Meituan moves delivery forecasting into a fixed latent space: unseen hotspots predicted without retraining, online MAPE down 12.38%
Synopsis
Meituan has put GLIDE into production to forecast order volume at scale-effect hotspots (SEHs, groups of delivery flows that can be pooled), recasting the task as inductive learning in a fixed-size latent space: a logistics subgraph of any size is compressed into a fixed set of latent tokens, and temporal modeling runs on those tokens, so hotspots never seen in training can still be predicted. Across three Meituan cities GLIDE is best on 10 of 12 metrics, and online weekday MAPE in the critical interval falls 12.38% against the incumbent baseline. Event disruptions are handled by multiplicative corrections from an offline LLM-built knowledge bank, with the backbone frozen and no retraining.
Interpretation
The forecasting target shifts from a fixed node set to logistics hotspots that continuously appear and vanish, so the model can predict hotspots absent from training. Earlier transductive graph models such as STGCN and DSTSGNN bind parameters or representations to fixed node identities and must be trained per subgraph, so they cannot generalize to newly emerging SEHs. Across 12 MAE/RMSE metrics on three Meituan cities in weekday and weekend settings, GLIDE is best on 10; on City A weekday RMSE it reaches 1.691 versus 1.899 for STGCN.
In cross-city zero-shot transfer, GLIDE's error comes close to a version fully supervised on the target city. Trained on City A and applied directly to City B weekdays, GLIDE zero-shot reaches MAE 1.048, versus 1.221 for the general spatiotemporal foundation model UniFlow and 1.055 for full supervision on City B itself. In the same transfer comparison, City A to City C weekdays gives GLIDE zero-shot MAE 1.022 versus 1.196 for UniFlow.
After online deployment, weekday MAPE in the critical interval drops 12.38% relative to the incumbent industrial baseline, and weekend MAPE is 0.444 versus 0.484. The incumbent baseline is the ensemble model currently deployed for traffic forecasting, with weekday critical-interval MAPE of 0.404. In the same online comparison table, weekday all-range MAE falls from 2.354 to 2.191, while weekend all-range RMSE rises from 3.715 to 3.770.
Systematic event-period bias is offset by multiplicative corrections on a frozen backbone, with no online LLM call and no retraining. Injecting event features into the backbone and retraining risks damaging learned regular patterns and is hard to operate as new event types keep appearing. In City A backtests, CRPS improves 7.3% on Labor Day, 5.0% on Qingming and 1.6% on sustained heat; Qixi is the only negative case at -0.7%.
Perspective
The demonstrated scope is three cities of Meituan platform data from July 1 2025 to July 1 2026, plus an online A/B test on Meituan's production dispatch system; the prediction target is each SEH's order volume in a 30-minute planning interval. Online subgraphs typically contain fewer than 300 nodes, while offline benchmark subgraphs range from 1 to 5,915 nodes.
The event correction is not uniformly positive: Qixi shows a CRPS change of -0.7%, which the authors attribute to the knowledge bank's impact vector overestimating the demand shift. Weekend all-range RMSE rises from 3.715 to 3.770, with gains concentrated in MAPE. The raw order-count thresholds for the six demand buckets are withheld for commercial confidentiality.
