Pith. sign in

REVIEW 5 major objections 7 minor 31 references

Spatio-Temporal Demand Prediction for Food Delivery Using Attention-Driven Graph Neural Networks

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 An attention-driven graph neural network predicts both food delivery demand per zone and the origin-destination flow of orders, and on a real Meituan dataset it reports lower MAPE and MAE than GEML, LSTNet, and AR baselines.

desk verdict Sensible application of standard attention-GNN machinery to food delivery demand prediction, but the empirical core is internally inconsistent: the headline numbers in Table 5 contradict the full-model rows in the paper's own ablation tables. read the letter →

arxiv 2507.15246 v1 pith:CVWOVR4N submitted 2025-07-21 cs.LG cs.AI

Rabia Latief Bhat , Iqra Altaf Gillani This is my paper · ORCID
classification cs.LGcs.AI
keywords fooddeliverydemandpredictionspatio-temporalforecastinggraphneuralnetworkattentionmechanismorigin-destinationurbanmobility
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 claims that modeling a city's food delivery zones as a directed graph, with edges weighted by historical order counts, lets a graph neural network with attention layers predict both next-slot demand per zone and the origin-destination matrix of orders. The attention mechanism is the key: it weights forward, backward, and geographically adjacent neighbors differently, and a temporal attention layer separates recurring meal-time patterns from short-term shocks. If the claim holds, platforms can position couriers, batch orders, and rebalance fleets based on where demand is headed, not just how much there is. The reported evidence is a single 312-hour Meituan dataset on which the model achieves the lowest errors across all reported horizons.

What carries the argument

The central object is a directed graph $G=(V,E,\Delta)$ whose nodes are 2.5 km grid cells and whose edge weights $\Delta_{ij}$ are historical order counts from cell $i$ to cell $j$. Three mechanisms carry the argument: a spatial attention layer that aggregates embeddings from forward, backward, and geographical neighbors with learned weights plus pre-weighted aggregators; a temporal attention layer with four channels (same hour across $N$ days, surrounding hours across previous days, and a recent $h$-hour context) fused by self-attention; and a transferring attention layer that computes destination probabilities $p_{ij}$ and forms the predicted OD matrix as $\hat{\delta}_i \cdot p_{ij}$. The prediction is blended with historical averages by a weighted aggregator.

What would settle it

Run the same architecture on a second city's food delivery data or on a multi-month span from a different platform, with hyperparameters fixed from the Meituan tuning; if the MAPE and MAE advantages over GEML and LSTNet shrink or reverse, the reported superiority was dataset-specific. A cheaper check: verify whether removing the linear recurrence layer degrades demand predictions as reported, since that ablation is the paper's main evidence for the layer's role.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a spatio-temporal attention GNN can forecast both total order volume and directional order flow between grid cells more accurately than existing demand predictors, with Table 5 reporting MAPE-0 0.1972, MAPE-3 0.1597, MAPE-5 0.1492, and MAE-0 25.7749 for the proposed model versus 1.0005, 0.3488, 0.3025, and 81.0424 for GEML and 0.2573, 0.1817, 0.1751, and 27.7420 for LSTNet. The model does this by exchanging node embeddings with three kinds of spatial neighbors and with temporal neighbors across previous days and recent hours, then using a transferring attention layer to distribute predicted demand across destinations.

Load-bearing premise

The claimed superiority rests on the assumption that a 312-hour slice of one platform's data, split 80/20 with hyperparameters (grid size, time slot, N=5 days, h=6 hours) chosen on that same data, is representative enough for the gains to generalize to other cities, platforms, or longer time spans.

Editorial extensions

If this is right

  • Platforms can use the model's OD predictions to reposition couriers toward zones that are about to receive orders, not just zones with high current volume.
  • The four-channel temporal attention gives an explicit decomposition: recurring meal-time demand comes from the same hour across past days, while short-term context from the past six hours captures event-driven surges.
  • The model is trained with Smooth L1 loss and reports MAPE-0, MAPE-3, and MAPE-5, so it can be evaluated separately on low- and high-demand regions.
  • The 15-minute time slot and 2.5 km grid size define a practical operating point that balances prediction accuracy against computational load.

