Precision in Motion: Deep Learning for Smarter ETA Predictions
Article Summary
DoorDash handles 2 billion orders annually, and every ETA prediction matters. Their old tree-based models couldn't keep up with the complexity.
The DoorDash ML team rebuilt their ETA prediction system from the ground up using deep learning. They combined multiple neural network architectures with probabilistic modeling to handle the intricate patterns across merchants, Dashers, and delivery stages.
Key Takeaways
- 20% relative improvement in ETA accuracy across all delivery scenarios
- MLP-gated mixture of experts with DeepNet, CrossNet, and transformer encoders
- Time series features improved accuracy 20% during high demand periods
- Multitask learning ensures consistency between explore and checkout stages
- Weibull distribution modeling captures long-tail delivery time uncertainty
Critical Insight
DoorDash achieved 20% better ETA accuracy by replacing tree-based models with a sophisticated deep learning system that combines specialized encoders, embeddings, and probabilistic predictions.
