Pith. sign in

REVIEW 4 major objections 6 minor 16 references

ACA-Net: Future Graph Learning for Logistical Demand-Supply Forecasting

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Learning a future order graph from two static graph snapshots beats long-sequence spatial-temporal forecasters for on-demand food-delivery pressure.

desk verdict A plausible industrial forecasting architecture whose reported gains are not attributable to its novel components because a black-box pretrained simulator sits in the prediction path and is never ablated. read the letter →

arxiv 2509.01997 v1 pith:WBH3EPE7 submitted 2025-09-02 cs.LG cs.AI

Jiacheng Shi , Haibin Wei , Jiang Wang , Xiaowei Xu , Longzhi Du , Taixu Jiang This is my paper
classification cs.LGcs.AI
keywords demand-supplyforecastingfuturegraphlearningcrossattentionadaptivespatiotemporalfooddeliverylogisticsneuralnetworkslightweight
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

ACA-Net claims that future order distribution—the main driver of delivery pressure—can be learned from just two graph snapshots: an ongoing graph of current orders and a global graph of all historical order flows. This replaces the long sequences of past spatial-temporal graphs used by current spatiotemporal models, which are inefficient and poorly suited to the strong randomness of online ordering. The paper reports that ACA-Net achieves the best accuracy among all compared methods while using one graph with two adjacency matrices and roughly a tenth of the input bytes of graph-based spatiotemporal baselines. Because the method is compact and has been deployed in production, the two-graph representation is presented as both more accurate and more practical for online logistics forecasting.

What carries the argument

The load-bearing objects are the ongoing graph and the global graph: two directed graphs whose nodes are origin/destination Areas of Interest. The ongoing graph encodes the current order state; the global graph aggregates historical order flow over all known nodes. The argument is carried by the learned future adjacency matrix Afuture, generated by adaptive graph learning from cross-attention output and supervised by the true future adjacency matrix, so the future graph can be checked against reality. Two cross-attention transformers—inter-graph CAT (ongoing attending to global) and influence-learning CAT (supply/environment features attending to graph embeddings)—produce the features from w

What would settle it

Retrain ACA-Net end-to-end with the pre-trained pressure simulation model replaced by a randomly initialized decoder on the same Shanghai/Nanjing data, and compare MAE on the held-out test set; if the gap to the spatial-temporal baselines disappears, the claimed benefit of the two-graph representation is not independent of the simulator.

Watch

Extended reading notes

Core claim

The paper's central claim is that logistical demand-supply pressure can be forecast more accurately by explicitly learning the future order graph from two static structures—the ongoing graph (real-time order flow) and the global graph (historical order statistics over all possible origin-destination Areas of Interest)—than by modeling long time-series dependencies. ACA-Net computes inter-graph cross-attention from ongoing nodes to global nodes, applies an influence-learning cross-attention to incorporate rider and environment features, and then generates the future adjacency matrix through supervised adaptive graph learning against the true future adjacency matrix. The learned future graph,

Load-bearing premise

The reported gains assume the unpublished pre-trained pressure simulation model is available, accurate, and not already encoding the prediction target, and that the global graph contains every node that may appear in the future.

Editorial extensions

If this is right

  • Forecasting pipelines can replace long historical input windows with two graph snapshots, cutting data-collection latency and online inference cost.
  • The future adjacency matrix is supervised by ground truth, making the model's prediction of order flow directly inspectable and auditable by operators.
  • The same compact graph recipe could extend to other on-demand logistics settings where future distribution is volatile, such as ride-hailing or instant delivery.
  • The reported accuracy with smaller input bytes suggests that lightweight deployment is compatible with state-of-the-art accuracy in this setting.

Reading between the lines

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

  • Editorial: because the pre-trained pressure simulation model is not described or released, replacing it with a randomly initialized learned decoder and re-running the comparison would isolate how much of the gain comes from the two-graph future learning rather than from the fixed simulator.
  • Editorial: the assumption that the global graph contains all future nodes is testable by holding out nodes that first appear after the training period and measuring whether the model still predicts their demand-supply pressure.
  • Editorial: the ongoing/global graph split is a general template for time-series-insensitive forecasting problems; applying it to other origin-destination demand datasets (for example ride-hailing) would show whether the result is specific to food delivery or general.
  • Editorial: the ablation pattern suggests the supervised future-graph loss is the largest single contributor; a natural next experiment is varying lambda to see at what point pressure accuracy trades off against graph fidelity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes ACA-Net, a spatiotemporal model for logistical demand-supply pressure forecasting in online food delivery. Instead of feeding long sequences of past graph snapshots, the model uses two static graphs—an ongoing graph of current orders and a global graph of historical order statistics—and learns a future graph via inter-graph cross-attention, influence-learning cross-attention, and supervised adaptive graph learning. The learned future graph is then combined with supply/environment features and passed through a pre-trained pressure simulation model to predict average delivery time over the next five minutes. On a proprietary dataset from two Chinese cities, ACA-Net reports lower MAE/RMSE/MAPE and lower input bytes than several STGNN baselines, and the authors state the system is deployed in production.

Significance. If the reported results are reliable, the paper makes a useful applied contribution: replacing long historical time-series graphs with two compact graphs (ongoing and global) is an elegant idea that could reduce online inference cost while improving accuracy. The adaptive future graph learning with auxiliary supervision is also interesting, and the claim of production deployment suggests practical value. The paper deserves credit for targeting an important operational problem and for showing consistent improvements over a range of baselines. However, the central attribution of the gains is currently not verifiable because the pre-trained simulation model f2 is undisclosed and never ablated, and because the experiments lack error bars and significance tests. These issues are load-bearing for the headline claim and must be addressed before the contribution can be assessed.

major comments (4)
  1. [Section 3.4, Eq. (9)] The pre-trained pressure simulation model f2 is a black box: no architecture, training data split, features, hyperparameters, or release information is provided beyond the statement that it was pre-trained on demand distribution, supply distribution, environment factors, and the 5-minute target index. Since every pressure prediction in Eq. (9) passes through f2, and since Table 2's ablation never replaces or removes f2, the reported gains in Table 1 cannot be attributed to the two-graph future learning modules. Moreover, the comparison is potentially unfair because the STGNN baselines cannot use f2. Please add (a) a full specification of f2, or (b) an ablation where f2 is replaced by a trainable head (e.g., MLP/GNN), and (c) an experiment where the same f2 is provided to all baselines as an additional input.
  2. [Section 3.3, Eq. (8) and Section 2, Eq. (1)] The auxiliary graph target Atruth is constructed by multiplying the two edge attributes, delivery time and order count, while the pressure label p_a in Eq. (1) is the average delivery time of orders in the district. Thus the supervised graph loss Lgraph may be training the future graph to encode a quantity derived from the same delivery-time signal that the main loss LP predicts. This creates a risk of label-shortcut learning: the model may learn to reproduce Atruth and let f2 do the forecasting, rather than learning a genuinely transferable future distribution. Please provide evidence that Lgraph is not simply recovering the target, for example by ablating the graph loss while keeping f2, or by constructing Atruth from held-out/independent order data.
  3. [Tables 1 and 2] All accuracy numbers are reported as single point estimates with no standard deviations, number of seeds, or significance tests. Given that the differences among the top STGNN baselines (e.g., ASTGNN 136.1 vs. SAGDFN 135.1) are much smaller than the gap to ACA-Net, the central SOTA claim would be much stronger with mean±std over at least 5 runs. The hyperparameter tuning protocol for baselines is also not documented, which makes it hard to judge whether the comparison is apples-to-apples. Please report multi-seed statistics and describe the hyperparameter budgets for all methods.
  4. [Section 3.4] The assumption that 'the global graph can provide information on all nodes that may appear in the future' is asserted without evidence. If the global graph misses rare or emerging AOIs, the constructed G_future will be incomplete and f2 will be applied to partial inputs, which could bias the pressure estimate. Please provide quantitative coverage statistics (e.g., percentage of future orders whose origin/destination AOIs appear in the global graph, across the test period) and discuss how the model behaves when a future node is absent.
minor comments (6)
  1. [Throughout] The manuscript contains several language and typographical errors, e.g., 'donated' should be 'denoted', 'we takes the advantage' should be 'we take advantage', and 'the pressure ratio signal' is used inconsistently. A thorough proofreading pass is needed.
  2. [Eq. (6)] The notation for the linear projections and the reshaping function χ is unclear. In particular, the subscript i in q_i, k_i, v_i is not connected to the node index in the attention formula, and the relationship between Eqs. (4)–(6) should be made explicit.
  3. [Table 2] The ablation table is hard to parse: the first three rows do not clearly indicate which components are present or absent. Please use explicit ✓/✗ marks for every component in every row, and include the MAE for the configuration without the pre-trained simulation model.
  4. [Figure 3] The figure panels are dense and the font sizes are very small, especially in the CAT block diagrams. Please enlarge or redraw the figure so that the cross-attention computations and data flow are readable.
  5. [References] Reference [9] (DN-DETR) is an object-detection paper; its relevance to 'auxiliary tasks accelerate training convergence' should be stated more explicitly, since the current citation is not self-explanatory in this context.
  6. [Table 1] The units and formatting of 'input bytes' are inconsistent (e.g., '3.0 × 10^3' vs. '1.5 × 10^7'). Specify the unit (bytes, KB, MB) and clarify whether the reported runtime is per batch or per sample.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the future-graph auxiliary target and pre-trained simulator create attribution/reproducibility questions, but the derivation is not equivalent to its inputs by construction.

