Title: UME: A Unified Meta-Generalization Framework for Cross-Domain ETA

Abstract: Precise Estimated Time of Arrival (ETA) forecasting on checkout interfaces is vital for instant logistics services, as it drives user satisfaction, streamlines dispatching processes, and manages operational expenditures. For international on-demand delivery networks, where ETA data stems from various countries and regions exhibiting distinct behavioral patterns, multi-domain modeling has become a standard practice. Nevertheless, deploying these models in real-world scenarios presents three significant hurdles. Firstly, contemporary multi-domain approaches lack the ability to generalize to entirely new domains, resulting in an inability to perform zero-shot predictions during the initial cold-start period. Secondly, while existing methods typically assume consistency across cross-domain feature spaces, emerging domains frequently experience structural gaps in offline (statistical) features due to insufficient historical records. Thirdly, this data scarcity often forces industrial systems to treat established and cold-start domains as separate entities, which impedes knowledge transfer and escalates maintenance burdens.

To overcome these obstacles, we introduce UME, a Unified Meta-generalization framework for ETA. UME combines a cohesive dual-branch architecture with an innovative meta-learning mechanism centered on a hypernetwork-based meta learner. By utilizing both domain-level insights and instance-level contexts, this meta learner enables three specific meta modules to dynamically adjust feature gating, expert attention, and final predictions. This approach effectively captures cross-domain correlations while supporting intra-domain adaptation. Additionally, a knowledge distillation strategy is incorporated to further boost performance. UME is currently live on the Meituan-Keeta delivery platform, recognized as the largest international food delivery service in China. Comprehensive offline evaluations and online A/B testing confirm that UME substantially surpasses existing baseline methods.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC