Deep Learning-Based ETA Estimation in Meituan's Delivery System

ETA (Estimated Time of Arrival) is a critical parameter in Meituan's delivery system, representing the time from order placement to delivery. Unlike ride‑hailing ETA, food‑delivery ETA must model many stages—reaching the merchant, food preparation, waiting, routing, and multi‑order batching—making accurate prediction challenging.

The model evolved from linear regression to tree‑based models, then to embedding‑enhanced FM and finally DeepFM as the base. Features were continuously enriched with user, rider, merchant, address, trajectory, area, temporal, sequential and order‑level attributes.

To align with business goals, an asymmetric absolute loss was designed: piecewise slopes (1.2× before the target, 1.8× after) encourage early arrivals while penalizing lateness more heavily. Business rules were later embedded directly into the TensorFlow graph so that model training accounts for rule‑time effects.

Long‑tail errors were addressed by adding a补时 rule that combines business‑tail factors (distance, price) with a model‑tail factor (random‑forest standard deviation). The final ETA equals the model prediction plus this补时 term.

Engineering practice uses Spark to assemble billions of raw records, writes TFRecord, trains with data‑parallel TF on multi‑GPU (single‑machine multi‑card for stability), and serves via TensorFlow Java API. To avoid GCC version upgrades on thousands of CPUs, a remote‑compute cluster handles SavedModel inference, delivering sub‑5 ms TP99 latency.