Contextual Property Embeddings for Corse-grained Personalization

Ioannis Partalas8 Jul 2025

Problem

A hotel search system ranks properties with the help of one fixed embedding per property. Every traveler gets the same property vector, although travelers differ by loyalty status: higher-tier travelers tend to prefer higher-quality properties. So the ranker cannot tailor results to the traveler's current tier, which limits relevance and conversion rate (CVR). The team wants coarse, cohort-level personalization built into the embeddings themselves.

Example from the article: at Expedia Group, property embeddings from the hotel2vec model feed the search ranker, and the team wants them to reflect the traveler's loyalty tier (blue, silver, gold).

Solution in brief

Add a traveler context embedding (loyalty tier) to the hotel2vec input and fuse it early with the other feature embeddings, so the model produces one property vector per tier. Export one embedding per property and tier, store them in a feature store, and join the one matching the traveler's tier to each impression at ranking time. The same pattern fits any coarse segment known at query time.

Questions this case answers
  • How to make item embeddings depend on a coarse user segment so a downstream ranker can personalize by cohort without per-user models
  • How to personalize item embeddings for traveler segments such as loyalty tiers?
  • How to add user context to an item embedding without training one model per segment?
  • How to feed context-dependent item embeddings into a search ranking model at inference time?

Tools and methods

  • hotel2vec with context input — property embedding trained on click sessions, extended with a loyalty-tier embedding
  • Negative sampling loss — the same loss as hotel2vec, with 2000 negatives drawn from the same market
  • Feature store — holds one embedding per property and tier for lookup at inference
  • Search ranking model — consumes the tier-specific embedding as an input feature
Continue reading — it's free

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