Reading between the lines

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

  • Because the model outputs full OD matrices, it could be extended to multi-step lookahead dispatch and dynamic batching without architectural change, yet the paper stops at one-slot-ahead forecasts.
  • The geographic-neighbor fallback suggests the framework should handle sparse zones, but a 1725-cell grid over 312 hours likely contains many zero-flow cells; a targeted test on low-demand cells would show whether the fallback actually works.
  • The transfer-attention decomposition $\hat{\Delta}_{ij} = \hat{\delta}_i \cdot p_{ij}$ implies that demand and flow errors can be attributed separately, so a platform could calibrate the demand output without retraining the flow distribution.
  • Feeding weather or event features into the non-linear context channel is a natural testable extension that the paper names as future work.
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

5 major / 7 minor

Summary. The paper proposes an attention-driven graph neural network for forecasting food delivery demand and origin-destination (OD) flows. The service area is partitioned into grid cells; a time-indexed directed graph is built from an OD matrix; node embeddings are updated by spatial attention over forward, backward, and geographical neighbors; a four-channel temporal attention layer captures linear and non-linear dependencies; and a transferring attention layer outputs demand and OD predictions. Experiments use 312 hours of Meituan order data, with 80% training and 20% validation/test, and compare against AR, LSTNet, and GEML. The paper reports lower MAPE and MAE for demand prediction in Table 5 and discusses OD results through a figure.

Significance. If the empirical claims were reproducible, the paper would offer a practical contribution to food delivery operations: it explicitly models directed OD flows, uses three types of spatial neighbors, and includes a systematic ablation of temporal attention channels. The model description is reasonably complete, and the use of a real-world Meituan dataset is a positive feature. However, the paper does not ship code, data, or machine-checked proofs, and the central numerical claim is contradicted by the paper's own ablation tables. The significance therefore rests entirely on an internally inconsistent empirical evaluation.

major comments (5)
  1. [§5.5.4, Table 5 versus §5.5.3, Tables 3 and 4] The load-bearing comparison is not internally reproducible. Table 5 reports 'Our Model' with MAPE-0 0.1972, MAPE-3 0.1597, MAPE-5 0.1492, and MAE-0 25.7749, but the identical full configuration ('Demand Yes') in Tables 3 and 4 is reported with MAPE-0 0.2196, MAPE-3 0.1835, MAPE-5 0.1724, and MAE-0 32.6926. Since §5.1 and §5.2 describe a single 80/20 split and a single 312-hour dataset, and no protocol difference is stated, at least one set of numbers is wrong or refers to a different experimental protocol. This discrepancy is larger than the reported margin over AR (MAPE-0 0.2501 in Table 5), so the claimed superiority over the baselines cannot be assessed until the numbers are reconciled.
  2. [§4.2.1, Eqs. (7)-(14)] The core attention equations are underspecified. Eq. (7) defines AN as a two-argument function, but Eqs. (8)-(10) call AN with three arguments (e_i^t, alpha_i^t, e_j^t), and Eq. (15) uses two arguments. The pre-weighted aggregators in Eqs. (12)-(14) have inconsistent index sets: for example, the numerator Delta_ij in Eq. (12) is not tied to a well-defined sum over the forward neighbors of i, and the denominator uses F_j^t rather than F_i^t. Because these equations define the spatial and transferring attention mechanisms, the architecture cannot be implemented as written.
  3. [§4.2, in-degree and out-degree formulas] The in-degree and out-degree formulas are reversed relative to their definitions. The text states that in-degree is total incoming requests and out-degree is outgoing requests, but the displayed formulas give in-degree as sum_j OD(g_k, g_j), which is the row sum and therefore outgoing, and out-degree as sum_i OD(g_i, g_k), which is the column sum and therefore incoming. Since the node embedding explicitly includes in-degree and out-degree, this correction changes the model's input features.
  4. [§5.5.4, OD task] The paper claims that the proposed model outperforms GEML for OD prediction and discusses graph-based models in this context, but it provides no numerical OD comparison against any baseline. The text refers to Fig. 7, yet no table or error metric is given for the OD task. Given that OD modeling is stated as a central contribution, the absence of a quantitative baseline comparison leaves that claim unsupported.
  5. [§5.1-§5.5.3] All hyperparameters (grid size, time slot, N=5, h=6) and the final comparison are selected on the same 312-hour Meituan dataset, with a single 80/20 split and no error bars, multiple seeds, or significance tests. A 312-hour window is only about 13 days, which is too short to validate weekly or event-driven patterns, and the chosen settings are not tested on a separate validation set or a second dataset. The reported gains over AR and LSTNet should be accompanied by confidence intervals and an external validation protocol before a general superiority claim can be supported.
minor comments (7)
  1. [§4.2] The preprocessing description states '360 graphs per day', which is inconsistent with 15-minute time slots; 24 hours at 15-minute granularity corresponds to 96 graphs per day, so this number should be corrected.
  2. [§2 and §5.3] Baseline citations are inconsistent: LSTNet is cited as [27] in §5.3 but associated with [22] in §2, and GEML is cited as [21] in §5.3 but [26] in §2.
  3. [§5.5.3] The abbreviations STPP and STPM are introduced in the ablation section but are not defined in the architecture section; they should be named and defined in §4.2.2 if they refer to the hour-before and hour-after layers.
  4. [§4.2.2] The sentence 'higher values indicate lesser relevance' after defining the similarity matrix is confusing, because softmax-normalized attention weights conventionally assign higher weights to more relevant neighbors.
  5. [Eq. (16)] The denominator y_i+1 makes this a smoothed MAPE; please state this explicitly or use the standard definition to avoid ambiguity.
  6. [Eqs. (12)-(14) and §4.2.2] The symbol h is used both as the smoothing constant in Eqs. (12)-(14) and as the non-linear context window in §4.2.2; these two quantities should use distinct symbols.
  7. [General] No code or data release is mentioned; providing them would allow the reported numbers to be checked and would substantially strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the model is trained on historical data and evaluated on held-out data, and no load-bearing claim reduces by construction to its inputs.

