Problem
At most airports Uber drivers wait in a first-in-first-out (FIFO) queue and are dispatched from it. When demand is high relative to the queue, it moves fast and riders are less likely to get a suitable ride. When demand is low, the queue crawls and drivers spend more time waiting per trip. Drivers cannot see which situation they are in, so the imbalance persists.
Example from the article: the driver app shows a short, medium or long wait next to the queue size and arriving flights.
Solution in brief
Estimate the time to request (ETR), the wait before a driver gets a trip request, from three pieces: a supply model for the true position of the last driver in the queue, a demand model for queue consumption in four 15-minute steps, and a simple simulation that combines them into a short, medium or long wait class.
Questions this case answers
- How to estimate the waiting time in a first-in-first-out queue when the queue length is observed with noise and the consumption rate depends on exogenous arrivals
- How to decompose a service-time estimate into separately maintained supply and demand models that other products can reuse
- How to tell drivers how long they will wait in a FIFO airport queue before the next trip request?
- How to forecast demand for a queue that depends on arriving flights, weather and app activity?
Tools and methods
- Gradient boosted trees — supply model (true queue position, every minute) and demand model (consumption per 15 minutes)
- Queue consumption simulation — compares cumulative demand with supply to return the wait class
- Near-real-time (NRT) features — flight, weather, app and queue signals for training and serving
- Apache Kafka, Apache Hive — queue event stream and batch history
Google or a one-time email code. No password.