters? the environmental impacts of shared dockless electric scooters,

    J. Hollingsworth, B. Copeland, and J. X. Johnson, “Are e-scooters polluters? the environmental impacts of shared dockless electric scooters,”Environmental Research Letters, vol. 14, p. 084031, 2019. [Online]. Available: https://doi.org/10.1088/1748-9326/ab2da8

    work page doi:10.1088/1748-9326/ab2da8 2019

  7. [7]

    Predicting demand for shared e-scooter using community structure and deep learning method,

    S. Kim, S. Choo, G. Lee, and S. Kim, “Predicting demand for shared e-scooter using community structure and deep learning method,”Sustainability, vol. 14, 2022. [Online]. Available: https: //www.mdpi.com/2071-1050/14/5/2564

    work page 2022

  8. [8]

    You’ll never share alone: Analyzing carsharing user group behavior,

    F. Baumgarte, T. Brandt, R. Keller, F. R ¨ohrich, and L. Schmidt, “You’ll never share alone: Analyzing carsharing user group behavior,” Transportation Research Part D: Transport and Environment, vol. 93, p. 102754, 2021. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S1361920921000584

    work page 2021

  9. [9]

    Spatiotemporal demand prediction model for e-scooter sharing services with latent feature and deep learning,

    S. W. Ham, J.-H. Cho, S. Park, and D.-K. Kim, “Spatiotemporal demand prediction model for e-scooter sharing services with latent feature and deep learning,”Transportation Research Record, vol. 2675, pp. 34–43,

    work page

  10. [10]

    Available: https://doi.org/10.1177/03611981211003896

    [Online]. Available: https://doi.org/10.1177/03611981211003896

    work page doi:10.1177/03611981211003896

  11. [11]

    Real-time forecasting of dockless scooter-sharing demand: A spatio-temporal multi-graph transformer approach,

    Y . Xu, X. Zhao, X. Zhang, and M. Paliwal, “Real-time forecasting of dockless scooter-sharing demand: A spatio-temporal multi-graph transformer approach,”IEEE Transactions on Intelligent Transportation Systems, vol. 24, pp. 8507–8518, 2023

    work page 2023

  12. [12]

    Sparse trip demand prediction for shared e-scooter using spatio-temporal graph neural networks,

    J.-C. Song, I.-Y . L. Hsieh, and C.-S. Chen, “Sparse trip demand prediction for shared e-scooter using spatio-temporal graph neural networks,”Transportation Research Part D: Transport and Environment, vol. 125, p. 103962, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1361920923003590

    work page 2023

  13. [13]

    Unet and unetr based frameworks for predicting the short-term spatiotemporal demand of e-scooter sharing services,

    M. Sahnoon, A. Manuel, M. G. Demissie, and R. Souza, “Unet and unetr based frameworks for predicting the short-term spatiotemporal demand of e-scooter sharing services,” in2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), 2024, pp. 2804–2811

    work page 2024

  14. [14]

    U-net: Convolutional networks for biomedical image segmentation,

    Philipp, B. T. R. Olaf, and Fischer, “U-net: Convolutional networks for biomedical image segmentation,” inMedical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Joachim, W. W. M., F. A. F. N. Nassir, and Hornegger, Eds. Springer International Publishing, 2015, pp. 234–241

    work page 2015

  15. [15]

    Unetr: Transformers for 3d medical image segmentation,

    A. Hatamizadeh, Y . Tang, V . Nath, D. Yang, A. Myronenko, B. Land- man, H. R. Roth, and D. Xu, “Unetr: Transformers for 3d medical image segmentation,” in2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022, pp. 1748–1758

    work page 2022

  16. [16]

    Predicting spatiotemporal demand of dockless e-scooter sharing services with a masked fully convolutional network,

    S. Phithakkitnukooon, K. Patanukhom, and M. G. Demissie, “Predicting spatiotemporal demand of dockless e-scooter sharing services with a masked fully convolutional network,”ISPRS International Journal of Geo-Information, vol. 10, 2021. [Online]. Available: https: //www.mdpi.com/2220-9964/10/11/773

    work page 2021

  17. [17]