full rationale

The paper's central claim is an empirical comparison: 'Our Model' achieves the lowest MAPE/MAE in Table 5 after training on 80% of the data and evaluation on the held-out 20% ('We used 80% of our data to train the model and set aside the remaining 20% for validation and testing'). The model inputs are historical OD matrices, embeddings, and graph structure; the outputs are future demand and OD predictions. No equation in Section 4 defines a predicted quantity as equal to an input, a fitted parameter, or a historical average in a way that would force the reported result. The pre-weighted aggregator and attention weights are learned functions of past observations, not of the target values. Hyperparameter choices such as grid size, time-slot duration, N=5 days, and h=6 hours are selected by experimentation on the same Meituan dataset, which is a legitimate generalization and overfitting concern but not a circular reduction. The internal numerical inconsistency between Table 5's 'Our Model' row and the 'Demand Yes' rows of Tables 3 and 4 is a reproducibility/correctness problem, not evidence that a prediction is equivalent to its input. Citations to prior work, including [27], are not load-bearing self-citations: the cited authors are not the present authors, and the cited mechanisms are standard components rather than results that define the paper's outcome. I therefore find no specific circular step that can be exhibited from the paper's text.

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

The model introduces no new physical entities. The central claim is empirical and rests on the dataset being representative, the chosen hyperparameters being appropriate, and the architectural decomposition being valid. All of these are domain assumptions rather than proven statements. The free parameters listed are tuned on the same single dataset used for evaluation.

