Skip to content

SYSTEM Cited by 1 source

Expedia Lodging Ranker

Definition

The Expedia Lodging Ranker is the ranking algorithm behind the search-results list on Expedia Group lodging surfaces (hotels, properties). It decides which of the candidate properties matching a user's search are shown, and in which order.

Expedia has not published the ranker's architecture (feature set, model family, serving stack); what has been published publicly is the experimentation harness built around it — specifically the use of interleaving as an accelerated alternative to A/B testing for evaluating candidate ranking changes ( Expedia, 2026-02-17).

What's disclosed

  • Treatment types evaluated — e.g., "pinning a random property to a slot between positions 5 and 10", "randomly reshuffling a number of top slots". Both are deliberately-deteriorated treatments used to validate the sensitivity of the interleaving framework against known-bad rankings.
  • Event types tracked — property-detail-page views (clicks) and booking transactions, separately. See patterns/interleaved-ranking-evaluation.
  • Experimentation substrate — interleaving with per-event attribution, lift metric reported at user level by default, significance via winning-indicator t-test (patterns/t-test-over-bootstrap).

What's not disclosed

  • Model family (gradient-boosted trees, deep ranking network, LTR, LambdaMART, Transformer-based) — unstated.
  • Feature set (price, location, user history, seasonality, property quality, conversion history, partner-commercial signals) — unstated.
  • Serving stack (offline batch scoring, online re-ranking, feature store, cache hierarchy, candidate-generation funnel) — unstated.
  • Latency / QPS / fleet size — unstated.
  • Interleaving variant used (team-draft, balanced, probabilistic) — unstated.
  • Relationship to Expedia's centralised embedding platform — not addressed in the public post; plausibly the embedding store feeds some candidate-generation or feature-extraction step for the ranker, but not confirmed.

Interleaving results on synthetic regressions

Expedia uses two deliberately-bad ranking treatments to validate their experimentation harness — the real-world utility is that the same sensitivity detects real, subtle ranking changes early:

Treatment A/B on CVR uplift Interleaving (clicks) Interleaving (bookings)
Pin random property to slot 5–10 Not detected at full sample size Detected in first day Detected within days
Reshuffle top slots Marginal Detected in first day Detected within days

(Qualitative summary of the post's Figure 5 and accompanying text.)

Caveats

  • This page is an experimentation-harness record, not a ranker architecture record. If Expedia publishes a ranker-architecture post later, expand this page accordingly.
  • The subject of measurement here is the lodging-search ranking algorithm; the experimentation platform that runs interleaving experiments against it is a separate, also-undisclosed subsystem.

Seen in

Last updated · 200 distilled / 1,178 read