Improving search ranking for Maps

Malay Haldar, Hongwei Zhang, Kedar Bellare, Sherry Chen18 Dec 2024

Problem

Search ranking at a marketplace usually sorts items by booking probability and assumes attention decays down a list. On a map, listings are scattered pins, so there is no top or bottom and the decay assumption breaks. Searchers click only a few pins, so showing many pins can hide the best options, while showing too few can drop the one the searcher wanted.

Example from the article: On Airbnb desktop, 18 results appear in a grid on the left and each needs a pin on the right, so the pin count cannot simply be cut.

Solution in brief

The team models how attention spreads over the map in three steps: uniform attention (drop low-probability pins), tiered attention (show weak listings as small mini-pins) and position-discounted attention (re-center the map on the best listings). Each step was validated in an online A/B test.

Questions this case answers
  • How to rank and select items for a spatial display where position in a list no longer predicts user attention
  • How to choose the number of displayed items when a limited click budget per user makes extra options harmful
  • How to rank and select listings when results are shown as pins on a map and there is no ranked list?
  • How to choose how many map pins to show so that searchers find a suitable option with few clicks?

Tools and methods

  • Booking-probability ranking model — the existing score, reused to filter, tier and place pins.
  • Probability-ratio pin filter — keeps only listings whose booking probability is within a bound of the best one.
  • Mini-pins — a second, low-attention pin tier for listings that must still be shown.
  • Click-through-rate surface over map coordinates — measures attention by position and drives map re-centering.
  • Online A/B tests — validate each attention model on bookings and discovery metrics.
Continue reading — it's free

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