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

Jiachen Wang; Song Han; Suining He; Xi Yang

REVIEW 5 major objections 7 minor 1 cited by

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

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that station-level bike traffic across an entire city system can be predicted with one attention-based encoder-decoder model, cutting next-hour forecast RMSE and MAE by over 40% relative to LSTM encoder-decoders.

desk verdict Applies a known attention architecture to NYC bike sharing with a useful data analysis, but the 40% improvement claim lacks statistical support and the architecture is not novel. read the letter →

arxiv 2507.16020 v1 pith:EQBFBIQE submitted 2025-07-21 cs.AI

Xi Yang , Jiachen Wang , Song Han , Suining He This is my paper

classification cs.AI

keywords bikesharingstation-leveldemandpredictionspatio-temporalattentionencoder-decoderLSTMmicromobilityNewYorkCityCitipointsofinterest

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes BikeMAN, a single neural network trained to forecast next-hour bike pick-ups and drop-offs for every station in a bike-sharing system at once, using two stacked attention mechanisms on top of an LSTM encoder-decoder. The claim is that this multi-level attention makes whole-city station-level prediction practical: on over ten million New York City trips across 766 stations, BikeMAN's RMSE and MAE are more than 40% lower than the plain LSTM encoder-decoder baseline for both demand and returns. The authors also show the gain grows with the number of stations included, suggesting the spatial attention layer is what carries the improvement. If true, the result matters because station-level demand imbalance is the main operational problem in docked bike sharing, and accurate full-system forecasts would let operators rebalance proactively.

What carries the argument

The load-bearing object is the multi-level attention pair inside an LSTM encoder-decoder. The spatial attention layer flattens the 766-station feature set into a vector, scores every feature against the encoder's previous hidden and cell states, and reweights the input with a softmax; it is meant to capture correlations both within a station's features and across stations. The temporal attention layer then scores decoder hidden states against encoder hidden states and takes a weighted sum, letting the decoder align its prediction with the most relevant past hours. Their interaction is what the paper credits for the accuracy gain, and removing it (the LSTM decoder-encoder baseline) is the main comparison.

What would settle it

Compute BikeMAN's and the LSTM encoder-decoder's per-station RMSE and MAE on the October 2019 test set and compare them within the decile of stations with the highest demand; if the over-40% error reduction shrinks or disappears for those stations, the paper's full-system accuracy claim is not established for the stations that matter most for rebalancing.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a spatial attention mechanism operating over the flattened feature vectors of all stations simultaneously, combined with a temporal attention mechanism over encoder hidden states in the decoder, gives materially more accurate station-level pick-up and return forecasts than the same encoder-decoder without attention. Tested on 766 New York City stations with June-August 2019 training and October 2019 testing, BikeMAN reports RMSE 3.366 and MAE 1.818 for demand and RMSE 3.369 and MAE 1.797 for returns, versus RMSE 5.678 and MAE 3.350 for the LSTM encoder-decoder baseline on demand; that is the over-40% drop the paper highlights. The same model on a single busy station (519) shows almost no gain, which the authors interpret as evidence that the spatial attention derives its power from pooling information across the whole station network.

Load-bearing premise

The central claim stands on the assumption that aggregate RMSE and MAE over all 766 stations is the right measure of prediction quality; the paper reports no per-station or demand-weighted error breakdown, and on the single busiest station BikeMAN's RMSE is 11.70, so the headline gain may not hold where rebalancing pressure is highest.

Editorial extensions

If this is right

A single trained model can produce next-hour pick-up and drop-off forecasts for all 766 stations of a large city bike-sharing system, removing the need to fit separate per-station predictors.
Forecast error should keep shrinking as more stations are included in the same model, since the spatial attention mechanism exploits cross-station feature correlations; the reported RMSE falls from 3.86 on 190 stations to 3.79 on 766 stations.
Demand and return prediction are nearly equally accurate (RMSE 3.366 vs 3.369), so the same architecture serves both sides of the rebalancing problem without a separate model.
The attention mechanism, not the recurrent cell choice, is the dominant contributor: swapping LSTM for GRU while keeping attention only slightly degrades results, while removing attention raises RMSE to about 70% higher than BikeMAN's.

