Listen Share In modern search and recommendation systems, embeddings play an important role as a tool for capturing complex relationships between entities such as items and users in a dense format. • Embeddings allow us to represent entities in the same vectorial space and measure similarity. • Embeddings compressed format offer efficient storage fast retrieval which can enable large-scale search and recommendation systems. • AtExpedia Group™ we have been leveraging property embeddings in our search systems which are learnt through the hotel2vec model [1]. • In this work we extend hotel2vec to encompass contextual information about the traveler. • Specifically, we are interested in adapting the property embedding based on the loyalty tier of the traveler.

Article Summaries:

  • Expedia Group has enhanced its hotel search system by extending the hotel2vec embedding model to incorporate traveler loyalty tier information. The updated architecture fuses property features (amenities, ratings, star level, location) with a loyalty‑tier embedding (blue, silver, gold) before projection. Trained on a year of click data, the contextual model shows higher hits@10 and hits@100 compared to the original hotel2vec. When integrated into a ranking model, the tier‑specific embeddings improve NDCG@10 and the quality of top‑20 results, demonstrating that tailoring property representations to loyalty status yields measurable gains in personalized hotel recommendations.

Sources: