Accurate forecasting of order volume at scale-effect hotspots (SEHs), spatially coherent groups of logistics flows that serve as distributed optimization units for real-time many-to-one dispatch, is a prerequisite for efficient order pooling in on-demand food delivery (OFD) platforms such as Meituan. However, real-world logistics networks present unique challenges that render standard spatiotemporal prediction model ineffective. (i) Dynamic Topology: Unlike static road networks, logistics graphs are ephemeral, with hotspot emerging and vanishing continuously; (ii) Scale--Volume Heterogeneity: Graph sizes vary from 102 to 104, while zero-inflated, long-tailed volumes create computational bottlenecks and bias predictions toward inactivity; and (iii) Event-Driven Disruption: Exogenous events (holidays, weather, academic calendars) induce systematic demand shifts that purely data-driven backbones cannot anticipate. In this paper, we present GLIDE, a production-ready framework that models evolving logistics graphs in a scalable latent space while adapting to heterogeneous demand regimes and event-driven shifts. GLIDE introduces a Graph-Guided Latent Aggregation mechanism, which dynamically projects varying physical subgraphs into a fixed-size, semantically stable latent space via a differentiable sparse assignment...
Figures & tables
Figure 1. The Framework of GLIDE (Graph-Latent Inductive Dynamics prediction with Event awarenes).
City ID
Type
# Subgraphs
Nodes per Subgraph ∣VS∣
City A
Wd / We
64
1 – 5,915
City B
Wd / We
79
1 – 3,224
City C
Wd / We
18
6 – 3,934
Table 1. Statistics of the Meituan OFD Datasets. The extreme variance in ∣VS∣ (e.g., 1 to 5,915) highlights the challenge for transductive baselines.
Table 3
Figure 2. Performance comparison of the GLIDE architecture and its ablation variants across City A-Wd and City B-Wd.
Event Type
Event
ATM%
N
Hist CRPS
P1 CRPS
P2 CRPS ( ↑ )
Public Holidays (City-wide)
Holiday
New Year
55.9%
640,927
1.3705
1.2946
1.2702 (+1.9%)
Holiday
Qingming
40.4%
625,165
1.1878
1.1243
1.0685 (+5.0%)
Holiday
Labor Day
66.0%
1,024,930
1.1407
1.0771
0.9990 (+7.3%)
Holiday
Dragon Boat
100%
647,190
1.6368
1.5435
1.5227 (+1.3%)
Weather Events (City-wide)
Table 4. Event-Scenario Evaluation on City A across diverse event types (affected timesteps). We compare History (quantile prior), GLIDE backbone (Phase-1, P1), and GLIDE+EventModulator (Phase-2, P2). ↑ denotes relative CRPS improvement of P2 over P1; positive means P2 is better.
Weekday Scenario
Weekend Scenario
Range
Model
MAE
RMSE
MAPE
MAE
RMSE
MAPE
[0,∞)
Baseline
2.354
3.759
0.493
2.522
3.715
0.484
GLIDE
2.191
3.478
0.445
2.504
3.770
0.444
Improv. (%)
+6.92%
+7.48%
+9.74%
+0.71%
-1.48%
+8.26%
Critical Interval
Baseline
3.465
4.884
0.404
3.507
4.683
0.353
GLIDE
3.139
4.225
0.354
3.465
4.593
0.333
Table 5. Online Regression Performance. We report MAE, RMSE, and MAPE. The “Critical Interval” represents the AFUs most sensitive to dispatch optimization.
Bundle size
1
2
3
4
5
6
7
Base (%)
46.42
37.95
11.16
3.76
0.71
0.00
0.00
Test (%)
13.96
39.32
26.07
13.87
4.34
1.48
0.96
Δ (pp)
−32.46
+1.37
+14.91
+10.11
+3.63
+1.48
+0.96
Table 6. Bundle-size distribution in the online A/B test. Base and Test denote the incumbent and GLIDE-guided dispatch strategies, respectively; Δ is Test − Base in percentage points.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3. Scalability and inference efficiency of GLIDE.
Subtype
Phase
Meal
Dominant AOI
Scalar
New Year
start
lunch
Residential (75%)
−3
New Year
start
lunch
Office+Industrial (53%)
−3
Rainstorm
start
lunch
Residential (100%)
+3
Rainstorm
middle
dinner
Residential (98%)
+3
Appendix
Table 7. Illustrative impact vectors from the knowledge bank. The AOI distribution is a 16-dim normalized vector over AOI types: [Residential, Office, University, K-12, Mall, Hospital, Public, Civic, Park, Hotel, Airport, Station, Other, Industrial, Market, Training].
(b) Window ordering at Th=5 (same backbone, same val split)
Appendix
Table 8. Input Reprogramming study. (a) Look-back window length Th trade-off: shorter windows reduce latency but degrade accuracy. (b) Symmetric vs. non-symmetric window under the same Th=5 backbone: the symmetric (reprogrammed) window outperforms the non-symmetric one, confirming that input ordering—not just window length or freshness—affects forecast quality. Both panels use the same offline validation split.