full rationale

ACA-Net's chain is: ongoing/global graphs and supply/environment features are embedded, passed through cross-attention, used to generate a future adjacency matrix A_future (Eq. 7), which is supervised against A_truth (Eq. 8), and then fed with F into a pre-trained simulation model f2 to produce pressure p (Eq. 9). A_truth is indeed built by multiplying edge attributes (order volume and delivery time), and the pressure label in Eq. 1 is the average delivery time over orders. So A_truth is mathematically related to the label (sum of A_truth over edges divided by total order count equals the average delivery time). However, this does not make the prediction circular: A_future is a learned function of the inputs, A_truth is an auxiliary multi-task target, and the final pressure is a composition of a learned predictor and a fixed pre-trained simulator. The supervised pressure loss (Eq. 10) and graph loss (Eq. 8) are complementary, not identical by construction. The paper does not invoke any self-citations, uniqueness theorems, or ansatz-by-citation; the references are external. The main weakness is that f2 is described only as 'a pre-trained pressure simulation model' (Section 3.4) with no architecture, training data split, or release, which makes the reported gains hard to attribute and reproduce. That is a serious reproducibility and fairness concern, but it is not circularity under the definitions used here: no fitted parameter is renamed as a prediction, and no equation reduces to its own input by definition. Therefore the honest finding is no significant circularity (score 0).

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claim depends on proprietary data, a black-box simulation model, and a hand-constructed graph target. The two-graph representation is a design choice rather than a first-principles derivation, and the pre-trained simulator is an unexposed fitted component that may carry much of the predictive power.

free parameters (4)
  • Loss weight lambda = 0.1
    Controls the weight of the future graph supervision loss in Eq. 11; set by hand, no sensitivity analysis.
  • Node embedding sizes m and M = m=100, M=500
    Output sizes for ongoing graph and global graph embeddings, Section 4.1; chosen without reported tuning.
  • Attention heads H and learning rate = H=8, lr=0.001
    Hyperparameters listed in Section 4.1; no tuning procedure or sensitivity study.
  • Ground-truth future adjacency Atruth = Product of order-volume and delivery-time edge attributes
    Defined in Eq. 8 as the supervision target for the future graph; it encodes the same quantities as the pressure label.
assumptions (4)
  • domain assumption The pre-trained pressure simulation model f2 accurately maps future graph and environment features to pressure
    Section 3.4 uses f2 as the pressure inference head; its architecture, training data, and parameters are not disclosed.
  • domain assumption The global graph contains all nodes and edges that can appear in the future
    Section 3.4 states: "We suppose that the global graph can provide information on all nodes that may appear in the future."
  • domain assumption Future order distribution is the main driver of demand-supply pressure and is learnable from ongoing and global graphs alone
    Motivates the whole design in Section 1 and Definitions 2-3; no quantitative evidence that longer histories add no information.
  • ad hoc to paper Atruth (product of edge attributes) is a valid target for the learned future graph
    Eq. 8 uses this constructed target; it is not validated against any external measure of future order distribution.
invented entities (1)
  • Future graph G_future
    purpose: Intermediate representation of predicted future order distribution that is fed to the pre-trained simulation model (Eq. 9).
    The only supervision is the paper-constructed Atruth on the same proprietary dataset; no falsifiable externally checkable prediction is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ACA-Net: Future Graph Learning for Logistical Demand-Supply Forecasting." pith.science (2026). https://pith.science/paper/WBH3EPE7