free parameters (6)
  • Grid cell length = 2.5 km
    Selected as optimal in Sec. 5.5.1; determines spatial granularity and the 1725-node graph.
  • Time slot duration = 15 minutes
    Chosen in Sec. 5.5.2 from tested 2 hours down to 15 minutes as best; below 15 minutes accuracy did not improve.
  • Linear recurrence window N = 5 days
    Tuned in Sec. 5.5.3 from 1 to 10 previous days; 5 days gave the best prediction performance.
  • Non-linear context window h = 6 hours
    Tuned in Sec. 5.5.3; past 6 hours produced the most reliable performance.
  • Geographical neighbor threshold L = not specified
    Eq. 3 defines geographical neighbors using a threshold distance L, but its value is never reported.
  • Training hyperparameters = learning rate 0.001, batch size 2, 100 epochs
    Reported in Sec. 5.1; chosen without an ablation or justification.
assumptions (6)
  • domain assumption A fixed 2.5 km grid preserves relevant spatial demand patterns.
    Sec. 3 partitions the service area into uniform grid cells; the paper assumes this granularity is sufficient without comparing other cell sizes.
  • domain assumption Demand decomposes into linear recurring and non-linear short-term components.
    Sec. 4.2.2 builds four temporal channels on this decomposition; no evidence is given that this factorization is complete.
  • domain assumption Transfer probabilities from learned embeddings remain valid for future time slots.
    Sec. 4.2.3 predicts OD as predicted demand times transfer probability, assuming destination distributions persist.
  • domain assumption The 13-day Meituan dataset is representative of food delivery demand.
    Sec. 5.2 uses a 312-hour sample; the superiority claim is drawn from this single dataset without external validation.
  • domain assumption The graph is complete, yet only edges with positive flow are used in attention.
    Sec. 4.2 calls the graph complete, while Eqs. 1-3 restrict neighbors to those with positive flow or within a distance threshold; the paper does not reconcile this.
  • standard math Gradient-based training in 100 epochs yields a converged model.
    Sec. 5.1 sets 100 epochs without reporting convergence criteria or validation curves.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatio-Temporal Demand Prediction for Food Delivery Using Attention-Driven Graph Neural Networks." pith.science (2026). https://pith.science/paper/CVWOVR4N

@misc{pith2026250715246,
  author       = {Pith},
  title        = {Pith review of: Spatio-Temporal Demand Prediction for Food Delivery Using Attention-Driven Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CVWOVR4N}},
  note         = {Machine review of arXiv:2507.15246}
}
read the original abstract

Accurate demand forecasting is critical for enhancing the efficiency and responsiveness of food delivery platforms, where spatial heterogeneity and temporal fluctuations in order volumes directly influence operational decisions. This paper proposes an attention-based Graph Neural Network framework that captures spatial-temporal dependencies by modeling the food delivery environment as a graph. In this graph, nodes represent urban delivery zones, while edges reflect spatial proximity and inter-regional order flow patterns derived from historical data. The attention mechanism dynamically weighs the influence of neighboring zones, enabling the model to focus on the most contextually relevant areas during prediction. Temporal trends are jointly learned alongside spatial interactions, allowing the model to adapt to evolving demand patterns. Extensive experiments on real-world food delivery datasets demonstrate the superiority of the proposed model in forecasting future order volumes with high accuracy. The framework offers a scalable and adaptive solution to support proactive fleet positioning, resource allocation, and dispatch optimization in urban food delivery operations.

Figures

Figures reproduced from arXiv: 2507.15246 by the authors.

Figure 1
Figure 1. Origin-Destination Matrix typically from customers to restaurants. Ignoring this directionality can lead to suboptimal demand forecasting and resource allocation. In our work, we aim to address the limitations of the above described demand prediction models. Specifically, we focus on capturing the spatial and temporal dependencies effectively that are critical for accurately predict￾ing both the demand volume and it… view at source ↗
Figure 2
Figure 2. Architecture of the spatio-temporal demand prediction framework. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Forward, backward, and geographical neighbors of grid cell [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: MAPE over Different Time Durations [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: OD and Demand MAPE Over Days Linear Recurrence Layer. In food delivery systems, user behavior tends to follow predictable, time-aligned routines, such as ordering lunch around 1 PM or dinner around 8 PM. These patterns usually repeat consistently across different days,…
Figure 6
Figure 6. Figure 6: Performance evaluation of GNN using data from previous time slots for demand prediction. [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: MAPE and MAE for OD tasks context alone is insufficient for OD prediction unless directional flows are also explicitly captured and utilized. Our model outperforms graph-based GEML and delivers reasonable results across forecasting horizons. This illustrates the genera…
Figure 8
Figure 8. Figure 8: Actual Vs. Predicted Demand Plot enhance operational strategies and resource efficiency in three real-world domains that are the food delivery platforms, ride-hailing services, and micro-mobility or courier logistics networks. 6.1 Food Delivery Platforms In online food…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 26 canonical work pages

  1. [1]

