How BlaBlaCar leverages machine learning to match passengers and drivers

Fabio Cecchi1 Mar 2023

Problem

In a carpooling marketplace, a passenger sends a booking request and the driver decides whether to accept it. BlaBlaCar creates "Boost rides": the driver's published trip plus a short detour so that it fits a passenger's pickup or drop-off point. Boost rides are about 45% of the results shown in France and bring 30% of bookings. Yet drivers accept only 50% of Boost requests, against 80% for other requests. Each refusal annoys the driver and disappoints the passenger, and both are less likely to come back.

Example from the article: a driver going from Paris to Lyon gets a Boost request from a passenger who also goes to Lyon but must be picked up in Auxerre.

Solution in brief

Train a binary classifier on past Boost requests to predict whether the driver will accept. At search time, score every Boost ride and hide those below a threshold set by product and business. Passengers whose first choice is hidden may find another ride that is more likely to be accepted.

Questions this case answers
  • How to predict whether the supplier will accept a request before it is sent, when the supplier's willingness depends on request convenience and no direct signal exists
  • How to decide which low-acceptance candidates to hide from search results in a two-sided marketplace without reducing total matches
  • How to stop showing search results that the other side of a marketplace will probably reject?
  • How to predict whether a driver will accept a booking request before the passenger sends it?

Tools and methods

  • XGBoost — binary classifier of driver acceptance, scored live at search time
  • Score threshold — hides Boost rides with low predicted acceptance; set by product and business
Continue reading — it's free

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