Figure 4. (a) Probability density of normalized order volume for weekdays and weekends. (b) Distribution of subgraph scales (number of nodes) generated during route generation.
Weekday Scenario
Weekend Scenario
Range
Model
R-MAE
Width ↓
Cov. ↑
CRPS
R-MAE
Width ↓
Cov. ↑
CRPS
[0,∞)
Baseline
2.145
6.001
81.7%
0.386
2.302
6.320
76.2%
0.408
GLIDE
1.758
5.014
89.5%
0.350
2.150
5.875
89.1%
0.384
Improv. (%)
+18.04%
+16.45%
+9.55%
+9.33%
+6.60%
+7.04%
+16.93%
+5.88%
Critical Interval
Baseline
3.556
9.345
76.6%
0.348
3.794
10.16
71.4%
0.376
GLIDE
2.918
7.775
85.3%
0.314
3.503
8.765
86.0%
0.350
Appendix
Table 9. Online Distribution Performance. Higher Coverage with lower Width indicates superior uncertainty estimation quality. The “Critical Interval” represents the AFUs most sensitive to dispatch optimization.
Panel B: Source Domain → City B (Medium-scale, 34k nodes)
Model
Target: City A (Large)
Target: City C (Small)
Weekday
Weekend
Weekday
Weekend
MAE
RMSE
MAE
RMSE
MAE
RMSE
MAE
RMSE
HA
1.335
1.895
1.455
2.091
1.186
1.689
1.182
1.649
UniFlow (Zero-shot)
1.345
1.893
1.336
2.052
1.147
1.691
1.122
1.647
GLIDE (Zero-shot)
1.147
1.714
1.291
1.947
0.989
1.531
1.006
1.547
Appendix
Table 10. Panel B: Zero-Shot Transfer from City B (Medium-scale, 34k nodes). “GLIDE (Supervised)” represents the upper-bound performance where the model is trained on the target city itself.
Model
Target: City A (Large)
Target: City B (Medium)
Weekday
Weekend
Weekday
Weekend
MAE
RMSE
MAE
RMSE
MAE
RMSE
MAE
RMSE
HA
1.335
1.895
1.455
2.091
1.246
1.781
1.293
1.892
UniFlow (Zero-shot)
1.331
1.994
1.369
2.044
1.242
1.767
1.296
2.001
GLIDE (Zero-shot)
1.166
1.755
1.293
1.955
1.093
1.690
1.140
1.836
Ref: GLIDE (Supervised)
1.131
1.691
1.284
1.939
1.055
1.624
1.132
1.819
Appendix
Table 11. Panel C: Zero-Shot Transfer from City C (Small-scale, 19k nodes).
City B (34k nodes)
City C (19k nodes)
Event
Hist
P1
P2
↑
Hist
P1
P2
↑
Weather
Rainstorm
1.4285
1.3812
1.3670
+1.0%
1.5849
1.5349
1.5184
+1.1%
Sust. Heat
—
—
—
—
1.6048
1.5555
1.5432
+0.8%
Platform Promotions
Double 11
1.4222
1.3785
1.3683
+0.7%
1.3895
1.3513
1.3388
+0.9%
Appendix
Table 12. Cross-city zero-shot transfer to City B and City C (unseen during training). The City A-trained backbone and modulator are applied without retraining. We report affected-sample CRPS and the relative improvement of Phase-2 (GLIDE+Event) over Phase-1 (GLIDE backbone). Positive ↑ means the event modulator helps.
Holiday
New 123-event
Legacy 27-event
Coverage
Naff
CRPS ↑
Naff
CRPS ↑
Legacy
New
New Year
640,927
+1.9%
640,927
+1.8%
55.9%
55.9%
Qingming
625,165
+5.0%
625,165
+5.0%
40.4%
40.4%
Labor Day
1,024,930
+7.3%
891,855
+7.8%
57.5%
66.0%
Dragon Boat
647,190
+1.3%
212,370
-0.4%
32.8%
100.0%
All
2,938,212
+4.0%
2,370,317
+4.4%
—
—
Appendix
Table 13. Ablation on event library coverage. Replacing the 123-event library with a legacy 27-event library reduces affected-sample coverage and degrades Phase-2 gains. On Dragon Boat, the CRPS improvement of Phase-2 over Phase-1 flips from −0.4% (legacy) to +1.3% (enriched), driven by newly added weather events that raise coverage from 32.8% to 100%.
Volume Interval
Bucket 1
Bucket 2
Bucket 3
Bucket 4
Bucket 5
Bucket 6
HA
1.142
3.585
5.051
6.076
6.859
6.867
w/o MoE
1.137
3.602
5.063
6.105
7.047
6.800
ZIP Baseline
1.128
3.597
5.074
6.387
6.965
6.867
GLIDE (MoE)
1.106
3.556
5.001
5.937
6.459
6.533
Appendix
Table 14. Performance comparison (MAE) across stratified volume buckets. Bold denotes the best performance, and underlined denotes the second best.