    Study on the relationship between the spatial distribution of shared bicycle travel demand and urban built environment,

    L. Yang, S. Fei, H. Jia, J. Qi, L. Wang, and X. Hu, “Study on the relationship between the spatial distribution of shared bicycle travel demand and urban built environment,”Sustainability, vol. 15, 2023. [Online]. Available: https://www.mdpi.com/2071-1050/15/18/13576

    work page 2023

  18. [18]

    Icn: Interactive convolutional network for forecasting travel demand of shared micromobility,

    Y . Xu, Q. Ke, X. Zhang, and X. Zhao, “Icn: Interactive convolutional network for forecasting travel demand of shared micromobility,” GeoInformatica, vol. 29, pp. 175–200, 2025. [Online]. Available: https://doi.org/10.1007/s10707-024-00525-9

    work page doi:10.1007/s10707-024-00525-9 2025

  19. [19]

    Estimating censored spatial- temporal demand with applications to shared micromobility,

    A. Paul, K. Flynn, and C. Overney, “Estimating censored spatial- temporal demand with applications to shared micromobility,”arXiv e- prints, p. arXiv:2303.09971, 3 2023

    work page arXiv 2023

  20. [20]

    Statistical comparisons of classifiers over multiple data sets,

    J. Dem ˇsar, “Statistical comparisons of classifiers over multiple data sets,”Journal of Machine Learning Research, vol. 7, pp. 1–30, 12

    work page

  21. [21]

    Available: http://jmlr.org/papers/v7/demsar06a.html

    [Online]. Available: http://jmlr.org/papers/v7/demsar06a.html

    work page

  22. [22]

    2020 Decennial Census: Austin City, Texas,

    U.S. Census Bureau, “2020 Decennial Census: Austin City, Texas,” 2020, accessed: Feb. 24, 2026. [Online]. Available: https://data.census. gov/

    work page 2020

  23. [23]

    American Community Survey 2024 1-Year Estimates: Austin City, Texas,

    U.S. Census Bureau, “American Community Survey 2024 1-Year Estimates: Austin City, Texas,” 2024, accessed: Feb. 24, 2026. [Online]. Available: https://data.census.gov/

    work page 2024

  24. [24]

    Dockless electric scooter-related injuries study,

    Austin Public Health, “Dockless electric scooter-related injuries study,” Epidemiology and Public Health Preparedness Division, Tech. Rep., 4 2019. [Online]. Available: https://www.austintexas.gov/sites/default/ files/files/Health/Epidemiology/Dockless Electric Scooter-Related Injury Study final version EDSU 5.9.19.docx.pdf

    work page 2019

  25. [25]

    Shared mobility services,

    City of Austin, “Shared mobility services,” accessed: Mar. 2, 2026. [Online]. Available: https://www.austintexas.gov/sharedmobility

    work page 2026

  26. [26]

    Austin-micromobility-dataset-processing- pipeline,

    M. Sahnoon, “Austin-micromobility-dataset-processing- pipeline,” 3 2026, accessed: Mar. 12, 2026. [Online]. Available: https://github.com/mohammadsahnoon/ Austin-Micromobility-Dataset-Processing-Pipeline.git

    work page 2026

  27. [27]

    Shared micromobility vehicle trips (2018-2022),

    City of Austin Open Data Portal, “Shared micromobility vehicle trips (2018-2022),” accessed: Mar. 2, 2026. [Online]. Available: https://data.austintexas.gov/Transportation-and-Mobility/ Shared-Micromobility-Vehicle-Trips-2018-2022-/7d8e-dm7r ARXIV PREPRINT 16

    work page 2018

  28. [28]

    2010 tiger/line shapefiles: Census tracts,

    US Census Bureau, “2010 tiger/line shapefiles: Census tracts,” accessed: Mar. 3, 2026. [Online]. Available: https://www.census.gov/cgi-bin/geo/ shapefiles/index.php?year=2010&layergroup=Census+Tracts

    work page 2010

  29. [29]

