Problem
A rider sees a pickup ETA before requesting a ride, and the driver may arrive much later than that. Low reliability goes with more rider cancellations, so an ETA is only useful if it is likely to be met. Whether it is met depends on things unknown at display time: which driver will accept, whether several requests compete for the same driver, whether drivers go offline, plus traffic, detours and GPS noise.
Example from the article: a rider with an early flight sees an estimated pickup time and needs to know whether the driver will arrive at that time.
Solution in brief
Lyft trains a gradient-boosted tree classifier that scores every candidate ETA bracket (for example 1 to 10 minutes) before the request. Features describe nearby drivers, recent regional history and live marketplace conditions. A stringent reliability SLA then decides which ETA can be shown. Training duplicates each ride once per candidate ETA, so the model sees every ETA it will score at inference.
Questions this case answers
- How to estimate the probability that a promised arrival time will be met before the customer commits, when supply availability and the matching outcome are still unknown
- How to train a classifier on logs shaped by the currently shown option without a feedback loop that degrades the model over time
- How to estimate the probability that a driver arrives within a reasonable window around a pickup ETA before the rider requests the ride?
- How to pick, from a fixed set of candidate ETAs, the most accurate one that still meets a reliability target?
Tools and methods
- Gradient-boosted trees — binary classifier of whether the driver arrives within the window around an ETA
- Reliability SLA threshold — rule that shows only ETAs whose predicted reliability exceeds the target
- Training-data duplication per candidate ETA — one row per ride and possible ETA, to avoid a feedback loop
- AUC and per-bracket checks — offline quality evaluation
- Drift alarms and retraining — keep the model current
Google or a one-time email code. No password.