@misc{pith2026250901997,
  author       = {Pith},
  title        = {Pith review of: ACA-Net: Future Graph Learning for Logistical Demand-Supply Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WBH3EPE7}},
  note         = {Machine review of arXiv:2509.01997}
}
read the original abstract

Logistical demand-supply forecasting that evaluates the alignment between projected supply and anticipated demand, is essential for the efficiency and quality of on-demand food delivery platforms and serves as a key indicator for scheduling decisions. Future order distribution information, which reflects the distribution of orders in on-demand food delivery, is crucial for the performance of logistical demand-supply forecasting. Current studies utilize spatial-temporal analysis methods to model future order distribution information from serious time slices. However, learning future order distribution in online delivery platform is a time-series-insensitive problem with strong randomness. These approaches often struggle to effectively capture this information while remaining efficient. This paper proposes an innovative spatiotemporal learning model that utilizes only two graphs (ongoing and global) to learn future order distribution information, achieving superior performance compared to traditional spatial-temporal long-series methods. The main contributions are as follows: (1) The introduction of ongoing and global graphs in logistical demand-supply pressure forecasting compared to traditional long time series significantly enhances forecasting performance. (2) An innovative graph learning network framework using adaptive future graph learning and innovative cross attention mechanism (ACA-Net) is proposed to extract future order distribution information, effectively learning a robust future graph that substantially improves logistical demand-supply pressure forecasting outcomes. (3) The effectiveness of the proposed method is validated in real-world production environments.

Figures

Figures reproduced from arXiv: 2509.01997 by the authors.

Figure 1
Figure 1. (a) It shows the logistical demand-supply forecasting performance of the predic￾tion model and simulation model. (b) It shows annular and astroid order distribution respectively, which has similar order count but different logistical pressure. demand-supply forecasting metrics. Hence, the accuracy and compactness of a logistical demand-supply pressure forecasting model are crucial for food delivery logistics dispatc… view at source ↗
Figure 2
Figure 2. Existing sequence-based method and our ongoing-global-based method. can be mitigated to some extent while it brings complexity challenge. b) Fu￾ture graph learning challenge. Existing methods utilize feature embedding or self-adaptive graph learning to derive the future order distribution informa￾tion. Feature embedding generates feature representations without robust future graph learning. In comparison, self-adapt… view at source ↗
Figure 3
Figure 3. The overview of the proposed ACA-Net. edges in the ongoing graph represent real-time order distribution are directed. Definition 4: Supply and Environment Features. In addition to order distribution, the key factors affecting the demand-supply forecasting of a spe￾cific business district include supply distribution and real-time environmental elements such as meteorological conditions, courier availability, and traf… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: (a)(b)(c)(d) Result curves of ACA-Net and contrast experiments. (e) Visual￾ization of the predicted future graph and the real graph. both spatial and temporal features. They utilize statistical features and the con￾tinuous order distribution graphs of past ∆t time slic…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 12 canonical work pages

  1. [1]

    Computers & industrial engineering143, 106435 (2020)

    Abbasimehr, H., Shabani, M., Yousefi, M.: An optimized model using lstm network for demand forecasting. Computers & industrial engineering143, 106435 (2020)

    work page 2020

  2. [2]

    In: Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining

    Chen, T., Guestrin, C.: Xgboost: A scalable tree boosting system. In: Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. pp. 785–794 (2016)

    2016

  3. [3]

    Journal of Tourism Futures5(1), 75–93 (2019)

    Ghalehkhondabi, I., Ardjmand, E., Young, W.A., Weckman, G.R.: A review of demand forecasting models and methodological developments within tourism and passenger transportation industry. Journal of Tourism Futures5(1), 75–93 (2019)

    work page 2019

  4. [4]

    arXiv preprint arXiv:1703.04247 (2017)

