Transforming Location Retrieval at Airbnb: A Journey from Heuristics to Reinforcement Learning

Dillon Davis, Huiji Gao, Thomas Legrand, Weiwei Guo, Malay Haldar, Alex Deng, Han Zhao, Liwei He, Sanjeev Katariya11 Nov 2024

Problem

Airbnb cannot rank every listing for every search, so before ranking it must choose a map area to pull listings from. This step is called location retrieval. The area must be wide enough to include listings guests actually book, yet small enough to stay relevant. A fixed rule per destination type cannot do both, because the best area depends on the guest and the destination.

Example from the article: a family searching for San Francisco may book a larger house in nearby Daly City, while a solo traveler wants a central apartment.

Solution in brief

The team moved through four stages: tuned heuristics, booking statistics, a neural network that predicts the map rectangle from search parameters, and finally the same network turned into a contextual bandit with Monte Carlo Dropout and an upper confidence bound. Statistics gave no gain; the other three did. Cumulative effect: +2.66% uncancelled bookers.

Questions this case answers
  • How to choose the candidate region to retrieve from, before ranking, when the full inventory is too large to rank and a fixed rule fits some queries badly
  • How to add controlled exploration to a retrieval model so that uncertain queries explore more and well-known queries explore less
  • How to decide which map area to retrieve listings from when a guest types a destination?
  • How to replace hand-written radius rules for location search with a learned model?

Tools and methods

  • Rule-based heuristics — administrative bounds, a 25-mile city radius, a log-scale expansion factor.
  • Booking-statistics areas — per-destination areas covering 96% of nearest booked listings (no gain).
  • Neural network regressor — two layers of 256 units; predicts four offsets of the retrieval rectangle.
  • Monte Carlo Dropout — 32 stochastic passes per search give a mean and spread.
  • Upper confidence bound (UCB) — widens the area where the model is unsure.
  • TensorFlow, Horovod — training stack (per the paper).
Continue reading — it's free

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