Problem
Expedia Group ranks lodging for visitors who search a destination and discover properties. Some visitors arrive from an external channel, such as an influencer's affiliate link, with one specific property in mind. They land on a results page where that property is pinned on top, followed by a ranking of other properties. If the pinned property is unavailable, the visitor needs a close alternative, and a ranking built only for relevance to the search context does not provide one.
Example from the article: a traveler follows a link to a romantic hotel for an anniversary, finds it fully booked for the planned dates, and needs a similar property for the same dates.
Solution in brief
Keep the destination-search ranking model, with the same architecture and ranking loss, and change two things for property searches. Train only on converted property searches, and add features that describe how similar each candidate is to the searched property. The ranking then blends relevance to the search context with similarity to the searched property.
Questions this case answers
- How to rank a candidate list when the user arrives with one specific item in mind that may be unavailable, so relevance alone is the wrong objective
- How to reuse a trained ranking model for a second query type by changing only the training data and the features
- How to rank search results for visitors who arrive from an external channel looking for one specific item?
- How to adapt an existing learning-to-rank model to a new search scenario without building a new architecture?
Tools and methods
- Pairwise ranking loss — a loss derived from LambdaRank, shared by both search types
- Property embeddings (Word2vec-style) — lodging-specific embeddings (hotel2vec) are part of the property features and the basis for similarity
- Two-stage ranking — candidate selection, then the relevance ranking that this post changes
Google or a one-time email code. No password.