Reading between the lines

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

The paper does not report per-station error distributions or demand-weighted metrics; because the 766-station average is dominated by many low-traffic stations, a testable extension is to check whether the over-40% gain survives on the busiest decile of stations, where rebalancing errors are most consequential.
A direct ablation extension would retrain BikeMAN with only the temporal attention and only the spatial attention to attribute the gain; the single-station result suggests the spatial layer carries most of it, but the paper does not run this ablation.
If the mechanism generalizes, the same architecture should transfer to other docked micromobility fleets or to finer forecast horizons such as 15-minute or 30-minute intervals, with the caveat that the model must be retrained for each city's station set and feature layout.

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.

Desk Editor's Note

read the letter

Colleague,

The paper is a competent application of the GeoMAN attention encoder-decoder to station-level bike sharing prediction. What's genuinely useful is the data analysis: the weather and POI correlations are clear, and the experiment covers 766 stations over five months with a large trip dataset. The authors are also honest that the attention mechanism gives no benefit for a single station (RMSE 11.70 vs. 11.78 at station 519), and they locate the gain in multi-station prediction.

The soft spots are real but not fatal. The architecture is essentially GeoMAN with a different application; Eqs. (3)-(8) match the prior work, so the 'first novel multi-level attention' claim is overstated. More importantly, the evaluation does not back the headline number. The 40% improvement over LSTM/GRU encoder-decoders rests on a single test month, no error bars, no significance tests, and only two baselines that are stripped versions of the same encoder-decoder. No ARIMA, GBDT, or graph-convolutional comparisons are included. The aggregate RMSE over 766 stations is likely dominated by low-demand stations; station 519, the busiest, shows no improvement, so the practical value for rebalancing high-demand hubs is not established.

The preprocessing choices (precipitation normalized to [0,10], coordinates to [0,100]) are also ad hoc and not justified, which weakens the claim that the model generalizes.

That said, the paper is clearly written, the experiments are reproducible in principle, and the authors do not hide the single-station result. It is a workshop-level empirical study, not a methodological breakthrough.

Who gets value? Researchers working on station-level micromobility prediction who want a sanity check for their own models and a description of a clean feature set. I would not cite the 40% figure in my own work, but I might cite the data analysis.

If I were an editor, I would send this to peer review with the expectation of major revision: add variance/significance, stronger baselines, demand-weighted metrics, and soften the novelty framing. As a colleague, I'd engage with it lightly—worth a skim, not a deep read.

Referee Report

5 major / 7 minor

Summary. The paper proposes BikeMAN, an encoder-decoder neural network with two attention mechanisms (spatial and temporal) for next-hour station-level bike pickup and drop-off prediction across all 766 NYC Citi Bike stations. Input features include historical traffic counts, weather (temperature, precipitation, wind speed), station coordinates, and counts of nearby POIs. The model is trained on June-August 2019 and tested on October 2019. The reported experiments compare BikeMAN with LSTM/GRU encoder-decoder baselines and claim an over-40% reduction in RMSE/MAE for both demand and return predictions (Table 2).

Significance. If the reported results are robust, the attention-based architecture would be a useful building block for system-wide micromobility forecasting, and the use of all 766 stations rather than a selected subset is a strength. The paper provides a real-world evaluation on over 10 million trips, and the ablation against no-attention encoder-decoders gives some evidence that the attention mechanisms contribute to accuracy. However, the evaluation is too narrow to establish the headline claim: a single test month, no variance or significance estimates, aggregate metrics that may be dominated by low-demand stations, and only self-implemented baselines. The significance is therefore moderate and conditional on strengthening the evaluation.

major comments (5)

[Section 5.2.2, Table 2] The central claim that 'the RMSE and MAE of BikeMAN drop by over 40% from LSTM Dec-Enc' rests on a single evaluation on October 2019 with one number per metric and no standard deviation, confidence interval, or significance test. Because the model is stochastic (random initialization, dropout, batching), the reported differences could be within run-to-run variation. The authors should report mean +/- std over multiple seeds and either evaluate additional test periods or use a paired statistical test (e.g., Diebold-Mariano or bootstrap over stations/hours).
[Section 5.2.1, Table 1] The aggregate RMSE/MAE over 766 stations does not show the per-station error distribution. Table 1 reports RMSE 11.70 for the busiest station (519) versus 3.79 averaged over 766 stations, so the 40% improvement in Table 2 may be concentrated in low-demand stations where absolute errors are small. Since the motivating application is rebalancing high-demand stations, please report demand-weighted errors, per-station error percentiles, or a stratification by station demand.
[Section 5.2, Tables 1-2] The only baselines are encoder-decoder models with the same architecture minus attention. While this is a valid ablation, it does not support the broader claim of 'high accuracy' relative to existing approaches. The paper should include at least one or two standard baselines (e.g., historical average, ARIMA, gradient-boosted trees, or a graph convolutional method such as [10]) to calibrate the improvement.
[Section 4.2.1, Eqs. (3)-(5)] The spatial attention notation is inconsistent and the mechanism is underspecified. The text defines f^k_{t-T:t} as the k-th feature of the flattened previous T timestamps, but Eq. (5) applies the attention weight to f^k_t, the current single-timestamp feature. Please clarify whether the score in Eq. (3) uses the full length-T series or a single value, and specify the dimensions of U_s, W_s, and v_s. This is needed for reproducibility.
[Section 4.2.1] The proposed 'spatial attention' computes a scalar weight per feature dimension and reweights each feature independently; it does not compute any pairwise or relational interaction between stations. The claim that it 'captures the correlation ... across all stations in the city' is therefore stronger than what Eqs. (3)-(5) implement. Please either revise the interpretation or add an explicit interaction term (e.g., a graph or covariance-based attention) if cross-station correlations are intended.