    Maryam Akbari-Moghaddam, Stephen Kelly, and Douglas Down. 2024. Demand Forecasting and Rebalancing in Shared Bike Systems Using Deep Learning and Evolutionary Computation. In 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC) . IEEE, 3333–3338

    work page 2024

  2. [2]

    Ramon Auad, Alan Erera, and Martin Savelsbergh. 2024. Dynamic courier capacity acquisition in rapid delivery systems: A deep q-learning approach. Transportation Science 58, 1 (2024), 67–93

    work page 2024

  3. [3]

    Jingyi Cheng and Shadi Sharif Azadeh. 2025. A Short-Term Predict-Then-Cluster Framework for Meal Delivery Services. arXiv preprint arXiv:2501.08466 (2025)

    work page arXiv 2025

  4. [4]

    Alessandro Crivellari, Euro Beinat, Sandor Caetano, Arnaud Seydoux, and Thiago Cardoso. 2022. Multi-target CNN- LSTM regressor for predicting urban distribution of short-term food delivery demand. Journal of Business Research 144 (2022), 844–853

    work page 2022

  5. [5]

    Jie Gao, Xiaoming Li, Chun Wang, and Xiao Huang. 2021. BM-DDPG: An integrated dispatching framework for ride-hailing systems. IEEE Transactions on Intelligent Transportation Systems 23, 8 (2021), 11666–11676

    work page 2021

  6. [6]

    Yongshun Gong, Zhibin Li, Jian Zhang, Wei Liu, Yu Zheng, and Christina Kirsch. 2018. Network-wide crowd flow prediction of sydney trains via customized online non-negative matrix factorization. In Proceedings of the 27th ACM international conference on information and knowledge management . 1243–1252

    work page 2018

  7. [7]

    Ruiying Guo, Zhihan Jiang, Jingchun Huang, Jianrong Tao, Cheng Wang, Jonathan Li, and Longbiao Chen. 2019. BikeNet: Accurate bike demand prediction using graph neural networks for station rebalancing. In 2019 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Cloud & Big Data Computing,...

    work page 2019

  8. [8]

    Yuhan Guo, Yu Zhang, Youssef Boulaksil, and Ning Tian. 2022. Multi-dimensional spatiotemporal demand forecasting and service vehicle dispatching for online car-hailing platforms. International Journal of Production Research 60, 6 (2022), 1832–1853

    work page 2022

Show all 31 references
  1. [9]

    Yuhan Guo, Yu Zhang, Junyu Yu, and Xueli Shen. 2020. A spatiotemporal thermo guidance based real-time online ride-hailing dispatch framework. IEEE Access 8 (2020), 115063–115077

    2020

  2. [10]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. Advances in neural information processing systems 30 (2017)

    2017

  3. [11]

    Jilin Hu, Bin Yang, Chenjuan Guo, Christian S Jensen, and Hui Xiong. 2020. Stochastic origin-destination matrix forecasting using dual-stage graph convolutional, recurrent neural networks. In 2020 IEEE 36th International conference on data engineering (ICDE) . IEEE, 1417–1428

    2020

  4. [12]

    Paulo Rotela Junior, Fernando Luiz Riêra Salomon, and Edson de Oliveira Pamplona. 2014. ARIMA: An applied time series forecasting model for the Bovespa stock index. Applied Mathematics 5, 21 (2014), 3383–3391

    2014

  5. [13]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

    2016 arXiv

  6. [14]

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2017. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926 (2017)

    2017 arXiv

  7. [15]

    Jian Liang, Jintao Ke, Hai Wang, Hongbo Ye, and Jinjun Tang. 2023. A Poisson-based distribution learning framework for short-term prediction of food delivery demand ranges. IEEE Transactions on Intelligent Transportation Systems 24, 12 (2023), 14556–14569

    2023

  8. [16]

    Yile Liang, Donghui Li, Jiuxia Zhao, Xuetao Ding, Huanjia Lian, Jinghua Hao, and Renqing He. 2023. Enhancing dynamic on-demand food order dispatching via future-informed and spatial-temporal extended decisions. InProceedings of the 32nd ACM International Conference on Informat...

