DeepETA: How Uber Predicts Arrival Times Using Deep Learning

Tanmay Binaykiya, Xinyu Hu, Olcay Cirit, Ramit Hora10 Feb 2022

Problem

Uber computes arrival times (ETAs) for fares, pickup estimates, rider-driver matching and delivery planning. A routing engine sums segment travel times along the best path, but a road graph cannot capture real conditions or the route a driver will pick. A gradient-boosted tree model corrected the engine for years, until it could no longer grow with the data. The correction model is the highest-QPS model at Uber, so it must answer within a few milliseconds and still cut mean absolute error (MAE) across rides and delivery worldwide.

Example from the article: the model takes a trip's origin, destination, time and traffic and predicts the gap between the engine's ETA and the real outcome.

Solution in brief

Keep the routing engine and train a deep model on its residual. Bucketize and embed every input, pass the embeddings through one linear-attention layer, and decode with segment bias adjustments. Train with a parameterized asymmetric Huber loss so one model gives different point estimates per use case.

Questions this case answers
  • How to refine the output of a physics-based or rule-based estimator with a learned residual model under a hard real-time latency limit
  • How to make a deep model on high-cardinality tabular and geospatial inputs fast enough for the highest-traffic prediction endpoint
  • How to correct a routing engine's arrival-time estimate with a machine learning model when the tree ensemble no longer scales?
  • How to move a tabular model from gradient boosting to a neural network without breaking a few-millisecond latency budget?

Tools and methods

  • Residual (post-processing) model — predicts observed ETA minus routing-engine ETA.
  • Quantile bucketing + embeddings — discretizes inputs.
  • Multiple feature hashing — compact location embeddings.
  • Linear transformer encoder — feature interactions.
  • Segment bias adjustment — per-segment correction.
  • Asymmetric Huber loss — outlier robustness, late vs early cost.
  • XGBoost, Horovod — the incumbent baseline and the data-parallel training route.
Continue reading — it's free

Google or a one-time email code. No password.