minor comments (7)

[Section 1] The statement 'we propose the first novel multi-level attention neural network' is difficult to reconcile with the cited GeoMAN work [9], which is itself a multi-level attention network; please qualify the novelty claim.
[Section 5.1.1] The choice to exclude September 2019 from both training and testing is not explained; please state the reason.
[Eq. (12)] The MAE formula is missing absolute value bars; it should be (1/M) * sum |y_i - yhat_i|.
[Eq. (10)] The loss is written as sqrt((y - yhat)^2), which is a per-sample absolute error, not the mean squared error minimized by Adam; please clarify the training objective.
[Section 5.1.1] The justification for the 150 m POI radius is not fully convincing; the text argues the radius must be small to avoid including another station, but then chooses 150 m because 91.25% of stations are more than 150 m apart. Please clarify the trade-off.
[Figures 6 and 7] The captions mention a 'red line' for ground truth, but the text says BikeMAN 'highly resembles the ground-truths'; please make the legend explicit and ensure the figures are legible in print.
[Tables 1 and 2] The LSTM encoder-decoder results in Table 1 (RMSE 5.64, MAE 3.26 for 766 stations) differ slightly from the LSTM Dec-Enc results in Table 2 (RMSE 5.678, MAE 3.350); please explain the source of the difference (e.g., different runs or settings).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: accuracy claims rest on a held-out October 2019 test set, and no fitted quantity is later renamed as a prediction.

full rationale

The paper's central claim is empirical: BikeMAN's RMSE/MAE on October 2019 held-out data is compared with baseline encoder-decoder models (Table 2). The model parameters are learned by minimizing MSE (Eq. 10) on June-August 2019 data, while the test month is disjoint and is not used in fitting. No parameter in the model is defined from the test target, and no reported prediction is a fitted input renamed as an output. The attention equations (Eqs. 3-9) define a feed-forward computation from input features and learned parameters, not an identity with the target. The paper contains no self-citations to prior work by the same authors and invokes no uniqueness theorem to force its architecture; the cited GeoMAN [9] is external related work. Concerns that the baselines are all internal variants and that the evaluation uses a single test month without error bars are correctness and robustness issues, not circularity under the specified definitions. Therefore no circular step can be exhibited.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central empirical claim rests on a fixed station set, hand-chosen normalization ranges, a 150 m POI radius, a 12-hour encoder window, and hyperparameters; the 19 features are not fully enumerated. No new physical or conceptual entities are introduced.

free parameters (6)

POI radius = 150 m
Chosen to include nearby POIs while avoiding most neighboring stations; Section 3.2.2.
Precipitation min-max normalization range = [0, 10]
Raw precipitation values are too small compared to trip counts; hand-rescaled in Section 5.1.1.
Coordinate min-max normalization range = [0, 100]
Raw lat/lon differences are about 1e-2; hand-rescaled in Section 5.1.1.
Encoder window length = 12 hours
Design choice for input history; Section 5.1.1.
Decoder horizon = 1 hour
Prediction target is next hour; Section 5.1.1.
Model hyperparameters = hidden 1024, layers 2, dropout 0.3, lr 0.001, batch 64, epochs 100
Reported in Section 5.1.2 without a tuning or sensitivity analysis.