    2023

  9. [17]

    Junming Liu, Weiwei Chen, and Leilei Sun. 2024. A Data-Driven Optimization Framework for Static Rebalancing Operations in Bike Sharing Systems. INFORMS Journal on Computing (2024). https://doi.org/10.1287/ijoc.2022.0182

    2024

  10. [18]

    Lingbo Liu, Zhilin Qiu, Guanbin Li, Qing Wang, Wanli Ouyang, and Liang Lin. 2019. Contextualized spatial–temporal network for taxi origin-destination demand prediction. IEEE Transactions on Intelligent Transportation Systems 20, 10 (2019), 3875–3887

    2019

  11. [19]

    Ji Ma, Shenggen Zheng, Shangjing Lin, and Yonghong Cheng. 2024. A city shared bike dispatch approach based on temporal graph convolutional network and genetic algorithm. Biomimetics 9, 6 (2024), 368

    2024

  12. [20]

    Connor Riley, Pascal Van Hentenryck, and Enpeng Yuan. 2020. Real-time dispatching of large-scale ride-sharing systems: Integrating optimization, machine learning, and model predictive control. arXiv preprint arXiv:2003.10942 (2020)

    2020 arXiv

  13. [21]

    Jingran Shen, Nikos Tziritas, and Georgios Theodoropoulos. 2022. A Baselined Gated Attention Recurrent Network for Request Prediction in Ridesharing. IEEE Access 10 (2022), 86423–86434

    2022

  14. [22]

    Alex Sherstinsky. 2020. Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network. Physica D: Nonlinear Phenomena 404 (2020), 132306. Spatio-Temporal Demand Prediction for Food Delivery Using Attention-Driven Graph Neural Networks 23

    2020

  15. [23]

    Sean J Taylor and Benjamin Letham. 2018. Forecasting at scale. The American Statistician 72, 1 (2018), 37–45

    2018

  16. [24]

    Yongxin Tong, Yuqiang Chen, Zimu Zhou, Lei Chen, Jie Wang, Qiang Yang, Jieping Ye, and Weifeng Lv. 2017. The simpler the better: a unified approach to predicting original taxi demands based on large-scale online platforms. In Proceedings of the 23rd ACM SIGKDD international co...

    2017

  17. [25]

    Yuandong Wang, Xuelian Lin, Hua Wei, Tianyu Wo, Zhou Huang, Yong Zhang, and Jie Xu. 2019. A unified framework with multi-source data for predicting passenger demands of ride services. ACM Transactions on Knowledge Discovery from Data (TKDD) 13, 6 (2019), 1–24

    2019

  18. [26]

    Yuandong Wang, Hongzhi Yin, Hongxu Chen, Tianyu Wo, Jie Xu, and Kai Zheng. 2019. Origin-destination matrix prediction via graph convolution: a new perspective of passenger demand modeling. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & ...

    2019

  19. [27]

    Yuandong Wang, Hongzhi Yin, Tong Chen, Chunyang Liu, Ben Wang, Tianyu Wo, and Jie Xu. 2021. Passenger mobility prediction via representation learning for dynamic directed and weighted graphs. ACM Transactions on Intelligent Systems and Technology (TIST) 13, 1 (2021), 1–25

    2021

  20. [28]

    Zhe Xu, Zhixin Li, Qingwen Guan, Dingshui Zhang, Qiang Li, Junxiao Nan, Chunyang Liu, Wei Bian, and Jieping Ye. 2018. Large-scale order dispatch in on-demand ride-hailing platforms: A learning and planning approach. In Proceedings of the 24th ACM SIGKDD international conferenc...

    2018

  21. [29]

    Huaxiu Yao, Fei Wu, Jintao Ke, Xianfeng Tang, Yitian Jia, Siyu Lu, Pinghua Gong, Jieping Ye, and Zhenhui Li. 2018. Deep multi-view spatial-temporal network for taxi demand prediction. In Proceedings of the AAAI conference on artificial intelligence, Vol. 32

    2018

  22. [30]

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

    2021

  23. [31]

    Jie Zheng, Ling Wang, Jing-Fang Chen, Zixiao Pan, Donghui Li, Yile Liang, and Xuetao Ding. 2023. A predictive-reactive optimization framework with feedback-based knowledge distillation for on-demand food delivery. IEEE Transactions on Intelligent Transportation Systems 24, 12 ...

    2023

Pith tools

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