    Guo, H., Tang, R., Ye, Y., Li, Z., He, X.: Deepfm: a factorization-machine based neural network for ctr prediction. arXiv preprint arXiv:1703.04247 (2017)

    arXiv 2017

  5. [5]

    IEEE Transactions on Knowledge and Data Engineering34(11), 5415–5428 (2021)

    Guo, S., Lin, Y., Wan, H., Li, X., Cong, G.: Learning dynamics and heterogene- ity of spatial-temporal graph data for traffic forecasting. IEEE Transactions on Knowledge and Data Engineering34(11), 5415–5428 (2021)

    work page 2021

  6. [6]

    SAGDFN: A Scalable Adaptive Graph Diffusion Forecasting Network for Multivariate Time Series Forecasting

    Jiang, Y., Li, X., Chen, Y., Liu, S., Kong, W., Lentzakis, A.F., Cong, G.: Sagdfn: A scalable adaptive graph diffusion forecasting network for multivariate time series forecasting. arXiv preprint arXiv:2406.12282 (2024)

    work page Pith review arXiv 2024

  7. [7]

    Information Sciences588, 315–330 (2022)

    Jin, G., Liu, C., Xi, Z., Sha, H., Liu, Y., Huang, J.: Adaptive dual-view wavenet for urban spatial–temporal event prediction. Information Sciences588, 315–330 (2022)

    work page 2022

  8. [8]

    Neurocomputing510, 79–94 (2022)

    Jin, G., Xi, Z., Sha, H., Feng, Y., Huang, J.: Deep multi-view graph-based network for citywide ride-hailing demand prediction. Neurocomputing510, 79–94 (2022)

    work page 2022

Show all 16 references
  1. [9]

    In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition

    Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition. pp. 13619–13627 (2022)

    2022

  2. [10]

    In: The Tenth International Conference on Learning Representa- tions, ICLR 2022, Virtual Event, April 25-29, 2022

    Liu, S., Yu, H., Liao, C., Li, J., Lin, W., Liu, A.X., Dustdar, S.: Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting. In: The Tenth International Conference on Learning Representa- tions, ICLR 2022, Virtual Event, April 25-29, ...

    2022

  3. [11]

    Proceedings of the VLDB Endowment15(11), 2733–2746 (2022)

    Shao, Z., Zhang, Z., Wei, W., Wang, F., Xu, Y., Cao, X., Jensen, C.S.: Decoupled dynamic spatial-temporal graph neural network for traffic forecasting. Proceedings of the VLDB Endowment15(11), 2733–2746 (2022)

    2022

  4. [12]

    Information Sciences 479, 214–230 (2019)

    Silva, E.S., Hassani, H., Ghodsi, M., Ghodsi, Z.: Forecasting with auxiliary in- formation in forecasts using multivariate singular spectrum analysis. Information Sciences 479, 214–230 (2019)

    2019

  5. [13]

    IEEE Transactions on Intelligent Transportation Systems 23(9), 16137–16147 (2022)

    Wang, H., Zhang, R., Cheng, X., Yang, L.: Hierarchical traffic flow prediction based on spatial-temporal graph convolutional network. IEEE Transactions on Intelligent Transportation Systems 23(9), 16137–16147 (2022)

    2022

  6. [14]

    In: Proceedings of the web conference 2020

    Wang, X., Ma, Y., Wang, Y., Jin, W., Wang, X., Tang, J., Jia, C., Yu, J.: Traffic flow prediction via spatial temporal graph neural network. In: Proceedings of the web conference 2020. pp. 1082–1092 (2020)

    2020

  7. [15]

    arXiv preprint arXiv:1906.00121 (2019)

    Wu, Z., Pan, S., Long, G., Jiang, J., Zhang, C.: Graph wavenet for deep spatial- temporal graph modeling. arXiv preprint arXiv:1906.00121 (2019)

    1906 arXiv

  8. [16]

    In: Proceed- ings of the AAAI conference on artificial intelligence

    Zhou, H., Zhang, S., Peng, J., Zhang, S., Li, J., Xiong, H., Zhang, W.: Informer: Beyond efficient transformer for long sequence time-series forecasting. In: Proceed- ings of the AAAI conference on artificial intelligence. vol. 35, pp. 11106–11115 (2021)

    2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.