assumptions (3)

domain assumption Bike traffic at each station can be predicted from the previous 12 hours of features of all stations plus weather and POIs.
This is the problem formulation in Sections 4.1 and 4.2.
domain assumption Spatial correlation between stations is captured by attention over the flattened feature vector of all stations, without explicit graph structure.
Assumed in the spatial attention design, Section 4.2.1.
domain assumption Stations appearing in all monthly datasets (766 stations) form a representative, unbiased subset.
Used when deleting non-common stations in Section 5.1.1.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Micromobility Flow Prediction: A Bike Sharing Station-level Study via Multi-level Spatial-Temporal Attention Neural Network." pith.science (2026). https://pith.science/paper/EQBFBIQE

@misc{pith2026250716020,
  author       = {Pith},
  title        = {Pith review of: Micromobility Flow Prediction: A Bike Sharing Station-level Study via Multi-level Spatial-Temporal Attention Neural Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EQBFBIQE}},
  note         = {Machine review of arXiv:2507.16020}
}

read the original abstract

Efficient use of urban micromobility resources such as bike sharing is challenging due to the unbalanced station-level demand and supply, which causes the maintenance of the bike sharing systems painstaking. Prior efforts have been made on accurate prediction of bike traffics, i.e., demand/pick-up and return/drop-off, to achieve system efficiency. However, bike station-level traffic prediction is difficult because of the spatial-temporal complexity of bike sharing systems. Moreover, such level of prediction over entire bike sharing systems is also challenging due to the large number of bike stations. To fill this gap, we propose BikeMAN, a multi-level spatio-temporal attention neural network to predict station-level bike traffic for entire bike sharing systems. The proposed network consists of an encoder and a decoder with an attention mechanism representing the spatial correlation between features of bike stations in the system and another attention mechanism describing the temporal characteristic of bike station traffic. Through experimental study on over 10 millions trips of bike sharing systems (> 700 stations) of New York City, our network showed high accuracy in predicting the bike station traffic of all stations in the city.

Figures

Figures reproduced from arXiv: 2507.16020 by the authors.

**Figure 2.** Bike demand and weather correlation in June, 2019. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Demand heatmap and distribution of commercial [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 5.** Figure 5: Core model architecture of BikeMAN. 4.2.1 Spatial Attention. The spatial attention mechanism is developed to describe the spatial correlation between the features of all stations. The intuition behind the spatial attention is that the traffic at each station is largel…

**Figure 4.** Figure 4: System architecture of BikeMAN. 4.1 Preliminaries We consider that for totally N stations in a city, each station i has s features at timestamp t : F𝑖,𝑡 = (𝐹𝑖1,𝑡, ..., 𝐹𝑖𝑗,𝑡, ..., 𝐹𝑖𝑠,𝑡), where 𝑖 ∈ [1, ..., 𝑁] and 𝑗 ∈ [1, ..., 𝑠], and let 𝐹 (𝑝) 𝑖1,𝑡 and 𝐹 (𝑑) 𝑖1,𝑡 be t…

**Figure 6.** Figure 6: Prediction of BikeMAN and BikeMAN-GRU on the bike demand of station 519 for the first week of October, 2019 (red line) compared with ground truth [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 7.** Figure 7: Prediction of BikeMAN and BikeMAN-GRU on the bike return of station 519 for the first week of October, 2019 (red line) compared with ground truth. 5 [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

STAGformer: A Spatio-temporal Agent Graph Transformer for Micro Mobility Demand Forecasting
cs.LG 2026-07 conditional novelty 4.0 of 10

STAGformer forecasts bike-station demand with linear-complexity spatial-temporal agent attention and reports lower RMSE/MAE than listed baselines on NYC and Chicago data.

Reference graph

Works this paper leans on

15 extracted references · 12 canonical work pages · cited by 1 Pith paper

[10]

Lei Lin, Zhengbing He, and Srinivas Peeta. 2018. Predicting station-level hourly demand in a large-scale bike-sharing network: A graph convolutional neural network approach. Transportation Research Part C: Emerging Technologies 97 (2018), 258–276

work page 2018
[1]

Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural ma- chine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473 (2014)

arXiv 2014
[2]

Dzmitry Bahdanau, Jan Chorowski, Dmitriy Serdyuk, Philemon Brakel, and Yoshua Bengio. 2016. End-to-end attention-based large vocabulary speech recog- nition. In Proc. IEEE ICASSP. IEEE, 4945–4949

work page 2016
[3]

Di Chai, Leye Wang, and Qiang Yang. 2018. Bike flow prediction with multi-graph convolutional networks. In Proc. ACM SIGSPATIAL. ACM, 397–400

work page 2018
[4]

Po-Chuan Chen, He-Yen Hsieh, Xanno Kharis Sigalingging, Yan-Ru Chen, and Jenq-Shiou Leu. 2017. Prediction of station level demand in a bike sharing system using recurrent neural networks. In Proc. IEEE VTC Spring . IEEE, 1–5

work page 2017
[5]

Jan K Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio. 2015. Attention-based models for speech recognition. In Proc. NIPS. 577–585

work page 2015
[6]

Pierre Hulot, Daniel Aloise, and Sanjay Dominik Jena. 2018. Towards Station- Level Demand Prediction for Effective Rebalancing in Bike-Sharing Systems. In Proc. ACM SIGKDD. ACM, 378–386

work page 2018
[7]

Zhong Ji, Kailin Xiong, Yanwei Pang, and Xuelong Li. 2019. Video summarization with attention-based encoder-decoder networks. IEEE Transactions on Circuits and Systems for Video Technology (2019)

work page 2019

Show all 15 references

[8]

Youru Li, Zhenfeng Zhu, Deqiang Kong, Meixiang Xu, and Yao Zhao. 2019. Learn- ing Heterogeneous Spatial-Temporal Representation for Bike-Sharing Demand Prediction. In Proc. AAAI, Vol. 33. 1004–1011

2019
[9]

Yuxuan Liang, Songyu Ke, Junbo Zhang, Xiuwen Yi, and Yu Zheng. 2018. Geo- MAN: Multi-level Attention Networks for Geo-sensory Time Series Prediction.. In Proc. IJCAI. 3428–3434

2018
[11]

Minh-Thang Luong, Hieu Pham, and Christopher D Manning. 2015. Effec- tive approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025 (2015)

2015 arXiv
[12]

Jingkuan Song, Zhao Guo, Lianli Gao, Wu Liu, Dongxiang Zhang, and Heng Tao Shen. 2017. Hierarchical LSTM with adjusted temporal attention for video captioning. arXiv preprint arXiv:1706.01231 (2017)

2017 arXiv
[13]

Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. 2017. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. In Proc. AAAI

2017
[14]

Shuai Wang, Tian He, Desheng Zhang, Yunhuai Liu, and Sang H Son. 2019. Towards Efficient Sharing: A Usage Balancing Mechanism for Bike Sharing Systems. In Proc. WWW. ACM, 2011–2021

2019
[15]

Albert Zeyer, Kazuki Irie, Ralf Schlüter, and Hermann Ney. 2018. Improved training of end-to-end attention models for speech recognition. arXiv preprint arXiv:1805.03294 (2018). 6

2018 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [10]

Lei Lin, Zhengbing He, and Srinivas Peeta. 2018. Predicting station-level hourly demand in a large-scale bike-sharing network: A graph convolutional neural network approach. Transportation Research Part C: Emerging Technologies 97 (2018), 258–276

work page 2018

[2] [1]

Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural ma- chine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473 (2014)

arXiv 2014

[3] [2]

Dzmitry Bahdanau, Jan Chorowski, Dmitriy Serdyuk, Philemon Brakel, and Yoshua Bengio. 2016. End-to-end attention-based large vocabulary speech recog- nition. In Proc. IEEE ICASSP. IEEE, 4945–4949

work page 2016

[4] [3]

Di Chai, Leye Wang, and Qiang Yang. 2018. Bike flow prediction with multi-graph convolutional networks. In Proc. ACM SIGSPATIAL. ACM, 397–400

work page 2018

[5] [4]

Po-Chuan Chen, He-Yen Hsieh, Xanno Kharis Sigalingging, Yan-Ru Chen, and Jenq-Shiou Leu. 2017. Prediction of station level demand in a bike sharing system using recurrent neural networks. In Proc. IEEE VTC Spring . IEEE, 1–5

work page 2017

[6] [5]

Jan K Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio. 2015. Attention-based models for speech recognition. In Proc. NIPS. 577–585

work page 2015

[7] [6]

Pierre Hulot, Daniel Aloise, and Sanjay Dominik Jena. 2018. Towards Station- Level Demand Prediction for Effective Rebalancing in Bike-Sharing Systems. In Proc. ACM SIGKDD. ACM, 378–386

work page 2018

[8] [7]

Zhong Ji, Kailin Xiong, Yanwei Pang, and Xuelong Li. 2019. Video summarization with attention-based encoder-decoder networks. IEEE Transactions on Circuits and Systems for Video Technology (2019)

work page 2019