A naukri.com initiative
Medium
2w
47
Image Credit: Medium
Embedding-Based Retrieval for Airbnb Search
- Airbnb has developed an Embedding-Based Retrieval (EBR) system to improve search accuracy and scalability for finding relevant homes for users.
- The system aims to narrow down the initial set of homes before using more compute-intensive models for ranking.
- Challenges in building the EBR system included constructing training data, designing the model architecture, and implementing an online serving strategy.
- Training data construction involved using contrastive learning to map homes and search queries into numerical vectors.
- User trips were grouped to identify positive and negative pairs for training the machine learning model.
- The model architecture consisted of a two-tower network design processing features of home listings and search queries separately.
- For online serving, an approximate nearest neighbor (ANN) solution like inverted file index (IVF) was chosen for scalability and performance.
- Using Euclidean distance in the similarity function improved cluster size uniformity and retrieval performance.
- The EBR system led to a significant increase in bookings and displayed more relevant results to users, particularly for queries with many options.
- The system was successfully launched in both Search and Email Marketing productions.
Read Full Article
2 Likes
For uninterrupted reading, download the app