Leveraging Knowledge Graphs in Real Estate Search

Jyoti Prakash Maheswari1 Apr 2025

Problem

Zillow listing data comes from text descriptions, photos, multiple listing service (MLS) feeds and agent input. The same feature appears as «pool», «swimming pool», «swimmingpool» or «has_pool: True», depending on the source. A user who types «large backyard» misses listings that say «huge backyard» or «fenced huge backyard», and keyword search returns too few homes, especially where inventory is small.

Example from the article: a search for «large backyard» should also show homes that mention «big backyard», «huge backyard», or the child concepts «covered large backyard» and «fenced huge backyard».

Solution in brief

Zillow keeps one knowledge graph (KG) of canonical home concepts connected by synonym and parent-child links. Both listings and queries are mapped to these concepts, so retrieval matches concepts instead of strings (concept search). New phrases are linked to the graph by a synonym list plus BERT-based classifiers, with human review.

Questions this case answers
  • How to match a query to catalog items when the same attribute is written in many surface forms across sources
  • How to discover synonym and hierarchy relations between phrases at scale with a model and limited human review
  • How to version and update a shared vocabulary that several downstream products depend on
  • How to make keyword search return listings that describe the same feature in different words?

Tools and methods

  • Knowledge graph of home concepts — canonical nodes with parent, child and synonym edges, shared by search, autocomplete and user profiles
  • Static synonym list — fast, controllable mapping from observed phrases to concepts
  • BERT pair classifier — decides whether two phrases are synonyms or parent-child
  • SBERT candidate generation — nearest-neighbor search that limits the pairs sent to the classifier
  • Human in the loop (HITL) — annotators verify uncertain links and major graph changes
Continue reading — it's free

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