Listing Embeddings for Similar Listing Recommendations and Real-time Personalization in Search Ranking
Overview
The article discusses the implementation of Listing Embeddings at Airbnb to enhance Similar Listing Recommendations and real-time personalization in search ranking. It details the methodology behind creating embeddings, their training process, and the significant impact they have on improving booking conversions.
What You'll Learn
1
How to implement Listing Embeddings for personalized recommendations
2
Why embeddings improve search ranking and user engagement
3
How to evaluate the effectiveness of embedding models
Key Questions Answered
What are Listing Embeddings and how are they used?
Listing Embeddings are vector representations of Airbnb homes that encode various features such as location and price into 32-dimensional vectors. They are used to measure similarities between listings, improving recommendations and search ranking.
How does Airbnb train its Listing Embeddings?
Airbnb trains its Listing Embeddings using a technique called Negative Sampling, which updates the embeddings through stochastic gradient descent based on user click sessions, optimizing for both positive and negative context listings.
What improvements were observed after implementing embeddings?
The embedding-based solution led to a 21% increase in Similar Listing carousel click-through rate (CTR) and 4.9% more guests discovering listings they ultimately booked, demonstrating significant enhancements in user engagement.
How are cold-start embeddings created for new listings?
Cold-start embeddings for new listings are generated by averaging the embeddings of the three geographically closest listings that share the same type and price range, ensuring relevant recommendations from the start.
Key Statistics & Figures
Active listings used for training
4.5 million
The embeddings were learned using a dataset of 4.5 million active listings on Airbnb.
Search click sessions used for training
800 million
The embeddings were trained on over 800 million search click sessions.
Increase in Similar Listing carousel CTR
21%
The embedding-based solution led to a 21% increase in click-through rate.
Increase in guests discovering listings
4.9%
4.9% more guests discovered the listings they ended up booking.
Key Actionable Insights
1Implementing Listing Embeddings can significantly enhance user engagement on your platform.By utilizing embeddings, platforms can provide personalized recommendations that align closely with user preferences, leading to higher conversion rates.
2Using Negative Sampling during the training of embeddings helps in optimizing the representation of listings.This technique not only improves the quality of the embeddings but also ensures that the model learns to differentiate between similar and dissimilar listings effectively.
3Regularly evaluate the performance of your embedding models to ensure they meet user needs.Conducting offline evaluations and A/B testing can provide insights into how well the embeddings are performing and where improvements can be made.
Common Pitfalls
1
Neglecting to update embeddings regularly can lead to outdated recommendations.
As new listings are added and user preferences change, failing to refresh the embeddings can result in less relevant suggestions.
2
Overfitting the model to historical data can limit its effectiveness.
If the model is too closely aligned with past user behavior, it may not adapt well to new trends or changes in user preferences.