Helping users discover their dream homes through home insights collections

Giuliano Janson14 Feb 2023

Problem

A real estate marketplace has far more listings than one screen can show, and each user cares about different things. Some want a pool or a remodeled kitchen, but those details sit in free-text descriptions and photos. Users find them only by opening listings one by one. Zillow wanted themed lists of homes ("Collections") that surface such features for each user. Two things made this hard: thousands of unique, very specific phrases ("olympic-sized swimming pool"), and clicks that say little about which feature a user liked.

Example from the article: a user interested in remodeled kitchens opens the Updates tab and finds a new "Homes with Remodeled Kitchens" Collection.

Solution in brief

Group the extracted phrases into a hierarchical taxonomy built semi-automatically. Build each user's preference histograms from strong signals (dwell time, favorites, explicit filters) and choose the taxonomy level for each Collection by inventory and user interest. Rank Collections and the homes inside them by the user's profile.

Questions this case answers
  • How to build personalized themed collections of items from attributes extracted from unstructured text, when clicks only weakly reflect interest in those attributes
  • How to group thousands of sparse, fine-grained item attributes into a hierarchy so that each themed list has enough items
  • How to build personalized themed lists of listings from features that appear only in free-text descriptions?
  • How to infer which item features a user cares about when clicks are a weak signal?

Tools and methods

  • Real-estate taxonomy — tree of home insights linked by hypernym edges, so niche phrases roll up into broader categories.
  • Sub-word embedding model — vectors for phrases trained on listing descriptions.
  • Hypernym classifier with human-in-the-loop review — removes noisy edges and retrains on reviewed edges.
  • User preference histograms — per-feature distributions from user activity.
  • A/B tests — compare Collection types by CTR, saves and right-scrolls.
Continue reading — it's free

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