    Boundaries jurisdictions,

    City of Austin Open Data Portal, “Boundaries jurisdictions,” accessed: Mar. 3, 2026. [Online]. Available: https://data.austintexas.gov/ City-Government/BOUNDARIES jurisdictions/vnwj-xmz9/about data

    work page 2026

  30. [30]

    Pedestrian walking speeds and conflicts at urban median locations,

    B. L. Bowman and R. L. Vecellio, “Pedestrian walking speeds and conflicts at urban median locations,”Transportation Research Record, pp. 67–73, 1994. [Online]. Available: https://api.semanticscholar.org/ CorpusID:8235028

    work page 1994

  31. [31]

    D. C. Montgomery, E. A. Peck, and G. G. Vining,Introduction to linear regression analysis, 5th ed. John Wiley & Sons, Inc.: Hoboken, 2012

    work page 2012

  32. [32]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” inProceedings of the 32nd International Conference on Machine Learning, F. Bach and D. Blei, Eds., vol. 37. PMLR, 3 2015, pp. 448–456. [Online]. Available: https://proceedings.mlr.press/v37/ioffe15.html

    work page 2015

  33. [33]

    rmsprop: Divide the gradient by a running average of its recent magnitude,

    G. Hinton, “rmsprop: Divide the gradient by a running average of its recent magnitude,” pp. 26–31, 2012. [Online]. Available: https: //www.cs.toronto.edu/∼tijmen/csc321/slides/lecture slides lec6.pdf

    work page 2012

  34. [34]

    An analysis of variance test for normality (complete samples)†,

    S. S. SHAPIRO and M. B. WILK, “An analysis of variance test for normality (complete samples)†,”Biometrika, vol. 52, pp. 591–611, 12

    work page

  35. [35]

    Available: https://doi.org/10.1093/biomet/52.3-4.591

    [Online]. Available: https://doi.org/10.1093/biomet/52.3-4.591

    work page doi:10.1093/biomet/52.3-4.591

  36. [36]

    Individual comparisons by ranking methods,

    F. Wilcoxon, “Individual comparisons by ranking methods,”Biometrics Bulletin, vol. 1, pp. 80–83, 1945. [Online]. Available: http://www.jstor. org/stable/3001968

    work page arXiv 1945

  37. [37]

    A simple sequentially rejective multiple test procedur e

    S. Holm, “A simple sequentially rejective multiple test procedure,” Scandinavian Journal of Statistics, vol. 6, pp. 65–70, 1979. [Online]. Available: http://www.jstor.org/stable/4615733

    work page arXiv 1979

  38. [38]

    Free-floating bike-sharing demand prediction with deep learning,

    Z. Zhang, L. Tan, and W. Jiang, “Free-floating bike-sharing demand prediction with deep learning,”International Journal of Machine Learning and Computing, 2022. [Online]. Available: https: //api.semanticscholar.org/CorpusID:247828628

    work page 2022

  39. [39]

    Improving short-term bike sharing demand forecast through an irregular convolutional neural network,

    X. Li, Y . Xu, X. Zhang, W. Shi, Y . Yue, and Q. Li, “Improving short-term bike sharing demand forecast through an irregular convolutional neural network,”Transportation Research Part C: Emerging Technologies, vol. 147, p. 103984, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0968090X22003977

    work page 2023

  40. [40]

    Predicting short-term bike-sharing demand at station level: A multi-task dynamic graph-based spatiotemporal approach,

    S. Nejadshamsi, J. Bentahar, C. Wang, and U. Eicker, “Predicting short-term bike-sharing demand at station level: A multi-task dynamic graph-based spatiotemporal approach,”Knowledge-Based Systems, vol. 333, p. 114986, 2026. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S0950705125020246

    work page 2026

  41. [41]

    Machine learning models for bike-sharing demand forecasting,

    D. Hosseinpanahi, P. Zadtootaghaj, J. Lin, A. K. Mohammadian, and B. Zou, “Machine learning models for bike-sharing demand forecasting,”Future Transportation, vol. 6, 2026. [Online]. Available: https://www.mdpi.com/2673-7590/6/1/26

    work page 2026