Precision in Motion: Deep learning for smarter ETA predictions

Ziqi Jiang, Qingyang Xu, Chi Zhang, Hubert Jenq, Jianzhe Luo, Pradeep Varma1 Oct 2024

Problem

A food-delivery platform with over 2 billion orders a year shows customers an estimated time of arrival (ETA): order creation time plus delivery duration. DoorDash's tree-based ETA model gave reasonable forecasts, but its predictions had less variance than real durations, especially in the long tail. It handled sparse categorical features badly, needed manual feature interactions, and lost quality when couriers became scarce.

Example from the article: in a region short of couriers, new orders join the queue and the shortage compounds, but the old features captured only coarse supply and demand levels and were too noisy to show it.

Solution in brief

Replace the trees with a neural model of parallel encoders (DeepNet, CrossNet, transformer) feeding an MLP decoder, using embeddings, bucketized numbers and minute-level time series. Train the explore and checkout stages as one multitask model, and predict Weibull distribution parameters with least squares on the log-log survival function. The article reports a 20% relative accuracy gain.

Questions this case answers
  • How to predict a positive, long-tailed duration with calibrated uncertainty when the model must react to rapid shifts in system load
  • How to train one model for several prediction stages that differ in available features, so that the predictions stay consistent and accuracy does not drop
  • How to predict delivery time more accurately when a tree-based model underfits the long tail and ignores feature interactions?
  • How to make an ETA model react quickly when courier supply suddenly drops?

Tools and methods

  • MLP-gated mixture of experts — parallel encoders, outputs combined by the decoder MLP
  • Embeddings and bucketization — dense features from categories and numbers
  • Time series features with a transformer encoder — minute-level market signals
  • CrossNet encoder (DCN v2) — explicit low-rank feature crosses
  • Multitask sequential training — one model for explore and checkout
  • Weibull distribution with interval regression — delivery-time distribution
Continue reading — it's free

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