Model
City A-Wd
City A-We
Informer
4h 33m 35s
5h 14m 09s
STGCN
3h 22m 45s
2h 53m 55s
Autoformer
1h 21m 27s
1h 13m 50s
DSTSGNN
49m 32s
48m 09s
DLinear
37m 27s
30m 33s
STID
21m 17s
22m 45s
Appendix
Table 15. Computational costs on the City A dataset.
Figure 5. Visualization of the Learned Routing Mechanism on City A (Weekday). (a) Expert Specialization Matrix: The heatmap shows the average activation probability of each expert across discretized volume buckets. (b) Dynamic Routing Mechanism: The curves illustrate the continuous evolution of expert dominance as order volume increases.
Modern logistics networks generate rich operational data streams at every warehouse node and transportation lane -- from order timestamps and routing records to shipping manifests -- yet predicting delivery delays remains predominantly reactive. Existing predictive approaches typically treat this problem either as a tabular classification task, ignoring network topology, or as a time-series anomaly detection task, overlooking the spatial dependencies of the supply chain graph. To bridge this gap, we propose a hybrid deep learning framework for proactive supply chain risk management. The proposed method jointly models temporal order-flow dynamics via a lightweight Transformer patch encoder and inter-hub relational dependencies through an Edge-Aware Graph Attention Network (E-GAT), optimized via a multi-task learning objective. Evaluated on the real-world DataCo Smart Supply Chain dataset, our framework achieves consistent improvements over baseline methods, yielding an F1-score of 0.8762 and an AUC-ROC of 0.9773. Across four independent random seeds, the framework exhibits a cross-seed F1 standard deviation of only 0.0089 -- a 3.8 times improvement over the best ablated variant -- achieving the strongest balance of predictive accuracy and training stability among all evaluated models.
Zhiming Xue, Menghao Huo, Yujue Wang
College of Engineering, Northeastern University Boston, USA · School of Engineering, Santa Clara University Santa Clara, USA · Department of Chemistry and Chemical Biology, University of New Mexico Albuquerque, USA
Delivery systems have become a core part of urban life, supporting the demand for food, medicine, and other goods. Yet traditional logistics networks remain fragile, often struggling to adapt to road closures, accidents, and shifting demand. Online Food Delivery (OFD) platforms now represent a cornerstone of urban logistics, with the global market projected to grow to over 500 billion USD by 2030. Designing delivery networks that are efficient and resilient remains a major challenge: fully connected graphs provide flexibility but are computationally infeasible at scale, while single Minimum Spanning Trees (MSTs) are efficient but easily disrupted. We propose the Union of Minimum Spanning Trees (UMST) approach to construct delivery networks that are sparse yet robust. UMST generates multiple MSTs through randomized edge perturbations and unites them, producing graphs with far fewer edges than fully connected networks while maintaining multiple alternative routes between delivery hotspots. Across multiple U.S. cities, UMST achieves 20--40× fewer edges than fully connected graphs while enabling substantial order bundling with 75--83% participation rates. Compared to learning-based baselines including MADDPG and Graph Neural Networks, UMST delivers competitive performance (88-96% success rates, 44-53% distance savings) without requiring training, achieving 30× faster execution while maintaining interpretable routing structures. Its combination of structural efficiency and operational flexibility offers a scalable and resilient foundation for urban delivery networks.
Md Nahid Hasan, Vishwam Tiwari, Aditya Challa +2
Department of Computer Science and Software Engineering, Miami University, Oxford, Ohio, USA · Department of Computer Science & Information Systems, BITS Pilani K. K. Birla Goa Campus, and APPCAIR, BITS Goa, India
Forecasting urban delivery demand becomes substantially more challenging when newly added service regions lack historical records. Existing spatiotemporal forecasters effectively model spatial dependence once sufficient node histories are available. Still, they remain parametric and therefore struggle to recover short-term operational dynamics in cold-start regions. Geospatial embeddings help identify where a region is and what function it serves, yet they do not directly reveal how a similar region behaves under a comparable temporal context. We propose Bridge, a retrieval-augmented spatiotemporal graph framework that combines an inductive contextual graph backbone with a time-aware memory of region-time windows. For each target region, Bridge retrieves future demand patterns from the memory using both regional context and recent dynamics, and refines the backbone forecast through a gated fusion mechanism. To align retrieval with forecasting utility, we further train the retriever with a future-aware objective that favors entries whose future trajectories best match the target. Experiments on four real-world delivery datasets show that Bridge consistently improves over competitive spatiotemporal baselines in both within-city cold-start and cross-city transfer with partial observations. The results show that retrieval augmentation provides a useful operational memory for cold-start urban demand forecasting when parametric graph generalization alone is insufficient.
Yihong Tang, Tong Nie, Junlin He +3
McGill University · The Hong Kong Polytechnic University · University of Toronto +1